HABILITATION A DIRIGER DES RECHERCHES ... - Eric Dumonteil

... lycéens, initiative relatée dans un article du "Monde de l'Education" en 2004) ...... Python. Cookbook, eds Martelli A, Ravenscroft A, Ascher D (O'Reilly, Paris).
10MB taille 0 téléchargements 214 vues
FACULTE DES SCIENCES D’ORSAY Ecole Doctorale : Particules, Noyaux, et Cosmos (ED517)

HABILITATION A DIRIGER DES RECHERCHES Sp´ecialit´e : Physique

Eric Dumonteil

Marches al´ eatoires branchantes et Simulation Monte-Carlo du transport des neutrons Soutenue le 22 mai 2014 devant la commission d’examen compos´ee de

Pr´esident :

Pierre DESESQUELLES

Rapporteurs :

Gilles BAN ´ Olivier BENICHOU Andreas PAUTZ

Examinateurs :

Sylvain DAVID Cheikh DIOP

2

E. Dumonteil

Université Paris-Sud

Remerciements

Ce manuscrit présente un travail engagé en 2005, au Service d’Etude des Réacteurs et de Mathématiques Appliquées du CEA Saclay. Je tiens par conséquent à remercier Richard Lenain et Patrick Blanc-Tranchant, successifs chefs de ce service, de m’avoir laissé l’opportunité de consacrer une partie de mon temps à ces travaux de recherche. Je suis vraiment très reconnaissant à Olivier Bénichou, Gilles Ban et Andreas Pautz d’avoir accepté d’être rapporteurs de cette habilitation à diriger des recherches, et à Sylvain David, Pierre Desesquelles, et Cheikh Diop d’avoir bien voulu faire partie du jury. Pour leur lecture si attentive des premières versions de mon mémoire, et pour bien d’autres raisons, je suis content de pouvoir également exprimer ici toute ma gratitude à Mireille Coste, Cyril Dieudonné, Marie-Paule Dumonteil, Cédric Jouanne, Guillaume Gouge, Fausto Malvagi, Alain Mazzolo et Odile Petit. Le troisième chapitre de ce mémoire, en particulier, est le fruit d’un travail commun mené au Laboratoire de Transport Stochastique et Déterministe par le petit groupe “Marches Aléatoires” composé d’Alain Mazzolo, d’Andrea Zoia et de moi-même. C’est très sincèrement que je remercie Andrea de porter avec tant de conviction ces activités, dans un monde où il ne semble plus rester de place qu’aux aspects numériques. Le quatrième et le cinquième chapitre auront en grande partie été le fruit de travaux communs avec Cheikh Diop et Fausto Malvagi, tous deux aussi compétents que disponibles : merci à tous les deux de votre bienveillance ! Enfin, la finalisation de cette HDR m’a beaucoup mobilisé lors de l’année passée, et cette mobilisation n’aurait pas été possible sans le soutien et le support continu de ma femme Caroline, qui a par ailleurs fait preuve d’une patience difficilement compréhensible. Cette dernière année m’a donné un nombre incalculable de raisons pour te remercier, vraiment. Pour faire court, je n’en cite qu’une... à la petite Louise, pour être entrée si vite dans mon coeur, à Caroline, pour m’avoir accepté dans le sien

3

4

E. Dumonteil

Université Paris-Sud

Propos Liminaire

L

a neutronique - l’étude du cheminement des neutrons dans la matière - est une branche de la physique née lors la seconde guerre mondiale, sous l’impulsion des militaires et des scientifiques pour la plupart réunis dans le projet Manhattan à Los Alamos. Son développement est associé depuis l’origine à celui des méthodes Monte-Carlo, mises au point par les mêmes mathématiciens et physiciens. Si les méthodes Monte-Carlo ont connu un essor très important lors des trois dernières décennies, et ont vu leur utilisation se généraliser à des domaines aussi variés que la météorologie ou l’étude des embouteillages, la neutronique est arrivée à maturité bien plus tôt : l’équation de Boltzmann, que l’on résout numériquement avec des codes Monte-Carlo mais aussi avec des codes déterministes, permet dans les domaines de la physique des réacteurs ou de la radioprotection, d’établir des prédictions moyennes solides, validées par des expériences auprès de réacteurs expérimentaux, et ce depuis les années 60. L’explosion conjointe de la capacité de calcul et de stockage des architectures informatiques actuelles est cependant en train de changer profondément cette donne : il est aujourd’hui possible de réaliser de véritables expériences numériques où les ordres de grandeur des statistiques des neutrons simulés sont très proches de celles mises en jeu dans des configurations physiques réalistes. Ainsi, les simulations Monte-Carlo du transport des neutrons permettent d’avoir accès non seulement à des grandeurs moyennes (pour les flux, les taux de réactions, les doses etc.) mais également à tous les autres moments des distributions statistiques de ces grandeurs. Afin d’accompagner ce changement de paradigme de nature technologique, la modélisation théorique et les schémas numériques doivent s’adapter afin de permettre l’accès à ces moments d’ordre supérieur que ne peut fournir une équation portant sur des grandeurs moyennes comme celle de Boltzmann. Mes travaux de recherche et d’ingénierie se situent au centre des ces préoccupations : après avoir présenté mon curriculum vitae détaillé au chapitre 1, et avoir fait une synthèse rapide de mes travaux au chapitre 2, j’expose au chapitre 3 les grandes lignes de mes recherches concernant la modélisation stochastique du transport des neutrons, via une approche fondée sur l’utilisation des intégrales de chemin de Feynman-Kac appliquées à l’étude des processus branchants (le branchement modélise le phénomène de fission induit par les neutrons dans les milieux fissiles). Le pendant numérique formel à cette approche théorique est la simulation Monte-Carlo du transport des neutrons, qui est abordée au chapitre 4. Enfin, au chapitre 5 je montre que l’approche théorique précédemment définie est cruciale pour les simulations numériques car elle permet d’encadrer leur utilisation et également de comprendre de manière qualitative leurs résultats. Pour ce faire, deux exemples importants d’applications de la théorie à la simulation sont considérés : on discute en particulier de la découverte d’un nouvel effet en neutronique dit de “clustering” des neutrons.

5

6

E. Dumonteil

Université Paris-Sud

Table des matières

Remerciements

3

Propos Liminaire

5

1 Curriculum Vitae Détaillé 1.1 Etat Civil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Titres Universitaires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Parcours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Activités d’enseignement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Activités liées à l’Administration, à la Recherche, au Transfert de Technologie . . . 1.5.1 Éléments de bibliométrie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2 Administration de la recherche (participation à des jurys, distinctions, etc.) 1.5.3 Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.4 Transfert de technologie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Encadrement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Travaux doctoraux et post-doctoraux . . . . . . . . . . . . . . . . . . . . . 1.6.2 Stages écoles d’ingénieurs, universitaires . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

11 12 12 12 13 13 13 13 14 14 14 14 14

2 Synthèse des travaux 2.1 Travaux réalisés comme Ingénieur-Chercheur au CEA Saclay . . . . . . . . . . . . . . . . . . . 2.2 Ingénierie et développement des codes, transfert de technologie . . . . . . . . . . . . . . . . . . 2.3 Aspects numériques : simulation des processus de transport, développement de méthodes numériques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Aspects théoriques : modélisation des processus de transport à l’aide des marches aléatoires . . 2.5 Originalité de la démarche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 Liste des Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.1 Articles dans des revues internationales avec comité de lecture . . . . . . . . . . . . . . 2.6.2 Livres ou chapitres de livres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.3 Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’orateur) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.4 Invitations et Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’auteur) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.5 Rapports de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17 . 18 . 18

3 Modélisation stochastique du transport des particules neutres 3.1 Hypothèses principales et modélisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Mouvement Brownien, équation de la diffusion et équation de Feynman-Kac . . . . . . . . . . . 3.2.1 De l’équation de champ moyen au propagateur . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Généralisation des équations de champ moyen aux équations de Feynman-Kac pour la probabilité . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25 . 26 . 27 . 27

7

. . . . . . .

19 19 20 22 22 22 22

. 23 . 23 . 24

. 28

TABLE DES MATIÈRES

8

3.3

Equation de Feynman-Kac pour un processus continu de transport . . . . . . . . . . . . . . . . 3.3.1 Equation de champ moyen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Généralisation des équations de champ moyen aux équations de Feynman-Kac pour la probabilité . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Prise en compte du branchement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Equation pour les moments et polynômes de Bell . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Traitement des processus discrets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Résultats théoriques principaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 Solution de l’équation de Boltzmann linéaire stationnaire en dimension 2 . . . . . . . . 3.7.2 Généralisation aux processus branchants de la formule de Cauchy pour les longueurs de corde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.3 La loi de l’arcsinus de Lévy avec absorption . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.4 Caractérisation de l’enveloppe convexe du déclenchement d’une épidémie . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 31 . 31 . . . . . .

31 34 36 36 39 39

. . . .

41 43 45 50

4 Simulation Monte-Carlo du transport des particules neutres 4.1 Principe de la simulation Monte-Carlo du transport des particules . . . . . . . . . . . . . . . . . 4.1.1 Principe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1.1 Equation de Boltzmann linéaire en neutronique . . . . . . . . . . . . . . . . . . . 4.1.1.2 Résolution de l’équation de Boltzmann par la méthode Monte-Carlo appliquée au transport des particules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1.3 Expression des opérateurs de transport et de collision . . . . . . . . . . . . . . . 4.1.1.4 Estimateurs usuels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.2 Algorithme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 TRIPOLI-4® . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Protection et méthodes de réduction de variance . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Principe de base de la réduction de variance . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2 Réduction de variance pour les problèmes de transport . . . . . . . . . . . . . . . . . . . . 4.2.2.1 Quantification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2.2 Une méthode générique : la capture implicite . . . . . . . . . . . . . . . . . . . . 4.2.2.3 Le schéma de la réduction de variance dans TRIPOLI-4® . . . . . . . . . . . . 4.2.2.4 Roulette russe, fractionnement et échantillonnage stratifié . . . . . . . . . . . . . 4.2.2.5 Utilisation de l’équation de Placzek . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2.6 Réduction de variance par apprentissage . . . . . . . . . . . . . . . . . . . . . . 4.3 Criticité et itération sur la puissance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Exemple illustratif : une cellule de réacteur à eau sous pression . . . . . . . . . . . . . . . 4.3.3 L’itération sur la puissance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Diagnostic de la convergence des sources à l’aide de l’entropie de Boltzmann . . . . . . . . 4.3.5 Diagnostic des corrélations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.5.1 Diagnostic des corrélations par l’ajustement de la relaxation du champ neutronique 4.3.5.2 Diagnostic des corrélations par la mesure de l’autocorrélation de rang k . . . . . 4.3.5.3 Diagnostic des oscillations à l’aide de la transformée de Fourier rapide . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53 54 54 54

5 Apports de la modélisation à la simulation 5.1 Le “clustering” des neutrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 Phénoménologie du clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.2 Modélisation et calcul des corrélations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.3 Développement d’un outil de diagnostic pour les simulations Monte-Carlo de criticité . . 5.2 Accélération des simulations de Monte-Carlo évoluant par la méthode des échantillons corrélés 5.2.1 Principe de la méthode, TRIPOLI-4-D . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.2 Mise en œuvre de l’échantillonnage corrélé pour accélérer les simulations . . . . . . . . . 5.2.3 Calcul des perturbations acceptables par intégrales de chemins . . . . . . . . . . . . . . 5.2.4 Perspectives technologiques, big data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77 78 78 79 87 90 90 91 92 95 97

E. Dumonteil

. . . . . . . . . .

55 56 56 57 57 58 58 59 59 59 59 60 61 61 65 65 66 68 69 71 71 72 73 75

Université Paris-Sud

TABLE DES MATIÈRES

9

6 Projet, perspectives 101 Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 A Démonstrations A.1 Rappel de la notion d’intervalle de confiance et le théorème de la limite centrale . . . . . . . . A.2 Réduction de variance dans les problèmes de transport . . . . . . . . . . . . . . . . . . . . . . A.3 Système de poids pour l’estimation d’une grandeur I . . . . . . . . . . . . . . . . . . . . . . . A.4 Jeu à variance nulle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.4.1 Problème adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.4.2 Jeu à variance nulle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.4.3 Equation de Placzek . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.5 Ordres de grandeur pour la population neutronique dans un cœur de réacteur nucléaire . . . A.6 Les histoires des neutrons : concept de traces . . . . . . . . . . . . . . . . . . . . . . . . . . . A.7 Calcul par intégrales de chemins des scores Monte-Carlo de la simulation initiale et perturbée A.7.1 Simulation initiale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.7.2 Simulation perturbée . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

105 105 105 106 108 108 109 110 111 111 112 112 112

B Publications principales 117 B.1 liées au Chapitre 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 B.2 liées au Chapitre 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 B.3 liées au Chapitre 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

E. Dumonteil

Université Paris-Sud

TABLE DES MATIÈRES

E. Dumonteil

10

Université Paris-Sud

Chapitre

1

Curriculum Vitae Détaillé Sommaire 1.1 1.2 1.3 1.4 1.5

Etat Civil . . . . . . . . . . . . . . . . . . . . . . . . Titres Universitaires . . . . . . . . . . . . . . . . . Parcours . . . . . . . . . . . . . . . . . . . . . . . . . Activités d’enseignement . . . . . . . . . . . . . . . Activités liées à l’Administration, à la Recherche, 1.5.1 Éléments de bibliométrie . . . . . . . . . . . . . . . 1.5.2 Administration de la recherche (participation à des 1.5.3 Communication . . . . . . . . . . . . . . . . . . . . 1.5.4 Transfert de technologie . . . . . . . . . . . . . . . 1.6 Encadrement . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Travaux doctoraux et post-doctoraux . . . . . . . 1.6.2 Stages écoles d’ingénieurs, universitaires . . . . . .

11

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . au Transfert de Technologie . . . . . . . . . . . . . . . . . . . jurys, distinctions, etc.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

12 12 12 13 13 13 13 14 14 14 14 14

CHAPITRE 1. CURRICULUM VITAE DÉTAILLÉ

1.1

12

Etat Civil

Eric Dumonteil Né le 17 avril 1977 à Thionville (57100) Nationalité française Marié, un enfant Tél. : 06 61 19 98 21 Email : [email protected] Adresse : 22 rue gabriel péri, 92120 Montrouge Ingénieur-Chercheur Commissariat à l’Energie Atomique et aux Energies Alternatives (CEA/DEN/DM2S/SERMA/LTSD) Maître de conférence Institut National des Sciences et Techniques Nucléaires Tél. Pro. : 01 69 08 55 76 (Fax : 01 69 08 45 72) Email Pro. : [email protected] Adresse Pro. : Bâtiment 470, 91191 Gif sur Yvette

1.2

Titres Universitaires

2001–2004

Doctorat ès Sciences, spécialité Constituants Elémentaires, Université de Caen Thèse soutenue le 13 septembre 2004 et effectuée au Service de Physique Nucléaire du CEA Saclay (DSM/IRFU) Sujet : "Etude des Résonances de la Famille du Υ dans les collisions d’ions lourds ultrarelativistes de l’expérience ALICE du LHC au CERN" Directeur : Bernard Tamain, Professeur Ecole Nationale Supérieure d’Ingénieurs de Caen, et Alberto Baldisseri, Ingénieur-Chercheur CEA Rapporteurs : Michel Gonin, Directeur de recherche CNRS, Professeur Ecole Polytechnique, et Pascal Dupieux, Directeur de recherche CNRS Mention Très honorable et Félicitations du Jury

2001–2002

Auditeur libre au DEA “Physique Théorique”, Ecole Normale Supérieure de Paris

2001 2001

Ingénieur diplômé de l’Ecole Nationale Supérieure d’Ingénieurs de Caen, cursus international DEA “Physique de la Matière et du Rayonnement”, Université de Caen en double diplôme

1.3

Parcours

Depuis 2012 Depuis 2011 Depuis 2010 Depuis 2009 Depuis 2005 2001–2004

2000

E. Dumonteil

Maître de conférence Institut National des Sciences et Techniques Nucléaires Membre du LABEX P2IO (Laboratoire d’Excellence "Physique des 2 Infinis et des Origines") Membre du LRC Manon (Laboratoire de recherche conventionné entre le Laboratoire JacquesLouis Lions (LJLL), unité mixte UPMC-CNRS, et le Département de Modélisation des Systèmes et Structures (DM2S) de la Direction de l’Energie Nucléaire (DEN) du CEA-Saclay) Chargé d’enseignements à Institut National des Sciences et Techniques Nucléaires Ingénieur-Chercheur au CEA de Saclay (Service d’Etudes des Réacteurs et de Mathématiques Appliquées, Laboratoire de Transport Stochastique et Déterministe) Allocataire de recherche (bourse ministérielle MENRT) détaché au CEA Saclay ; séjours réguliers au CERN (Genève, Suisse) pour la mise au point des chambres à muons du détecteur ALICE du LHC ainsi qu’au Brookhaven National Laboratory (BNL, New York, USA) pour participer à l’expérience PHENIX du RHIC Assistant de recherche à la George Washington University (Washington DC, USA), en détachement sur le site de l’accélérateur CEBAF en Virginie

Université Paris-Sud

CHAPITRE 1. CURRICULUM VITAE DÉTAILLÉ

1.4

13

Activités d’enseignement

Depuis 2011 Depuis 2007 Depuis 2006 Depuis 2006 Depuis 2006

1.5

Enseignant à l’Ecole Centrale de Paris (cours EN1700 "Eléments de Neutronique et de Physique des Réacteurs Nucléaires", 15h) Enseignant au Master Nuclear Energy (M2) de l’Université Paris-Sud (TP "code TRIPOLI4®", 18h) Enseignant au Génie Atomique de l’Institut National des Sciences et Techniques du Nucléaire (9h de TD "Neutronique" et 18h de TP "code TRIPOLI-4®") Enseignant formation CRISTAL dispensée à l’Institut de Radioprotection et de Sûreté Nucléaire (cours "Méthode de Monte-Carlo appliquée au transport de particules", 3h) Enseignant formation ICC (Ingénieurs Criticiens de Centre) dispensée conjointement par l’Institut de Radioprotection et de Sûreté Nucléaire et l’Institut National des Sciences et Techniques du Nucléaire (cours "Méthode de Monte-Carlo appliquée au transport de particules", 1h30)

Activités liées à l’Administration, à la Recherche, au Transfert de Technologie

1.5.1

Éléments de bibliométrie

Les données/statistiques complètes de mes publications sont accessibles sur le site Google Scholar dédié 1 . On donne cependant ici quelques “indicateurs académiques” : • Articles publiés dans des revues à comité de lecture : 40+ • Livres ou chapitres de livres : 3 • H-index : 8 • Nombre de citations : 1900+ • Reviewer pour les revues internationales : Journal of Computational Physics, Annals of Nuclear Energy, Nuclear Science and Engineering depuis 2006 La dernière section de ce chapitre liste l’ensemble de ces publications.

1.5.2 •

• •



• •



• •

• • • • •

Administration de la recherche (participation à des jurys, distinctions, etc.)

Membre du jury de la thèse de Antoine Collin à l’Université Paris-Sud portant sur l’"Etude des antineutrinos de réacteurs : mesure de l’angle de mélange leptonique θ13 et recherche d’éventuels neutrinos stériles" (2014) Membre du comité technique de la conférence Phytra3 (Tetouan, 2014) Co-lauréat du projet “Systèmes Nucléaires” de la mission interdisciplinarité NEEDS (Nucléaire : Energie, Environnement, Déchets, Société) commune au CEA, au CNRS et à l’ANDRA (2013) Membre du jury de la thèse de Cyril Dieudonné à l’Université Paris-Sud portant sur les "Accélération de la simulation Monte Carlo du transport des neutrons dans un milieu évoluant par la méthode des échantillons corrélés" (2013) Membre du comité technique de la conférence SNA&MC 2013 (Paris, 2013) Nomination en tant qu’«Expert du CEA» dans le domaine des mathématiques et de l’informatique, spécialité physique statistique (à compter du 1er janvier 2013) Membre du groupe d’experts “Advanced Monte-Carlo Techniques” du comité WPNCS (Working Party on Nuclear Criticality Safety) de l’Agence à l’Energie Nucléaire (2012) Co-organisateur de la conférence Phytra2 (Fez, Maroc, 2011) Membre du jury de la thèse de Julie Brizi à l’Université Paris-Sud portant sur les "Cycles uranium et thorium en réacteurs à neutrons rapides refroidis au sodium" (2010) Membre du comité technique de la conférence “Physics of Reactors 2010” (Pittsburgh, Etats-Unis, 2010) Chef de laboratoire (LTSD) par intérim sur quelques périodes courtes (2010/2011) Co-organisateur de la conférence Geant4 (Paris, France, 2007) Chairman des sessions "Transport Monte-Carlo" pour plusieurs conférences internationales Membre des projets européens NURESIM et NURISP depuis 2007

1. http ://scholar.google.fr/citations ?user=q44jHq0AAAAJ&hl=fr

E. Dumonteil

Université Paris-Sud

CHAPITRE 1. CURRICULUM VITAE DÉTAILLÉ

1.5.3 •

• •





Encadrement

1.6.1









Travaux doctoraux et post-doctoraux

Co-directeur de la thèse (encadrement à 100%) à l’Université Paris-Sud de Cyril Dieudonné intitulée "Mise en œuvre de l’échantillonnage corrélé pour la réduction de variance de calculs Monte-Carlo évoluants" (soutenue le 12 décembre 2013, Directeur de thèse : Pr. C. Diop). Ce travail a été en partie publié dans les proceedings de deux conférences (présentation orale et article acceptés à M&C 2013 (Sun Valley, EtatsUnis) et SNA&MC 2013 (Paris, France)), et a donné lieu à deux articles acceptés dans la revue Annals of Nuclear Energy. Co-encadrement (50% Eric Dumonteil, 50% Emeric Brun) du séjour doctoral (2 mois) au CEA Saclay de Ricardo Reyes Ramires de l’Universidad Nacionale Autonoma de Mexico visant à comparer les codes TRIPOLI-4®-D et MCNPX sur un cœur rapide refroidi au gaz (2010). Cette collaboration a mené à la publication commune d’un article dans la revue Annals of Nuclear Energy. Co-encadrement (50% Eric Dumonteil, 50% François Xavier Hugot) du post-doctorat (1 an) de Brian Nease intitulé "Accélération de la convergence des sources dans les calculs de criticité" (2009). Suite à ce stage post-doctoral, M. Nease et moi-même avons publié un article dans les proceedings de la conférence SNA&MC 2010 (Tokyo, Japon). Encadrement (75% Eric Dumonteil, 25% Odile Petit) lors de plusieurs séjours doctoraux courts (de 1 semaine à 1 mois) de Stavros Christoforou de la Delft University of Technology (Hollande). Ces séjours visaient à mettre en œuvre dans le code TRIPOLI-4® un schéma de biaisage de type zéro variance. Ce travail a notamment mené à la publication d’un article dans les proceedings de la conférence PHYSOR 2006 (Vancouver, Canada).

1.6.2 •

Transfert de technologie

Co-développeur du code neutronique Monte-Carlo TRIPOLI-4® (registered trademark of CEA), code Monte-Carlo de transport de particules, référence pour les études de neutronique chez EDF, et utilisé par AREVA. Ce code est distribué par l’AEN (Agence pour l’énergie nucléaire) de l’OCDE (voir http ://www.oecd-nea.org/tools/abstract/detail/nea-1716) et est utilisé par plus de 20 pays différents. Co-développeur avec Emeric Brun du code Monte-Carlo TRIPOLI-4®-D, code Monte-Carlo couplant transport de particules et évolution (élaboration de l’architecture de ce code). Support dans TRIPOLI-4® et TRIPOLI-4®-D de la librairie ROOT (librairie C++ pour la physique des particules développée par le CERN). L’utilisation de cette librairie ouvre de nombreuses possibilités en terme de pré-traitement des données (géométrie ROOT et GEANT4 compatibles avec TRIPOLI-4®) ainsi qu’en terme de post-traitement via la conception d’un module ad hoc, t4ROOT tools.

1.6



Communication

Publication d’un “Fait Marquant” du DM2S (CEA/DEN) sur une collaboration SERMA/SPP (Service de Physique des Particules) visant à étudier certains aspects de l’expérience de physique fondamentale double Chooz à l’aide d’outils de calcul neutronique (2011) Responsable des séminaires au SERMA (CEA/DEN/DM2S) depuis 2006 Vulgarisation scientifique (équipe de communication du CEA avec, notamment l’organisation de rencontres "café des sciences" pour les lycéens, initiative relatée dans un article du "Monde de l’Education" en 2004)

1.5.4 •

14

Stages écoles d’ingénieurs, universitaires

Encadrement (100%) du stage de fin d’études (5 mois) au Politecnico di Torino de Davide Artusio intitulé "Temporal and spatial correlations in Monte-Carlo criticality codes" (2012/2013). Suite à ce stage un papier a été accepté comme présentation orale et publié dans les proceedings de la conférence SNA&MC 2013 (Paris, France) Encadrement (100%) du stage Master 2 (6 mois) à l’université Paris 7 de Cyril Dieudonné intitulé "Echantillonnage corrélé et Monte-Carlo évoluant" (2010)

E. Dumonteil

Université Paris-Sud

CHAPITRE 1. CURRICULUM VITAE DÉTAILLÉ •













15

Encadrement (100%) du stage de 3ème année (5 mois) à l’Ecole Polytechnique de Zixian Jiang intitulé "Accélération par un réseau de neurones d’un calcul de transport Monte-Carlo" (2009) Encadrement (100%) du stage de 3ème année (6 mois) à l’UTC Compiègne de Mingchao Yang intitulé "Implémentation d’outils statistiques dans le code TRIPOLI-4®" (2009) – Encadrement (100%) du stage de 3ème année (5 mois) à l’ENSIMAG de Benjamin Baumann intitulé "Externalisation du scorer de TRIPOLI-4®" (2008) Co-encadrement (50% Eric Dumonteil, 50% Fausto Malvagi) du stage de 3ème année (5 mois) du Génie Atomique de Emeric Brun intitulé "Validation du Monte-Carlo évoluant dans le cas des RNR" (2008). Il est à noter que cet encadrement a conduit au recrutement de M. Brun au CEA puis à une publication commune dans la revue Progress in Nuclear Science and Technology. Encadrement (100%) du stage de 3ème année (6 mois) à l’ISIMA de Laurent Minot intitulé "Implémentation d’une interface graphique dans TRIPOLI-4®" (2007) Encadrement (100%) du stage de 3ème année de l’ENSPS de Julie Brizi (6 mois) intitulé "Etude de l’activation des structures d’un dispositif de fusion thermonucléaire" (2006) Encadrement (100%) du stage de 3ème année à l’ENSPG de Aurélien LePeillet (6 mois) intitulé "Contribution à la qualification du code Monte-Carlo TRIPOLI-4®.3 pour les études de criticité. Convergence des sources" (2005). Ce travail a été publié dans les proceedings de la conférence internationale Physor 2006. Encadrement (100%) du stage de 3ème année de l’UTC de Guillaume Pichon (8 mois) intitulé "Validation d’un nouveau moteur de géométrie 3D pour le code de transport Monte-Carlo TRIPOLI-4®" (2005). Les développements informatiques issus de ce stage ont été intégrés à la version industrielle du code MonteCarlo TRIPOLI-4® et sont largement utilisés au CEA et chez EDF.

E. Dumonteil

Université Paris-Sud

CHAPITRE 1. CURRICULUM VITAE DÉTAILLÉ

E. Dumonteil

16

Université Paris-Sud

Chapitre

2

Synthèse des travaux Sommaire 2.1 2.2 2.3

Travaux réalisés comme Ingénieur-Chercheur au CEA Saclay . . . . . . . . . . . . Ingénierie et développement des codes, transfert de technologie . . . . . . . . . . Aspects numériques : simulation des processus de transport, développement de méthodes numériques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Aspects théoriques : modélisation des processus de transport à l’aide des marches aléatoires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Originalité de la démarche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 Liste des Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.1 Articles dans des revues internationales avec comité de lecture . . . . . . . . . . . . . 2.6.2 Livres ou chapitres de livres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.3 Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’orateur) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.4 Invitations et Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’auteur) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.5 Rapports de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17

18 18 19 19 20 22 22 22 22 23 23 24

CHAPITRE 2. SYNTHÈSE DES TRAVAUX

18

D

ans ce chapitre je fais une synthèse de mes travaux, en présentant dans un premier temps les aspects ingénierie et développement de codes, puis les aspects numériques (portant principalement sur la simulation Monte-Carlo du transport des particules), et finalement les aspects théoriques (modélisation du transport des particules à l’aide de l’étude des marches aléatoires branchantes). Une dernière section est dévolue à la bibliographie de l’ensemble de mes travaux publiés.

2.1

Travaux réalisés comme Ingénieur-Chercheur au CEA Saclay

Mon travail doctoral reposant en partie sur la simulation Monte-Carlo de la génération de particules fondamentales et de leur transport (dans le cadre de l’expérience ALICE au LHC du CERN), un poste d’IngénieurChercheur au CEA Saclay dans un domaine proche m’a été offert en 2005, dans le Laboratoire de Protection et de Probabilités (LEPP) du Service d’Etudes des Réacteurs et de Mathématiques Appliquées de la Direction de l’Energie Nucléaire du CEA. Ce poste portait sur le développement des méthodes Monte-Carlo ainsi que sur le développement du code Monte-Carlo de transport TRIPOLI-4®, qui permet de réaliser des simulations de neutronique (propagation des neutrons et des gammas) à des énergies de 0 à 20 MeV. Ce code est principalement utilisé pour la physique des réacteurs nucléaires, la propulsion navale, et l’instrumentation. Suite à une réorganisation interne du service en 2007, mon activité s’est poursuivie au sein du LTSD, Laboratoire de Transport Stochastique et Déterministe où j’ai développé, en 2009/2010 un nouveau code de calcul - TRIPOLI-4®-D couplant le transport des neutrons et l’évolution des compositions, code dont le développement a été repris par la suite par Emeric Brun. Mes activités au sein de ce laboratoire se regroupent aujourd’hui en trois thématiques : • •



2.2

développements applicatifs et d’ingénierie sur les codes TRIPOLI-4® et TRIPOLI-4®-D développement de méthodes numériques autour de ces codes (techniques de réduction de variance, itération sur la puissance, etc.) travaux de modélisation et travaux théoriques portant principalement sur la compréhension qualitative (et partiellement quantitative) de certains aspects du transport des neutrons dans les milieux fissiles, ainsi que sur la théorie du Monte-Carlo elle-même à travers des problématiques comme la propagation des incertitudes, l’étude des biais, la théorie des estimateurs etc.

Ingénierie et développement des codes, transfert de technologie

Outre les tâches de développement applicatif routinières sur les codes TRIPOLI-4® et TRIPOLI-4®-D, ma démarche depuis 2005 a consisté à faire profiter ces codes des briques logicielles (et de la philosophie d’utilisation qui leur est associée) issues de la physique des particules, plus particulièrement des fonctionnalités offertes par la librairie C++ ROOT (sous license libre) développée au CERN. En effet, les expériences complexes de physique des particules du LHC ont nécessité la mise au point dans la librairie ROOT d’un moteur de géométrie performant qui propose des capacités de navigation optimisées pour des géométries comportant plusieurs centaines de millions de volumes physiques. Comme la description fine du cœur des réacteurs nucléaires implique elle aussi l’utilisation de géométries très complexes, j’ai mené dans TRIPOLI-4® des travaux permettant d’inclure cette fonctionnalité, ouvrant aux utilisateurs neutroniciens la possibilité d’en bénéficier. De la même manière, les analyses de physique des particules sont faites par des logiciels s’appuyant sur la notion de “trace” des particules dans les détecteurs (dont l’origine historique est liée à la trace laissée par des particules comme des muons cosmiques dans une chambre à bulles, cf appendice A.6). J’ai travaillé, dans TRIPOLI-4®, à la mise au point de ce même concept de trace pour les neutrons : pendant la simulation de l’histoire d’une particule, les différentes étapes de la vie de celle-ci sont écrites sur le disque dur de la machine. Le code ROOT offre alors de très nombreuses possibilités nouvelles d’analyse a posteriori (c’est-à-dire après la simulation). Ce type de post-traitement permet par exemple à un utilisateur de chercher des événements rares dans sa simulation Monte-Carlo, à l’origine de sauts de variance lors du déploiement des méthodes de réduction de variance.

E. Dumonteil

Université Paris-Sud

CHAPITRE 2. SYNTHÈSE DES TRAVAUX

2.3

19

Aspects numériques : simulation des processus de transport, développement de méthodes numériques

Concernant le développement des méthodes numériques, mes travaux se sont concentrés autour de trois types d’utilisation du transport Monte-Carlo : la radioprotection, la criticité, et l’évolution. En radioprotection, les enjeux sont liés aux fortes atténuations du flux des particules (typiquement de l’ordre de 10−12 − 10−16 ), qui imposent des temps de calcul très longs pour obtenir des précisions correctes sur l’estimation des grandeurs d’intérêt ou de leurs incertitudes. Pour accélerer les simulations, il s’agit alors de mettre en œuvre des méthodes de réduction de variance. Celles-ci consistent à augmenter la concentration de particules simulées dans les volumes de détection, afin que l’erreur statistique associée à la mesure soit la plus faible, et ce sans modifier les valeurs moyennes des grandeurs d’intérêt. Ces méthodes utilisent souvent un précalcul déterministe approché de la configuration physique étudiée. A l’aide des fonctionnalités liées à l’enregistrement des traces permettant une étude collision-par-collision des neutrons, j’ai proposé un autre type de méthode reposant sur la mise en place d’une stratégie d’apprentissage de la simulation qui peut être accélérée à l’aide, par exemple, d’un réseau de neurones [8]. Les calculs de criticité, où les neutrons qui fissionnent sont transportés génération par génération, posent quant à eux des problèmes d’une autre nature : ce suivi génération par génération induit des corrélations dans les suites permettant d’estimer les grandeurs d’intérêt du calcul, brisant ipso facto les hypothèses du théorème central limite utilisé pour l’estimation de l’erreur statistique du calcul. J’ai ainsi pu proposer un outil de diagnostic de ces corrélations ainsi qu’une méthode (dite de super-cycle) permettant de corriger le biais dans l’estimation de cette erreur [14]. Enfin, j’ai développé un code Monte-Carlo évoluant (TRIPOLI-4®-D [13]) permettant de prendre en compte l’effet dans le temps des irradiations sur les compositions, en alternant des successions de simulations de transport spatial (résolution de l’équation de Boltzmann) et d’évolution temporelle (résolution de l’équation de Bateman). Ces dernières équations sont non-linéaires par rapport aux flux, et par conséquent le couplage entre le transport et l’évolution est lui-même non-linéaire, ce qui induit des biais dans l’utilisation des estimateurs Monte-Carlo usuels. Un travail exposé dans les articles [12, 11] propose une méthode originale de correction de ces biais.

2.4

Aspects théoriques : modélisation des processus de transport à l’aide des marches aléatoires

Concernant les travaux théoriques, ceux-ci ont été initiés grâce à une collaboration intra-laboratoire entre MM. Zoia Andrea, Mazzolo Alain et moi-même. Ils reposent sur l’utilisation de la mécanique statistique et des intégrales fonctionnelles, particulièrement les intégrales de chemin de Feynman-Kac. Notre objectif est de décrire au mieux la marche aléatoire des neutrons dans les milieux fissiles, afin de la comprendre de manière qualitative, et éventuellement de suppléer de manière quantitative les codes de type Monte-Carlo - capables de décrire de manière très fine et avec très peu d’approximations le comportement des neutrons. Il s’agit ainsi d’utiliser la théorie des marches aléatoires pour généraliser les processus de diffusion (représentant la limite en loi de tous les sauts à variance finie) aux processus de transports branchants (on cherche à décrire le transport dans l’espace des particules ainsi que la fission des noyaux à l’origine du “branchement” des neutrons). Par exemple, dans le cas particulier des milieux homogènes, la marche des particules neutres suit une loi de sauts exponentiels (les sauts exponentiels sont une classe particulière de marches aléatoires, dite de Pearson, définie sur Rd et dont la distance inter-collisions est distribuée exponentiellement). Dans un premier temps, nous nous sommes concentrés sur la construction du propagateur de ces sauts, en privilégiant sa forme discrète (le propagateur discret donne la probabilité qu’une particule ayant collisionné ou étant émise au point x0 se rende en un point x1 en n sauts précisément). La somme du propagateur sur toutes les valeurs possibles de n permet alors de construire la densité de collision, qui est solution de l’équation de Boltzmann stationnaire et ce pour des domaines finis ou infinis, avec ou sans fuites. Cette démarche, publiée dans [10], a permis de dégager plusieurs résultats originaux comme la solution de l’équation de Boltzmann stationnaire en dimension 2. Puis nous nous sommes intéressés au lien entre le temps de résidence d’une particule dans un volume et le nombre de collisions qu’elle y subit [9]. Ce thème présente un intérêt théorique mais également appliqué car il permet d’établir la relation entre les estimateurs du flux dits “cordes” et “collisions” utilisés par les codes Monte-Carlo de transport de particules. Si l’on considère le cas particulier des sauts exponentiels, comme ceux-ci sont markoviens dans les variables position et direction en tous points de leur parcours, nous avons alors utilisé le théorème de Kac pour mettre en équation puis résoudre le "rod model", modèle simple de sauts en 1

E. Dumonteil

Université Paris-Sud

CHAPITRE 2. SYNTHÈSE DES TRAVAUX

20

dimension permettant d’établir de solides prédictions qualitatives [8]. Une généralisation de ces travaux dans le cas où la marche aléatoire est anisotrope est par ailleurs exposée dans [7] . L’ensemble de ces développements nous a alors permis de proposer une formulation discrète de la formule de Feynman-Kac, permettant d’établir des statistiques de comptage. Après avoir montré que dans la limite d’un grand nombre de collisions (où s’opère le passage du discret au continu) nous retrouvions les résultats de la diffusion, cette formulation nous a permis par exemple - outre les problèmes de transport neutronique - de formuler une généralisation de la loi de l’arcsinus prenant en compte des absorptions ou de revisiter le problème de la ruine du joueur en y apportant des résultats originaux [6]. Enfin, l’étude dans ce cadre des processus branchants (dits super-processus, et qui permettent de décrire la fission induite par les neutrons) nous a permis non seulement de comprendre certains aspects physiques de la neutronique (pour des systèmes de dimension 3 [5, 4, 3]) mais aussi d’établir un modèle original de description du déclenchement des épidémies en biologie (système en dimension 2 [9]) ou de décrire le phénomène de clustering des neutrons découvert récemment [1].

2.5

Originalité de la démarche

Les différentes thématiques présentées ci-dessus s’articulent en fait très naturellement, comme le montre la figure 2.1.

Figure 2.1 Articulation de mes différentes thématiques de recherche. Les différents domaines impliqués apparaissent en bleu, et mes contributions en rouge. En effet, les marches aléatoires sont au centre de mes activités. Bien que celles-ci soient décrites formellement par les mathématiques, ce sont les outils de la mécanique statistique et de la physique statistique des systèmes hors-équilibre qui permettent de modéliser au mieux les processus à l’œuvre en neutronique, afin de bâtir des modèles qualitatifs. Dans ce cadre, l’équation de Feynman-Kac revêt une importance toute particulière car elle donne accès à tous les moments d’une distribution de probabilité, à la différence de l’équation de Boltzmann qui n’est qu’une équation de champs moyens. Le chapitre 3 présentera ainsi mes travaux sur l’établissement d’une

E. Dumonteil

Université Paris-Sud

CHAPITRE 2. SYNTHÈSE DES TRAVAUX

21

équation “à la Feynman-Kac” du transport des neutrons, en donnant des exemples d’applications dans d’autres domaines comme l’étude de la propagation des épidémies ou celle de la théorie des jeux. Le pendant quantitatif de cette équation est la simulation Monte-Carlo de la marche aléatoire, c’est-à-dire celle des vols exponentiels des neutrons dans les milieux homogènes dans notre cas. Comme pour l’équation de Feynman-Kac, les simulations Monte-Carlo donnent accès à tous les moments des distributions de probabilité des grandeurs d’intérêt. De plus, l’évolution de la puissance de calcul des ordinateurs permet à présent d’effectuer de manière routinière (quelques heures) le transport de milliards de neutrons, et ce avec un nombre extrêmement limité d’hypothèses puisque les principaux codes Monte-Carlo de neutronique sont des codes en 3 dimensions, permettant de suivre le transport de neutrons en espace, en temps et en énergie de manière très fine. Cela motive l’utilisation de ces codes par les différents acteurs de l’industrie nucléaire pour établir, par exemple, le plan de chargement d’un réacteur nucléaire, ou pour dimensionner une installation de radioprotection. Je présenterai donc au chapitre 4 certaines de mes contributions au développement du code TRIPOLI-4®, en essayant de montrer toute la pertinence de l’utilisation, en neutronique, d’outils issus de la physique des particules (et de l’ingéniérie informatique du CERN en particulier). Enfin on montrera au chapitre 5 l’apport de l’étude des marches aléatoires à la simulation du transport des particules à travers deux exemples concrets : le clustering des neutrons et le Monte-Carlo dit “évoluant”.

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

22

Références 2.6 2.6.1

Liste des Publications Articles dans des revues internationales avec comité de lecture

[1] E. Dumonteil, F. Malvagi, A. Zoia, D. Artusio, C. Dieudonné, "Particle clustering in Monte-Carlo criticality simulations", Annals of Nuclear Energy 63, 612-618 (2014) [2] E. Dumonteil, S. Majumdar, A. Rosso, A. Zoia, "Spatial extent of an outbreak in animal epidemics", Proceedings of the National Academy of Sciences 110(11), 4239-4244 (2013) [3] A. Zoia, E. Dumonteil, A. Mazzolo, "Properties of branching exponential flights in bounded domains", EPL 100, 40002 (2012) [4] A. Zoia, E. Dumonteil, A. Mazzolo, S. Mohamed, "Branching exponential flights : travelled lengths and collision statistics", J. Phys. A : Math. Theor. 45 425002 (2012) [5] A. Zoia, E. Dumonteil, A. Mazzolo, "Discrete Feynman-Kac formulas for branching random walks", EPL 98, 40012 (2012) [6] A. Zoia, E. Dumonteil, A. Mazzolo, "Counting statistics : a Feynman-Kac perspective", Physical Review E 85, 011132 (2012) [7] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision statistics for renewal processes with anisotropic scattering and absorption", Physical Review E 84 061130 (2011) [8] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision densities and mean residence times for d-dimensional exponential flights", Physical Review E 83 041137 (2011) [9] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision number statistics for transport processes", Physical Review Letters 106, 220602 (2011) [10] A. Zoia, E. Dumonteil, A. Mazzolo, "Residence time and collision statistics for exponential flights : the rod problem revisited", Physical Review E 84 021139 (2011) [11] E. Brun, E. Dumonteil, F. Malvagi, "Uncertainties propagation in Monte-Carlo burnup codes. Implementation in TRIPOLI-4-D", Progress in Nuclear Science and Technology, Vol2, P879-885 (2011) [12] E. Dumonteil, C. Diop, "Biases and statistical error bars in Monte-Carlo burnup calculations : an unbiased stochastic scheme to solve Boltzmann/Bateman coupled equations", Nuclear Science and Engineering, Vol 167, N2, p165-170 (2011) [13] R. Reyes Ramirez et al., "Comparison of MCNPX-C90 and TRIPOLI-4-D for fuel depletion calculations of a gas-cooled fast reactor", Annals of Nuclear Energy, Vol37, 8, p1101-1106 (2010) [14] E. Dumonteil and T. Courau, "Dominance ratio assessment and Monte-Carlo criticality calculations : dealing with high dominance ratio systems", Nuclear Technology, Vol172, 2, p120-131 (2009) [15] E. Dumonteil, "On a new variance reduction technique : neural network biasing. Applications on two test cases in the Monte-Carlo code TRIPOLI-4", Nuclear Technology (2009) [16] Aamodt K. et al., "The ALICE experiment at the CERN LHC", Journal of Instrumentation 3, S08002 (2008) [17] E. Dumonteil et al., "An Overview on the Monte-Carlo Particle Transport Code TRIPOLI-4", Transactions of the American Nuclear Society vol97 p694-695 (2007) [18] C. Diop et al., "TRIPOLI-4 : A 3D Continuous energy Monte-Carlo transport code", Transactions of the American Nuclear Society, vol95, p661 (2006) [19] F. Carminati et al., "Alice Physics performance report Volume 1", Journal of Physics. G : Nucl. Part. Phys., 30, 1517 (2004) [20] C. Cicalo et al., "The tracking system of the ALICE Dimuon Spectrometer", IEEE NSSCR, Vol1, p660-664 (2003)

2.6.2

Livres ou chapitres de livres

[21] "La neutronique", Monographie de la direction de l’énergie nucléaire, Editions Le Moniteur

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

23

2.6.3 Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’orateur) [22] E. Dumonteil et al., "Neutron clustering in Monte-Carlo criticality simulations", Monte-Carlo and Super Computing in Nuclear Applications 2013, Paris, France (2013) [23] E. Dumonteil and F. Malvagi, "Automatic Treatment of the Variance Estimation Bias in TRIPOLI-4 Criticality Calculations", International Congress On the Advances in Nuclear Power Plants, Chicago, Etats-Unis (2012) [24] E. Dumonteil and F. Malvagi, "Variance estimation in Monte-Carlo criticality calculations", 2d International Conference on Physics and Technology of Reactors and Applications, Fez, Morroco (2011) [25] E. Dumonteil, E. Brun and F. Malvagi, "Uncertainties propagation in Monte-Carlo burnup codes. Implementation in TRIPOLI-4-D", Joint International Conference of the 7th Supercomputing in Nuclear Application and the 3rd Monte-Carlo, Tokyo, Japan (2010) [26] E. Dumonteil and C.M.B. Diop, "Unbiased minimum variance of a matrix exponential function. Application to Boltzman/Bateman coupled equations solving", International Conference on Mathematics, Computational Methods and Reactor Physics 2009, Saratoga Springs, New York, USA (2009) [27] E. Dumonteil, "On a new variance reduction technique", 11th International Conference on Radiation Shielding, Pine Mountain Georgia, USA (2008) [28] E. Dumonteil and T. Courau, "A New Method to Compute the Dominance Ratio in Monte-Carlo Simulations Application to a Simple Pin Cell with The 3-D Monte-Carlo Code TRIPOLI-4", Proceedings of International Congress on Advances in Nuclear Power Plants, Anaheim, CA, USA (2008) [29] E. Dumonteil et al., "Source convergence diagnostics using Boltzmann entropy criterion : application to different OCDE/NEA criticality benchmarks with the 3-D Monte-Carlo code TRIPOLI-4", Physics of Reactors 2006 : Advances in Nuclear Analysis and Simulations, Vancouver, Canada (2006)

2.6.4 Invitations et Articles dans les Proceedings de conférences internationales avec comité de lecture (en tant qu’auteur) [30] A. Zoia, E. Dumonteil, A. Mazzolo, "Residence times and survival probabilities for branching exponential flights", Invited talk at the "Groupe de travail Modèles stochastiques en finance", CMAP, Ecole Polytechnique, France (2014) [31] A. Zoia, E. Dumonteil, A. Mazzolo, "Residence times and survival probabilities for branching exponential flights", Invited talk at the Workshop on Hitting times and exit problems for stochastic models, Université de Bourgogne, France (2013) [32] A. Zoia, E. Dumonteil, A. Mazzolo, "Residence time and collision number statistics : a Feynman-Kac approach", Invited talk at the International Conference on "Random Processes, Conformal Field Theory and Integrable Systems", Moscow, Russia (2011) [33] B. Nease, E. Dumonteil, "A decorrelation technique for iterated Monte-Carlo source calculation", Joint International Conference of the 7th Supercomputing in Nuclear Application and the 3rd Monte-Carlo, Tokyo, Japan (2010) [34] E. Dumonteil and C.M.B. Diop, "Unbiased minimum variance of a matrix exponential function. Application to Boltzman/Bateman coupled equations solving", International Conference on Mathematics, Computational Methods and Reactor Physics 2009, Saratoga Springs, New York, USA (2009) [35] C.M.B. Diop et al., "TRIPOLI-4 : A 3D Continuous energy Monte-Carlo transport code", 1st International Conference on Physics and Technology of Reactors and Applications, Marrakech, Morroco (2007) [36] A. L’Abbate, T. Courau and E. Dumonteil, "Monte-Carlo criticality calculations : source convergence and dominance ratio in an infinite lattice using TRIPOLI-4 and MCNP", 1st International Conference on Physics and Technology of Reactors and Applications, Marrakech, Morroco (2007) [37] S.Christoforou et al., "A zero variance based scheme for variance reduction in Monte-Carlo criticality" , Physics of Reactors 2006 : Advances in Nuclear Analysis and Simulations, Vancouver, Canada (2006)

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

2.6.5

24

Rapports de recherche

[38] E. Dumonteil, P. Crochet, "Measuring the p_t dependence of the Upsilon’/Upsilon ratio with the ALICE muon spectrometer", CERN-Note ALICE-INT-2005-002 (2005) [39] A.Khanzadeev et al., "Complementary approach to Beam Test Analysis", CERN-Note ALICE-INT-2003048 (2003) [40] K. Boudjemline et al., "Results of Slat CPC Prototype Test for ALICE Dimuon Spectrometer", CERNNote ALICE INT-2002-023 (2002) [41] S. Salasca et al., "Summary of the Cooling Calculations for Stations 345", CERN-Note ALICE INT-2001042 (2001) [42] E. Dumonteil et al., "Neutron detection efficiency in CLAS", CLAS-Note 01-006 (2001) Auteur ou co-auteur d’environ 10 rapports internes CEA portants principalement sur le code TRIPOLI-4

E. Dumonteil

Université Paris-Sud

Chapitre

3

Modélisation stochastique du transport des particules neutres Sommaire 3.1 3.2

Hypothèses principales et modélisation . . . . . . . . . . . . . . . . . . . . . . . . . Mouvement Brownien, équation de la diffusion et équation de Feynman-Kac . . 3.2.1 De l’équation de champ moyen au propagateur . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Généralisation des équations de champ moyen aux équations de Feynman-Kac pour la probabilité . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Equation de Feynman-Kac pour un processus continu de transport . . . . . . . . 3.3.1 Equation de champ moyen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Généralisation des équations de champ moyen aux équations de Feynman-Kac pour la probabilité . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Prise en compte du branchement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Equation pour les moments et polynômes de Bell . . . . . . . . . . . . . . . . . . . 3.6 Traitement des processus discrets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Résultats théoriques principaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 Solution de l’équation de Boltzmann linéaire stationnaire en dimension 2 . . . . . . . 3.7.2 Généralisation aux processus branchants de la formule de Cauchy pour les longueurs de corde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.3 La loi de l’arcsinus de Lévy avec absorption . . . . . . . . . . . . . . . . . . . . . . . . 3.7.4 Caractérisation de l’enveloppe convexe du déclenchement d’une épidémie . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

26 27 27 28 31 31 31 34 36 36 39 39 41 43 45 50

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 26

C

e chapitre vise à introduire les principaux résultats des travaux menés par Andrea Zoia, Alain Mazzolo et moi-même, sur la modélisation du transport des particules neutres à l’aide des marches aléatoires [1, 2, 3, 4, 5, 6, 7, 8, 9]. On présentera dans un premier temps le formalisme adapté à l’étude de la neutronique que nous avons développé (équation de Feynman-Kac discrète pour les processus branchants [7]) puis nos principaux résultats seront passés en revue, en ouvrant sur des applications à la neutronique [2, 3, 4], ou à d’autres domaines comme la théorie des jeux [5] ou la biologie [9].

3.1

Hypothèses principales et modélisation

La neutronique peut être définie comme l’étude du transport de populations de neutrons dans la matière. Comme la charge électrique du neutron est nulle, et que l’on néglige les interactions neutrons-neutrons (de sorte que les neutrons n’interagissent qu’avec le milieu), la distribution de la longueur qu’il parcourt entre deux collisions dans un milieu homogène suit une loi exponentielle. Ainsi, la probabilité p(l) qu’un neutron fasse une collision à la distance l est donnée par la loi p(l) = Σt e−Σt l , où Σt est la section efficace macroscopique totale du milieu traversé, grandeur proportionelle à la probabilité d’interaction avec le milieu traversé par unité de longueur (dont on montre qu’elle est égale à l’inverse du libre parcours moyen en calculant le premier moment de p). Lorsque la particule fait une collision avec un noyau du milieu traversé, elle est réorientée aléatoirement. La trajectoire de cette particule est donc décrite par une marche aléatoire. L’étude des marches aléatoires a débuté en 1905 suite à un courrier dans la rubrique des lecteurs de la revue Nature [10] où Karl Pearson introduisait le problème de la marche aléatoire dans le plan d’une personne effectuant des parcours de distance fixée entre des réorientations au hasard (on connait cette marche sous le nom de marche de l’ivrogne). Ainsi, parallèlement à la mise en place de la description quantitative de la théorie du mouvement Brownien par Albert Einstein, les marches aléatoires ont été étudiées tout d’abord sur des réseaux. Par la suite, l’étude des propriétés de ce type de marche a été généralisée à des espaces continus (la direction après réorientation est continue) puis à d’autres espaces que le plan, et enfin pour d’autres distributions possibles de distance entre les réorientations (on parle de collisions pour les particules) [11, 12]. Dans notre cas, et de manière très générale, on modélisera le transport des neutrons en introduisant les hypothèses suivantes : • on considère un milieu M infini, homogène et isotrope, dont la dimension d n’est pas a priori fixée. • un neutron est une particule qui réalise une marche aléatoire dans ce milieu, composée successivement de parcours en ligne droite entre deux collisions, et de réorientations aux lieux de collision. Ce type de marche s’appelle un vol exponentiel (“exponential jumps” en anglais). Cette marche est donc caractérisée par son noyau de transport T (r~0 , ω ~ |~r, ω ~ ) (représentant la probabilité qu’une particule se trouvant en ~r avec une direction ω ~ soit transportée en r~0 en conservant sa direction) et par son noyau de collision C(~r, ω~0 |~r, ω ~) (représentant la probabilité qu’une particule diffusée en ~r voit sa direction modifiée pour passer de ω ~ à ω~0 en conservant sa position). Nous garderons ces notations par la suite et, si besoin est, nous pourrons donner une forme particulière à ces fonctions. Par exemple, pour les vols exponentiels, on aura   ~0    ˆr  0 0 ~ ~ (3.1) T (r , ω ~ |~r, ω ~ ) = Σt (r ) exp − Σt (s)ds ,     ~ r

expression qui se simplifie lorsque le milieu est caractérisé par une section efficace macroscopique uniforme Σt pour prendre la forme suivante o n T (r~0 , ω ~ |~r, ω ~ ) = T ( r~0 − ~r ) = Σt exp −Σt r~0 − ~r . (3.2) De même, si les particules sont émises de manière isotrope en sortie de collision, on peut simplifier la forme du noyau de collision qui s’écrit 1 , (3.3) C(~r, ω~0 |~r, ω ~) = ωd



où ωd est la surface d’une sphère unitaire en dimension d (la sphère est de dimension d − 1) et s’écrit de 2π d/2 la manière suivante ωd = Γ(d/2) , avec Γ la fonction gamma. la vitesse des neutrons est constante et notée v.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 27 •



on se donne un volume V de détection dans M éventuellement traversé par la marche aléatoire. En toute généralité, ce volume sera traité comme un volume de l’espace des phases, mais dans les applications on se contentera de faire des mesures dans l’espace physique, comme représenté Figure 3.1. on s’intéressera par la suite à l’étude de la contribution de la particule au “détecteur” V, en mesurant la longueur parcourue par la particule dans le détecteur, ou en recensant le nombre discret de collisions faites par la particule dans le détecteur. En effet, nous verrons dans la section 3.7 que de manière assez surprenante, de nombreux problèmes peuvent être formulés sous cette forme. Comme la longueur parcourue dans le détecteur ou le temps de résidence sont des grandeurs stochastiques, on les caractérisera par leur distribution de probabilité. Ainsi p(lV (t)) est la distribution de probabilité de la longueur du parcours de la particule dans le détecteur, en l’observant jusqu’au temps t. On a alors ˆ

t

1V (~r(τ ), ω ~ (τ ))vdτ,

lV (t) = 0

où 1V est la fonction indicatrice du volume V et où v est la vitesse de la particule (cette vitesse est constante et identique pour toutes les particules). De même, le temps de résidence dans V d’un neutron est noté TV et est donné par ˆ t TV (t) = 1V (~r(τ ), ω ~ (τ ))dτ. 0

Comme la vitesse v des neutrons est constante, TV et lV sont trivialement liés, et on se concentrera dans la suite sur lV . Dans la suite du présent chapitre on cherchera dans un premier temps, et à des fins pédagogiques, à construire une équation pour ce type de grandeur dans le cas du mouvement Brownien à une dimension. On s’intéressera dans un deuxième temps aux processus de transport en dimension quelconque, et donc à l’équation de Boltzmann linéaire, en étudiant les marches aléatoires de type “vols exponentiels”. On montrera que cette équation peut être généralisée à une classe de super-processus : les vols branchants, décrivant au mieux le transport des neutrons dans des milieux fissiles. Enfin, on présentera une généralisation de ces résultats aux marches aléatoires discrètes caractérisées uniquement par la donnée de leur noyau de transport T et de collision C.

3.2 3.2.1

Mouvement Brownien, équation de la diffusion et équation de Feynman-Kac De l’équation de champ moyen au propagateur

Considérons un grand nombre de particules N effectuant un mouvement Brownien en 1 dimension, en commençant à l’origine. Le nombre moyen de particules N (x, t)dx en x, à dx près, au temps t est donné par l’équation de la diffusion ∂2N ∂N =D 2, ∂t ∂x

(3.4)

où D est le coefficient de diffusion propre au milieu M. Une analyse de Fourier permet d’en extraire le noyau G(∆x, 4t) donné par la relation (4x)2 1 G(∆x, 4t) = √ exp− 4D4t . 2 πD4t

Ce noyau représente donc le propagateur (i.e. la fonction de Green) d’une particule Brownienne, c’est-à-dire d’une particule effectuant une marche aléatoire avec des sauts infiniments petits et obéissant à cette loi de probabilité. Un exemple de mouvement Brownien est donné figure 3.1.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 28

Figure 3.1 Simulation d’un mouvement brownien uni-dimensionnel (axe x) représenté en fonction du temps t. les unités sont arbitraires. Au début du XXème siècle, Einstein et Smoluchowski démontrent à partir de ce dernier résultat que la probabilité qu’une particule Brownienne partant de l’origine se trouve entre a1 et b1 au temps t1 , entre a2 et b2 au temps t2 , ..., entre an et bn au temps tn est donnée par ˆb1

ˆbn ...

a1

an

G(x1 , t1 |0, 0)G(x2 , t2 |x1 , t1 )...G(xn , tn |xn−1 , tn−1 )dx1 dx2 ...dxn ,

(3.5)

en utilisant la notation G(∆x = xf − xi , 4t = tf − ti ) = G(xf , tf |xi , ti ).

3.2.2

Généralisation des équations de champ moyen aux équations de FeynmanKac pour la probabilité

Suivant une démarche initiée en 1949 [13, 14], Mark Kac rédige un article fondateur en 1957 en s’appuyant sur ce résultat et sur le formalisme des intégrales fonctionnelles (dites aussi intégrales de chemin) de Feynman pour calculer la distribution de probabilité p(lV (t)) définie précédemment, dans le cas de cette particule Brownienne. L’idée consiste à utiliser la transformée de Laplace de la distribution de probabilité, que l’on note ˆ∞ Qt (s|x0 ) = 0

p(lV (t)|x0 ) exp {−slV } dlV ,

où s est la variable transformée de lV , et où la dépendance de p(lV (t)) par rapport au point de départ x0 de la marche aléatoire a été rendue explicite. Pour alléger la notation on introduit h ...| x0 i, la moyenne sur les toutes les trajectoires possibles démarrant en x0 , ce qui permet de noter cette transformée  Qt (s|x0 ) = h exp {−slV }| x0 i =

E. Dumonteil

 exp −s

ˆ 0

t

  1V (x(τ ))vdτ x0 .

(3.6)

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 29 Kac démontre alors que cette moyenne sur les trajectoires peut justement se réécrire, à l’aide de 3.5 , de la manière suivante  ˆ t  ˆ∞ ˆ∞ 1V (x(τ ))vdτ G(x1 , t1 |x0 , t0 )G(x2 , t2 |x1 , t1 )...G(xn , tn |xn−1 , tn−1 )dx1 dx2 ...dxn . Qt (s|x0 ) = ... exp −s 0

0

0

(3.7) Cette notation rend extrêmement intuitive l’intégrale sur tous les chemins possibles puisque l’on considère la probabilité que la particule aille de x0 à x1 en un temps t1 − t0 puis de x1 à x2 en un temps t2 − t1 etc., et que l’on intègre sur toutes les positions possiblesn pour x0 , x1 , x2 ,... oChaque trajectoire est alors pondérée ´t par sa contribution au détecteur V, c’est-à-dire exp −s 0 1V (x(τ ))vdτ (plus exactement ce terme génère les

différents moments de la probabilité p(lV (t)) initiale). Dans la suite, on notera donc toujours l’intégrale sur tous les chemins possibles démarrants en x0 par le sigle h ...| x0 i. Kac cherche dans un second temps à formuler une équation pour cette transformée. Le raisonnement, exposé par exemple dans [17], consiste à considérer la transformée à un temps d’observation t + dt * ( ˆ ) + t+dt Qt+dt (s|x0 ) = exp −s 1V (x(τ ))vdτ x0 . 0 On remarque alors que la trajectoire aléatoire se scinde en deux parties : la première partie de la trajectoire va de 0 à dt (et donc de x0 à x0 + dx0 ) et la seconde partie de dt à t + dt. On cherche alors à lier l’équation pour la transformée considérée entre 0 et t et la transformée considérée entre dt et t + dt, comme représenté sur la figure 3.2.

Figure 3.2 Simulation d’un mouvement brownien en deux dimensions contribuant à un détecteur V : l’équation pour la probabilité est dérivée en considérant les différences entre les deux représentations présentées ci-dessus. Pour ce faire, on remarque la dernière équation peut se réécrire * ( ˆ ) ( ˆ dt Qt+dt (s|x0 ) = exp −s 1V (x(τ ))vdτ exp −s 0

E. Dumonteil

t+dt

dt

) + 1V (x(τ ))vdτ x0 .

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 30 Or le premier terme exponentiel ne dépend que de x0 et n’est donc pas aléatoire. On a donc * Qt+dt (s|x0 ) = exp {−s1V (x0 )vdt}

ˆ

( exp −s

t+dt

dt

) + 1V (x(τ ))vdτ x0 .

Le second terme exponentiel de l’expression de Qt+dt peut être vu comme commencant en x0 + dx0 et ayant une “longueur” t Qt+dt (s|x0 ) = exp {−s1V (x0 )vdt}



 exp −s

ˆ 0

t

  1V (x(τ ))vdτ x0 + dx0 .

Comme la position de départ comporte une partie aléatoire dx0 , le second terme ne peut simplement se réécrire Qt (s|x0 + dx0 ) mais doit être moyenné sur cette incertitude sur la position de départ Qt+dt (s|x0 ) = exp {−s1V (x0 )vdt} hQt (s|x0 + dx0 )idx0 . A l’aide d’un développement limité au premier ordre en temps et au second ordre en espace on a   ∂Qt (s|x0 ) ∂ 2 Qt (s|x0 ) 2 Qt+dt (s|x0 ) = (1 − s1V (x0 )vdt + ...) Qt (s|x0 ) + hdx0 i + dx0 + ... , ∂x0 ∂x20 et en remarquant que les collisions sont isotropes, et que par conséquent hdx0 i = 0, il vient  ∂ 2 Qt (s|x0 ) 2 Qt+dt (s|x0 ) = (1 − s1V (x0 )vdt + ...) Qt (s|x0 ) + dx0 + ... , ∂x20 

Qt+dt (s|x0 ) − Qt (s|x0 ) ∂ 2 Qt (s|x0 ) dx20 = − s1V (x0 )vQt (s|x0 ). dt ∂x20 dt hdx2 i Et par définition du coefficient de diffusion D = dt on obtient finalement l’équation de Feynman-Kac “backward”, également connue sous le nom d’équation de Fokker-Planck ∂ 2 Qt (s|x0 ) ∂Qt (s|x0 ) =D − s1V (x0 )vQt (s|x0 ). ∂t ∂x20

(3.8)

Elle est assortie de sa condition initiale Q0 (s|x0 ) = 1, donnée par Eq.3.6 (qui traduit que la probabilité de trouver la particule en x0 au temps t = 0 est de 1). Le terme “backward” fait référence au fait que l’équation porte sur une fonction de la variable de départ x0 , et non pas la variable d’arrivée comme il est de coutume en physique (en fait dans le papier original de Kac, c’est l’équation “forward” qui est dérivée). Bien que proche, en apparence, de l’équation de la diffusion, cette équation est très générale. En effet, l’équation Eq.3.4 est une équation pour la moyenne de la concentration, quand l’équation Eq.3.8 donne accès à toute la distribution de probabilité ! De plus, cette équation fait apparaître le terme s1V (x0 )vQ(s|x0 ) qui permet de rendre compte de l’histoire passée de la particule du temps 0 au temps t. Enfin, on remarquera que toute la démonstration faite à partir de la fonction indicatrice 1V (x0 ) se généralise à un potentiel quelconque V (x0 ). Ainsi, l’équation 3.8 donne accès, une fois résolue, à toute la distribution de probabilité d’une grandeur d’intérêt, pour la marche aléatoire d’une particule Brownienne dans un potentiel V . On notera également au passage qu’à une rotation de Wick près t → it, l’équation de Feynman-Kac est exactement l’équation de Schrödinger, qui donne la probabilité de trouver une particule dans une position donnée, à un moment donné. Nous pouvons à présent résumer notre démarche. Nous sommes partis d’une équation de champ moyen (i.e. donnant accès à la moyenne d’un champ) : l’équation de la diffusion. On a cherché à associer à cette équation sa fonction de Green, en l’occurence le propagateur Gaussien associé au mouvement Brownien. Ce propagateur a permis d’établir une équation pour la transformée de Laplace de la distribution de probabilité du temps de résidence ou de la longueur parcourue dans un volume détecteur V par une particule Brownienne. Nous allons dans la suite tenter de construire la même équation pour le transport des neutrons, dont les hypothèses ont été précisées Section 3.1.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 31

3.3 3.3.1

Equation de Feynman-Kac pour un processus continu de transport Equation de champ moyen

Considérons à présent un nombre N0 de particules effectuant des vols exponentiels commençant en (r~0 , ω~0 ) au temps t = 0. Si N (~r, ω ~ , t|r~0 , ω~0 ) est la densité de particules en (~r, ω ~ , t), on a alors N (~r, ω ~ , t, r~0 , ω~0 )d~rd~ ω particules situées en (~r, ω ~ ) au temps t, à d~r et d~ ω près. Afin d’établir l’équation de Boltzmann linéaire [18], on effectue un bilan dans un volume élémentaire de taille d~rd~ ω centré en (~r, ω ~ ) sur la variation en temps de ce nombre de particules. La variation en temps du nombre de particules dans ce volume est donnée par le terme ∂ N (~r, ω ~ , t|r~0 , ω~0 ). ∂t Le terme de source q est lié au départ de la marche aléatoire et s’écrit donc q = N (~r, ω ~ , 0|r~0 , ω~0 ) = N0 δ(~r − r~0 )δ(w ~ − w~0 ). Les particules se trouvant d’ores et déjà dans le volume élémentaire (de l’espace des phases) et effectuant une collision quittent ce volume élémentaire (puiqu’une collision modifie les propriétés cinématiques de la particule). Le recensement de ces particules se fait par conséquent en multipliant le taux de réaction total vΣt et la densité N (~r, ω ~ , t|r~0 , ω~0 ). De même, deux termes supplémentaires doivent apparaître. D’une part les particules quittant ~ ~r N (~r, ω le volume élémentaire par transport sont caractérisées par le terme v~ ω .∇ ~ , t|r~0 , ω~0 ) qui doit être lui aussi retranché du bilan. A l’inverse, certaines particules se trouvaient à l’extérieur du volume de l’espace des phases, et se déplacent à l’intérieur à la suite d’une collision (une diffusion). Ces particules se recensent à l’aide ´ ~ 0 vΣt C(~r, ω de l’opérateur de collision : dw ~ |~r, ω ~ 0 )N (~r, ω~0 , t|r~0 , ω~0 ). Cette expression doit être multipliée par le nombre moyen ν1 de neutrons émis en sortie d’une collision. L’équation finale, dite équation du transport, s’écrit finalement ˆ ∂ ~ ~r N + vΣt N = ν1 dw ~ 0 vΣt C(~r, ω N + v~ ω .∇ ~ |~r, ω~0 )N (~r, ω ~ 0 , t|r~0 , ω~0 ) + q, (3.9) ∂t où l’on a contracté l’écriture de la densité N (~r, ω ~ , t|r~0 , ω~0 ). C’est une équation de type Boltzmann linéaire pour la densité des particules dans l’espace des phases (on a allégé la notation). On notera également qu’elle peut également être vue comme une équation de conservation de la probabilité, et dans ce contexte elle porte aussi le nom d’équation de Chapman-Kolmogorov. En neutronique, il est important de pouvoir calculer certains taux de réaction induits par le transport des neutrons (dose reçue par un patient, taux de fission mesuré par un détecteur etc). Comme un taux de réaction est le produit d’une section efficace macroscopique par un flux, on utilise plus naturellement l’équation de Boltzmann pour le flux des particules ψ = N v, obtenue en multipliant et en divisant 3.9 par v ˆ 1 ∂ ~ ~r ψ + Σt ψ = ν1 dw ~ 0 Σt C(~r, ω ψ+ω ~ .∇ ~ |~r, ω~0 )ψ(~r, ω ~ 0 , t|r~0 , ω~0 ) + q. (3.10) v ∂t

3.3.2

Généralisation des équations de champ moyen aux équations de FeynmanKac pour la probabilité

Dans le cas d’un processus de transport, plusieurs propagateurs peuvent être définis. Le noyau de transport, donnant la distance séparant deux collisions, peut être vu comme une fonction de Green. Si l’on s’intéresse aux processus discrets (ici, les collisions), on peut être amené à considérer la probabilité G(~r, n|~0, 0) de se trouver en ~r à la n-ième collision, sachant que la marche aléatoire a démarré en ~0 à n = 0. On peut également s’intéresser à G(~r, t|~0, 0), c’est-à-dire à la probabilité que la particule passe en ~r au temps t après avoir été émise en ~0 au temps t = 0. Ces différentes grandeurs dépendent de la dimension d de l’espace [1]. On cherche par conséquent à dériver directement l’équation pour la probabilité, en construisant la génératrice des moments Qt (s|~r, ω ~ ) caractérisant à présent le processus de transport. Ce processus est bien Markovien pour les variables (~r, ω ~ ), ce qui permet de diviser l’intervalle en temps t + dt en deux sous-intervalles (de 0 à dt et de dt à t + dt) en s’assurant qu’au temps dt la particule n’a plus de “mémoire” du passé. Dans cette sous-section, on ne considére que trois événements possibles dans cet intervalle de temps de 0 à dt : • il ne se passe rien, avec probabilité p1 = 1 − Σt vdt, et la particule poursuit son parcours en ligne droite. On a alors d~r = v wdt ~

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 32 la particule subit une diffusion avec probabilité p2 = Σs vdt : sa position est alors inchangée mais sa direction est modifiée • la particule disparaît avec probabilité p3 = Σa vdt. La condition p1 + p2 + p3 = 1 impose alors Σs + Σa = Σt La figure 3.3 donne le résultat de la simulation d’un vol exponentiel ainsi défini. •

z

O

y

x

. Figure 3.3 Simulation du vol exponentiel en 3 dimensions d’une particule partant de ~r0 avec la direction w ~ 0 , avec Σs = 0.8Σt et Σa = 0.2Σt . Le volume détecteur V représenté ici ne possède qu’une composante spatiale. Comme pour la sous-section 3.2.2 on écrit l’équation pour la génératrice en partant de sa définition   ˆ t   Qt (s|r~0 , ω~0 ) = h exp {−slV }| r~0 , ω~0 i = exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 , (3.11) 0

ce qui permet d’assurer, à t + dt, que * Qt+dt (s|r~0 , ω~0 ) = * Qt+dt (s|r~0 , ω~0 ) =

(

ˆ

exp −s

ˆ

( exp −s

) + 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 , ) ( ˆ

t+dt

0

dt

t+dt

1V (~r(τ ), ω ~ (τ ))vdτ exp −s

0

dt

) + 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 .

Cette expression de la génératrice peut se décomposer sur la base des trois événements exclusifs évoqués ci-dessus * Qt+dt (s|r~0 , ω~0 )

=

X

pi

i

ˆ

(

exp −s

( exp −s

ˆ

t+dt

dt

0

)

dt

1V (~r(τ ), ω ~ (τ ))vdτ

) +# 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 ,

(3.12)

pi

où le signe .]pi indique que l’expression est conditionnée à l’événement associé à la probabilité pi . Par conséquent, • si aucun événement n’advient, la contribution à la génératrice est un terme qui s’écrit * ( ˆ ) ( ˆ ) +# dt t+dt (1 − Σt vdt) exp −s 1V (~r(τ ), ω ~ (τ ))vdτ exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 . 0 dt p1

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 33 Le premier terme exponentiel dans l’intégrale stochastique n’est, en fait, pas aléatoire car de 0 à dt la particule suit une trajectoire en ligne droite (événement p1 ). La contribution s’écrit alors +# * ( ˆ ) t+dt . (1 − Σt vdt) exp {−s1V (r~0 , ω~0 )vdt} exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 dt p1

De plus, comme il est équivalent de considérer une trajectoire de dt à t + dt en moyennant sur (r~0 , ω~0 ) ou de la considérer de 0 à t mais en la moyennant sur (r~0 + dr~0 , ω~0 ), on peut écrire la contribution en fonction de Qt (s|r~0 + dr~0 , ω~0 ) (1 − Σt vdt) exp {−s1V (r~0 , ω~0 )vdt} Qt (s|r~0 + dr~0 , ω~0 )

(le trajet en ligne droite assure que la direction est inchangée de 0 à dt). Un développement limité au premier ordre de l’exponentielle en dt ainsi qu’un développement en série de Taylor de la génératrice en r~0 donnent alors   ~ r~ Qt dt . (1 − Σt vdt) (1 − s1V (r~0 , ω~0 )vdt) Qt (s|r~0 , ω~0 ) + v ω~0 .∇ 0 En ne gardant que les premiers termes en dt, on simplifie cette expression ~ r~ Qt dt − s1V (r~0 , ω~0 )vdtQt (s|r~0 , ω~0 ) − Σt vdtQt (s|r~0 , ω~0 ). Qt (s|r~0 , ω~0 ) + v ω~0 .∇ 0





si la particule disparaît entre 0 et dt, la longueur parcourue dans le détecteur est nulle, et la contribution à la génératrice est Σa vdt. si l’on considère un événement de type collision, en reprenant ce dernier raisonnement, on obtient une contribution à la génératrice de la forme +# +* ( ˆ ) * ( ˆ ) t+dt dt . exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 Σs vdt exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 dt 0 p2

Comme il n’y a pas de transport pour cet événement, la contribution de la première intégrale à la distance parcourue est nulle (l’exponentielle vaut 1). Le traitement de la seconde intégrale de chemin est plus délicat, puisque pour passer d’une intégration entre dt et t + dt à une intégration entre 0 et t, il faut moyenner sur toutes les directions possibles en sortie de la collision (la direction de la particule a été modifiée de ω~0 à ω~00 ). Ainsi, la notation h.iω~0 désigne dans la suite une moyenne sur toutes les directions ω~00 (à ne pas 0 confondre avec l’intégrale de chemins). Si l’on moyenne sur toutes les directions ω~0 sachant que la direction 0

de départ était ω~0 , on peut utiliser la notation h.|ω~0 iω~0 . Le théorème de Dynkin permet alors d’écrire cette 0 moyenne à l’aide du noyau de collision C D E n o ˆ Qt (s|r~0 , ω~00 ) r~0 , ω~0 ~0 = C † Qt (s|r~0 , ω~00 ) = C † (r~0 , ω~0 |r~0 , ω~00 )Qt (s|r~0 , ω~00 )dω~00 , (3.13) ω0

où C † est la densité adjointe à la densité C (attention, la notation utilisant le noyau de collision adjoint masque la dépendance aux variables de départ). La somme des trois termes permet d’assurer la relation Qt+dt (s|r~0 , ω~0 )

=

~ r~ Qt dt Σa vdt + Qt (s|r~0 , ω~0 ) + v ω~0 .∇ 0

n o −s1V (r~0 , ω~0 )vdtQt (s|r~0 , ω~0 ) − Σt vdtQt (s|r~0 , ω~0 ) + Σs vdtC † Qt (s|r~0 , ω~00 ) .

Si l’on fait apparaître l’expression de la dérivée temporelle de la génératrice, et en rendant implicite ses dépendances aux variables, on aboutit finalement à l’équation suivante 1 ∂ ~ r~ Qt − s1V (r~0 , ω~0 )Qt − Σt Qt + Σs C † {Qt } + Σa . Qt = ω~0 .∇ 0 v ∂t

(3.14)

Cette équation est proche de l’équation 3.8. C’est une équation pour la probabilité qui généralise l’équation de Boltzmann linéaire 3.9, au sens où elle permet de construire tous les moments de la distribution de probabilité des longueurs parcourues dans un domaine d’intérêt, et ce pour un processus de transport. Cela se fait au prix ~ r~ {.} − Σt {.} + Σs C † {.} + Σa , qui se substitue de l’introduction d’un opérateur de transport de la forme ω~0 .∇ 0 E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 34 2

∂ à l’opérateur de diffusion D ∂x 2 . On note cependant que cet opérateur de transport est de type “backward” à 0 la différence de l’opérateur de l’équation 3.9 qui est de type “forward”. Cette équation est importante pour le transport des particules neutres, car elle ne fait pas l’approximation de la diffusion (le mouvement Brownien constitue la limite en loi de toutes les marches aléatoires à variance finie pour lesquelles la taille des sauts tend vers 0). Il est possible à présent de faire un pas supplémentaire en direction d’une équation pour la neutronique, en cherchant à généraliser cette équation aux proccessus branchants qui prennent en considération une distribution de probabilité du nombre de particules produites en sortie d’une réaction, et qui permettent donc au mieux de décrire la fission nucléaire.

3.4

Prise en compte du branchement

Considérons donc à présent qu’à chaque collision d’un neutron avec un noyau, k neutrons fils sont produits avec probabilité bk (voir figure 3.4).

z

O

y

x

Figure 3.4 simulation du vol exponentiel branchant en 3 dimensions d’une particule partant de ~r0 avec la direction w ~ 0 , avec b0 = 0.1, b1 = 0.8 et b2 = 0.1. Le volume détecteur V représenté ici ne possède qu’une composante spatiale.

Ainsi, une absorption correspond à l’événement 0 fils produit avec probabilité b0 = Σa /Σt et une diffusion simple à l’événement 1 fils produit avec probabilité b1 = Σs /Σt . Plus généralement, le taux de production d’un

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 35 événement à k fils vaut bk Σt vdt. On peut donc généraliser l’équation 3.12 de la manière suivante +# ) ) ( ˆ * ( ˆ t+dt dt 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 1V (~r(τ ), ω ~ (τ ))vdτ exp −s Qt+dt (s|r~0 , ω~0 ) = (1 − Σt vdt) exp −s dt 0 p1 +# * ( ) ( ) ˆ dt ˆ t+dt ∞ X + bk Σt vdt exp −s 1V (~r(τ ), ω ~ (τ ))vdτ exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 0 dt k=0

,

bk

(3.15) Qt+dt (s|r~0 , ω~0 )

=

(1 − Σt vdt) exp {−s1V (r~0 , ω~0 )vdt} Qt (s|r~0 + dr~0 , ω~0 ) +# * ( ˆ ) ∞ t+dt X + bk Σt vdt exp −s 1V (~r(τ ), ω ~ (τ ))vdτ r~0 , ω~0 dt k=0

,

(3.16)

bk

où on rappelle que p1 = 1 − Σt vdt est la probabilité qu’il ne se passe rien. Le terme de gauche se traite de la même manière qu’en sous-section précédente. Afin de simplifier le terme de droite, et pour comprendre l’action des événements à particules multiples en voie de sortie sur la génératrice, on peut raisonner directement sur la probabilité d’effectuer un trajet de longueur lV : si aucune particule n’est émise (probabilité b0 ), la trajectoire ne se poursuit pas et on n’ajoute pas de contribution supplémentaire à lV . Dans le cas ou k ≥ 1 particules sont émises, la probabilité que les trajets de chacune des particules s’additionnent pour donner précisément lV est égale à la convolution de la probabilité que la première particule fasse un trajet l1 , que la deuxième particule fasse un trajet l2 , etc, et que la kième fasse un trajet lV − l1 − l2 − .... Dans l’espace transformé (celui des génératrices), la convolution de probabilité devient un simple produit. Ainsi, la contribution à l’équation précédente dans le cas d’une collision à 0 fils vaut b0 Σt vdt, car la particule disparaît en dt et par conséquent l’argument de l’exponentielle conditionnée par b0 s’annule. Aussi, pour les mêmes raisons qu’en section précédente la contribution à un fils doit être moyennée sur toutes les directions de sortie possibles et s’écrit D E b1 Σt vdt Qt (s|r~0 , ω~00 ) r~0 , ω~0 ~0 ω0

(la modification de la direction est indiquée par le signe ’). Pour une contribution à deux fils, on moyenne de même sur toutes les directions de sortie possibles le produit des génératrices (convolution des probabilités) de chacune des particules émises en dt D E 0 )Q (s|r~ , ω~ 0 b2 Σt vdt Qt (s|r~0 , ω~0,1 ~0 ~0 ~0 , t 0 0,2 ) r~0 , ω w0,1 w0,2

par conséquent on a une moyenne portant sur les directions des 2 particules (le deuxième indice de la direction correspond à l’identifiant de la particule). Par indépendance des directions des particules en sortie de collision, cette contribution à deux fils se met sous la forme D E2 b2 Σt vdt Qt (s|r~0 , ω~00 ) r~0 , ω~0 ~0 . w0

On peutPrépèter ce raisonnement pour un nombre arbitraire k de particules en voie de sortie. Si l’on introduit ∞ G(z) = k=0 bk z k , la génératrice des probabilités du nombre de fils en sortie, et que l’on utilise 3.13, l’ensemble des contributions se somme à n o G(C † Qt (s|r~0 , ω~00 ) ), et permet d’établir la généralisation de l’équation de Feynman-Kac pour les vols exponentiels branchants (et donc pour l’équation des probabilités associée à l’équation de Boltzmann linéaire critique)  1 ∂ ~ r~ Qt − s1V (r~0 , ω~0 )Qt − Σt Qt + Σt G C † {Qt } . Qt = ω~0 .∇ 0 v ∂t

(3.17)

La résolution de cette équation demande comme condition initiale Q0 (s|r~0 , ω~0 ) = 1, afin de traduire la présence certaine de la particule en (r~0 , ω~0 ) à t = 0.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 36

3.5

Equation pour les moments et polynômes de Bell

Pour les processus branchants, l’équation 3.17 est une équation aux dérivées partielles non-linéaire, car la fonction G est un polynôme de degré k > 1. De plus, elle donne accès à la fonction génératrice de la probabilité d’intérêt. Ainsi, le calcul de la probabilité d’intérêt suppose de résoudre une équation différentielle non-linéaire, puis de calculer la transformée de Laplace de la solution. Dans la pratique, ces considérations rendent le calcul direct de la distribution de probabilité quasiment impossible. En revanche, il est possible de construire une équation linéaire pour les moments de cette distribution de probabilité [6, 7, 8], ce qui est d’autant plus pertinent que l’on se satisfait souvent de la connaissance des deux ou trois premiers moments. Par exemple, les codes Monte-Carlo de transport des particules permettent souvent de calculer la moyenne, la variance et parfois la variance de la variance de la distribution des longueurs parcourues par une particule dans un volume donné. Par définition de Qt , le moment d’ordre m de la distribution de probabilité s’écrit hlVm it (r~0 , ω~0 ) = (−1)m

∂m Qt (s|r~0 , ω~0 )|s=0 . ∂sm

En dérivant une fois l’équation 3.17 on obtient la longueur moyenne parcourue





 1 ∂ 1 ~ r~ lV1 + 1V (r~0 , ω~0 ) lV0 − Σt lV1 + ν1 Σt C † lV1 = ω~0 .∇ l , 0 t t t t v ∂t V t

(3.18)

et en dérivant deux fois, on accède aux fluctuations





 1 ∂ 2 ~ r~ lV2 + 21V (r~0 , ω~0 ) lV1 − Σt lV2 + ν2 Σt C † lV2 l = ω~0 .∇ . 0 t t t t v ∂t V t

(3.19)

Plus généralement, en dérivant m fois l’équation 3.17 et en utilisant la formule de Faà di Bruno-Arbogast on accède à l’équation linéaire pour les moments de la longueur via m X

   1 ∂ m ~ r~ hlVm i + m1V (r~0 , ω~0 ) lm−1 − Σt hlVm i + Σt hlV it = ω~0 .∇ νj Bm,j C † lVi t , 0 V t t t v ∂t j=1

(3.20)

où les termes Bm,j sont les polynômes de Bell, et νj = hk(k − 1)...(k − j + 1)i sont les moments factoriels descendants du nombre de fils k, avec ν0 = 1. Cette expression peut également se réécrire m X

   1 ∂ m hlV it = L† hlVm it + m1V (r~0 , ω~0 ) lVm−1 t + Σt νj Bm,j C † lVi t , v ∂t j=2

(3.21)

~ r~ − Σt + Σt ν1 C † est l’opéateur de transport adjoint. où L† = ω~0 .∇ 0

3.6

Traitement des processus discrets

A la différence du mouvement Brownien, les vols exponentiels sont constitués d’un ensemble fini de collisions (la distance entre deux collisions successives est non nulle). Ces collisions représentent dans notre cas les interactions de la particule avec le milieu, et leur recensement est par conséquent d’un intérêt primordial. C’est pourquoi les codes Monte-Carlo de transport des particules proposent bien souvent deux types d’estimateurs des quantités d’intérêt : l’un fondé sur la mesure de la longueur parcourue par une particule et ses descendants dans un volume V - c’est-à-dire la distribution de probabilité p(lV ) à laquelle nous nous sommes intéréssé jusqu’à présent - et l’autre utilisant le recensement du nombre discret de collisions faites par une particule et ses descendants dans un volume V. Dans cette section on cherchera donc à généraliser les résultats précédents dans le cas du recensement d’événements discrets [2, 7, 8]. Cette approche généralise l’approche Brownienne car elle ne requiert du processus étudié que d’être Markovien aux points de collision. Ainsi, un processus peut il être défini en toute généralité par son noyau de transport T et son noyau de collision C. Le comportement diffusif peut alors être recouvré dans la limite où la taille des sauts donnée par T tend vers 0 et où C est isotrope. On définit par conséquent le nombre nV (n) comme le nombre de collisions faites par une particule et ses descendants dans le volume V, quand on considère la marche aléatoire jusqu’à la “génération” n (la génération n joue le rôle d’un “temps”). Ce nombre s’écrit donc

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 37

nV (n) =

n X

1V (~ri , ω~i ),

i=0

où le couple (~ ri , ω~i ) désigne les positions (dans l’espace des phases) des neutrons aux points de collisions. Par convention, la source n’est pas prise en compte dans le comptage des collisions (les neutrons issus de la source sont transportés jusqu’au premier lieu de collision). La quantité nV (n) est une variable stochastique, à laquelle on peut associer sa distribution de probabilité notée p(nV (n)). Comme cette variable est discrète, on associe à sa distribution de probabilité une génératrice discrète de la forme * ( + ) n X Qn (u|r~0 , ω~0 ) = h exp {−unV }| r~0 , ω~0 i = exp −u 1V (~ri , ω~i ) r~0 , ω~0 , i=0

où u est la variable de Laplace transformée du nombre de collisions nV . Dans le cas des processus discrets, la source joue un rôle particulier car on observe le processus aux lieux de collision, et le premier lieu de collision (où le processus est donc Markovien) est en r~1 . Pour cette raison, on s’intéressera par la suite à la génératrice ˜ n (u|r~1 , ω~0 ) et qui est liée à Qn (u|r~0 , ω~0 ) par le noyau de transport au premier point de collision, que l’on note Q adjoint T † (l’action de l’opérateur de transport adjoint consiste à transporter les particules de r~1 à r~0 c’est-à-dire à revenir dans le passé) via ˆ ˜ n (u|r~1 , ω~0 )T † (r~0 , w~0 |r~1 , w~0 )dr~1 . Qn (u|r~0 , ω~0 ) = Q (3.22) ˜ n est la génératrice au premier point de collision, par définition Comme Q ) + * ( n X ˜ n (u|r~1 , ω~0 ) = h exp {−unV }| r~1 , ω~0 i = exp −u 1V (~ri , ωi−1 ~ ) r~1 , ω~0 , Q i=1

où les lieux de collisions sont paramétrés par le même décalage des indices de position et de direction, et où l’on considère que la position du premier lieu de collision est fixée. Le raisonnement est alors sensiblement similaire ˜ à la génération n + 1 et cette à celui des processus continus : on cherche à établir le lien entre la génératrice Q même génératrice à la génération n, en s’intéressant aux différents événements aléatoires au point r~1 de première collision. Comme précédemment, à ce point k particules sont créées avec probabilité bk . Par conséquent : ˜ • avec probabilité b0 , la particule disparaît (k = 0) et on a une contribution à la génératrice Qn+1 donnée par le terme +# ) * ( n+1 X = b0 exp {−u1V (r~1 , ω~0 )} , 1V (~ri , ωi−1 ~ ) r~1 , ω~0 b0 exp −u i=1



b0

car le premier lieu de collision n’est pas aléatoire ˜ n+1 donnée par avec probabilité b1 , la particule diffuse (k = 1) et on a une contribution à la génératrice Q le terme D n o Ei Pn+1 b1 exp −u i=1 1V (~ ri , ωi−1 ~ ) r~1 , ω~0 b1 D n o Ei Pn+1 = b1 exp {−u1V (r~1 , ω~0 )} exp −u i=2 1V (~ ri , ωi−1 ~ ) r~1 , ω~0 , b1

qui peut être transformé de la même manière que pour les processus continus en introduisant la moyenne sur les positions et directions au deuxième lieu de collision r~2 , w~1 en remarquant que o D n E Pn+1 exp −u i=2 1V (~ ri , ωi−1 ~ ) r~1 , ω~0 Pn = h exp {−u i=1 1V (ri+1 ~ , ω~i )}| r~1 , ω~0 i D E ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 = Q . r~2 ,w~1

E ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 Cette contribution vaut donc b1 exp {−u1V (r~1 , ω~0 )} Q D

r~2 ,w~1

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 38 •

avec probabilité b2 , la particule branche (k = 2) et Ei D n o Pn+1 b2 exp −u i=1 1V (~ ri , ωi−1 ~ ) r~1 , ω~0 b2 Ei o D n Pn+1 ri , ωi−1 ~ ) r~1 , ω~0 = b2 exp {−u1V (r~1 , ω~0 )} exp −u i=2 1V (~

b2

.

Dans ce cas, à l’issue de la première génération deux particules ont été émises, et le terme de droite dans cette dernière expression doit donc être interprété comme une intégrale sur tous les chemins possibles, et en l’occurence sur les chemins associés aux deux particules (conditionnement par b2 ). Comme dans le traitement des processus discrets branchants ce terme peut se réécrire comme un produit des génératrices à une particule et la dernière expression prend alors la forme E D ˜ n (u|r~2,1 , ω~1,1 )Q ˜ n (u|r~2,2 , ω~1,2 ) r~1 , ω~0 , b2 exp {−u1V (r~1 , ω~0 )} Q r2,1 ~ ,ω~ ~ ,ω~ 1,1 ,r2,2 1,2

où le premier indice du terme des vecteurs position et direction désigne la génération et le deuxième indice désigne l’identité de la particule. Par indépendance des particules en sortie de collision ce terme devient D E2 ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 b2 exp {−u1V (r~1 , ω~0 )} Q . r~2 ,ω~1

En généralisant le raisonnement ci-dessus, on a la relation suivante ˜ n+1 (u|r~1 , ω~0 ) Q

=

b0 exp {−u1V (r~1 , ω~0 )}

D E ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 +b1 exp {−u1V (r~1 , ω~0 )} Q r~2 ,w~1 D E2 ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 +b2 exp {−u1V (r~1 , ω~0 )} Q r~2 ,ω~1

+... qui se simplifie à l’aide de la définition de la génératrice G du processus branchant D E ˜ n+1 (u|r~1 , ω~0 ) = exp {−u1V (r~1 , ω~0 )} G Q ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 Q

(3.23)

Afin que cette équation porte sur les mêmes variables, on cherche à exprimer les monômes de la génératrice (de D Ek ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 ) en fonction de Qn (u|r~1 , ω~0 ). Pour ce faire, on utilise la définition la forme Q r~2 ,ω~1

D E ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 Q

r~2 ,ω~1

ˆ =

ˆ C † (r~1 , w~0 |r~1 , w~1 )

˜ n (u|r~2 , ω~1 )dr~2 dw~1 , T † (r~1 , w~1 |r~2 , w~1 )Q

où l’on remarque que le noyau de transport adjoint donne la probabilité d’aller de la position (r~2 , w~1 ) à la position (r~1 , w~1 ), et où le noyau de collision adjoint a une action similaire sur la direction puisque (r~1 , w~1 ) devient (r~1 , w~0 ). Ainsi l’action combinée de ces deux opérateurs permet de revenir “une collision en arrière” dans le temps. La définition de l’opérateur de transport adjoint donnée par l’équation 3.22 permet alors d’assurer que ˆ D E ˜ Qn (u|r~2 , ω~1 ) r~1 , ω~0 = C † (r~1 , w~0 |r~1 , w~1 )Qn (u|r~1 , ω~1 )dw~1 , r~2 ,ω~1

et en utilisant de même la définition de l’opérateur de collision adjoint donnée par l’équation 3.13, on s’assure que D E ˜ n (u|r~2 , ω~1 ) r~1 , ω~0 Q = C † {Qn } . r~2 ,ω~1

En injectant cette dernière relation dans l’équation 3.23, il vient  ˜ n+1 (u|r~1 , ω~0 ) = exp {−u1V (r~1 , ω~0 )} G C † {Qn } , Q

(3.24)

et en exprimant la génératrice au premier point de collision en fonction de la génératrice originelle, on trouve finalement la forme intégrale d’une équation de type Feynman-Kac discrète pour les vols exponentiels ˆ  Qn+1 (u|r~0 , ω~0 ) = dr~1 T † (r~0 , w~0 |r~1 , w~0 ) exp {−u1V (r~1 , ω~0 )} G C † {Qn } , (3.25) E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 39 On peut également dériver à partir de cette expression une forme intégro-différentielle de l’équation en utilisant la relation ~ r~ f (r~0 , w~0 ) − Σt f (r~0 , w~0 ) + Σt g(r~0 , w~0 ) = 0, w~0 .∇ 0 vraie pour des fonctions f et g liées par

f (r~0 , w~0 ) =

ˆ dr~1 T † (r~0 , w~0 |r~1 , w~0 )g(r~1 , ω~0 ).

Par conséquent, l’équation 3.25 peut également s’écrire  ~ r~ Qn+1 (u|r~0 , ω~0 ) − Σt Qn+1 (u|r~0 , ω~0 ) + Σt exp {−u1V (r~1 , ω~0 )} G C † {Qn } = 0. w~0 .∇ 0

(3.26)

Cette forme est à rapprocher de celle de l’équation 3.17 pour les processus continus. En suivant une démarche similaire à celle de la section 3.5, on peut également dériver l’équation pour les moments du nombre de collisions hnm ~0 ) = (−1)m V it (r~0 , ω

∂m Qn (u|r~0 , ω~0 )|u=0 . ∂um

En dérivant m fois l’équation 3.26 et en utilisant la formule de Faà di Bruno on obtient en effet m ~ r~ hnm −ω~0 .∇ V in+1 +Σt hnV in+1 = Σt 0

m X

m   m−k X X       m νj Bm,j C † niV n +Σt 1V (r~0 , ω~0 ) νj Bm−k,j C † niV n , k j=1 j=0 k=1

pour m ≥ 1.

3.7

Résultats théoriques principaux

L’approche par intégrales de chemins pour les processus discrets branchants est très générale car elle permet de caractériser de très nombreux systèmes décrits par des marches aléatoires. On s’attachera dans cette section à étudier quelques résultats probants de cette approche pour différents types de problèmes. Les deux premières sous-sections sont dévolues à la neutronique, puisque l’on étudiera respectivement les grandes lignes de la dérivation d’une solution analytique pour l’équation de Boltzmann linéaire stationnaire en dimension 2 [1], puis une généralisation de la formule de Cauchy pour les processus branchants [6, 7, 8]. Les deux sous-sections suivantes ouvrent le champ des applications tout d’abord à la théorie des jeux, en proposant une généralisation de la loi de l’arcsinus aux processus avec absorption [5], et à l’épidémiologie en caractérisant l’enveloppe convexe d’une épidémie animale lors de son déclenchement [9]. On ne cherchera pas ici à être exhaustif dans les démonstrations qui sont accessibles dans les publications, mais plutôt à en donner les grandes lignes. On replacera aussi ces différentes thématiques dans le cadre d’étude présenté dans la section précédente. De même, cette section ne présente pas l’ensemble des résultats publiés mais se concentre sur certains d’entre eux. Aussi, le lecteur intéressé par des applications de cette démarche à la théorie des estimateurs pourra compulser les références [2, 3, 4].

3.7.1

Solution de l’équation de Boltzmann linéaire stationnaire en dimension 2

Parmis les différents phénomènes physiques décrits par les sauts exponentiels, on trouve la propagation des neutrons, la migration des espèces chimiques et biologiques [19] ou l’étude des puits quantiques et des couches d’inversion en théorie des semi-conducteurs [20]. Ainsi, on s’intéresse ici à ce type de marche aléatoire restreinte au plan (d = 2), sans absorption, et caractérisée par un noyau de transport exponentiel et un noyau de collision isotrope. La question qui se pose est alors la suivante : si l’on considère une “particule” démarrant à l’origine, quel est le taux de fréquentation d’un point donné dans l’espace ? En d’autres termes, quelle est la densité de collision en ce point ? On sait que les marches aléatoires sont récurrentes en une et deux dimensions, ce qui signifie qu’un site, quelqu’il soit, est visité indéfiniment par la particule. Mais cela n’est plus vrai lorsque l’on ne considère qu’une partie de l’espace, avec des conditions de fuite (absorption totale) sur les bords. D’où l’intéret de la connaissance de la densité de collision. Une approche directe, consistant à trouver une solution à l’équation de Boltzmann linéaire ne permet pas de dériver de solution analytique. L’idée consiste donc à utiliser les marches aléatoires. L’originalité de la démarche exposée par la référence [1] pour résoudre ce problème repose sur l’utilisation d’un propagateur en espace-collisions noté Ψ(~r | n) donnant la probabilité de trouver la particule en ~r à la collision n.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 40 On construit une marche aléatoire, où la position de la particule à la n-ième collision est donnée par r~n =

n X

r~i .

i=0

On appelle πd (~r|r~0 ) la densité de probabilité de trouver une particule en ~r sachant que celle-ci se trouvait en 0 0 ~ ~ r à la collision précédente. La probabilité d’échantillonner un parcours de longueur l = ~r − r est donnée par 3.2. La probabilité d’échantillonner un point sur la (d − 1)-sphère unitaire étant donnée par 3.3, πd s’écrit πd (| ~r − r~0 |) =

T (| ~r − r~0 |) . | ~r − r~0 |d−1 ωd

(3.27)

Comme πd peut être vu comme le propagateur à une collision, il est lié à Ψ(~r, n), le propagateur à n collisions (qui représente la probabilité de trouver la particule en ~r à la n-ième collision) par l’équation ˆ Ψ(~r | n) = πd (| ~r − r~0 |)Ψ(r~0 | n − 1)dr~0 . (3.28) Si l’on développe cette dernière formule, elle prend l’expression d’une simple intégrale de chemin de type 3.5 ˆ Ψ(~r | n) = πd (| ~r − ~rn−1 |)πd (| ~r − ~rn−2 |)...πd (| ~r1 − ~0 |)d~r0 . (3.29) On cherche à réécrire cette intégrale de chemin dans l’espace d-dimensionel de Fourier (on reste en dimension d quelconque pour l’instant), accessible par le couple de transformée/anti-transformée f (z) = Fd {f (r)} et f (r) = Fd−1 {f (z)} [21] : ˆ +∞ d/2 f (z) = z 1−d/2 (2π) rd/2 Jd/2−1 (zr)f (r)dr, 0

−d/2

ˆ

+∞

f (r) = r1−d/2 (2π)

z d/2 Jd/2−1 (rz)f (z)dz, 0

avec Jν la fonction de Bessel modifiée de premier ordre et d’index ν [22]. Dans cet espace, l’équation 3.28 devient, après transformation de Fourier et récurrence implicite : (3.30)

Ψ(z|n) = πd (z)n

Cette dernière équation motive tout l’intérêt du travail dans l’espace de Fourier : les produits de convolution de l’équation 3.28 qui permettent de construire de manière itérative les probabilités de collision deviennent de simple produit dans cet espace. De plus, la Markovianité de la marche aléatoire en chaque point de collision permet de construire une récurrence triviale sur ces produits et d’obtenir 3.30. Pour expliciter la densité de collision stationnaire en ~r, on somme sur toutes les collisions le propagateur Ψ(r|n) : Ψ(r) = lim

N →∞

N X

Ψ(r|n) = lim

N →∞

n=1

N X n=1

Fd−1 {πd (z)n }

(3.31)

En utilisant la forme exponentielle du propagateur à un pas πd ( ~r − r~0 ) en dimension d, celui-ci peut être calculé explicitement dans l’espace de Fourier :   1 d z2 πd (z) = 2 F1 , 1, ; − 2 , (3.32) 2 2 σ avec 2 F1 la fonction hypergéométrique de Gauss [22]. En dimension d = 2, ce dernier prend la forme simplifiée 1 , 1 + z2 et permet le calcul du propagateur libre Ψ(r|n) via (3.30) π2 (z) = √ n

Ψ(r|n) =

E. Dumonteil

(3.33)

n

2− 2 r−1+ 2 K−1+ n2 (r)  , πΓ n2

(3.34)

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 41 associé à sa condition initiale Ψ(z, 0) = 1. La figure 3.5 compare ce résultat à des simulations Monte-Carlo, et indique un accord parfait.

Figure 3.5 Propagateur libre Ψ(r|n) pour d = 2. les simulations monte-carlo sont données par les croix bleues (n = 1), les cercles rouges (n = 5), les signes plus verts (n = 10) et les points noirs (n = 50). Les lignes solides sont les résultats théoriques. les lignes pointillées donnent la limite de la diffusion. Une des techniques utilisées pour le calcul de la densité de collision divergente en dimension d = 2 consiste à isoler la divergence en N . Le calcul de la densité de collision dans l’espace de Fourier est donné en effet par : −N/2 N X 1 − 1 + z2 √ Ψ(z|n) = lim Ψ(z) = lim . (3.35) N →∞ N →∞ 1 + z2 − 1 n=1

Cette formule autorise un retour dans l’espace direct, au terme de développements fastidieux non exposés ici, pour trouver log (N ) e−r Ei (−r) − 2 log (r) Ψ(r) ' + + 2π 2πr 2π (avec Ei la fonction exponentielle intégrale [22]). La divergence en N de cette intégrale indique qu’en milieu infini, dans le plan, le taux de fréquentation d’un site quelconque diverge (d’où l’expression “tous les chemins mènent à Rome dans le plan”). Il est alors possible de calculer la densité de collision ΨR (r) pour un point quelconque d’un disque de rayon R (les particules fuient sur les bords), en appliquant la méthode des images :   Ei (−r) − Ei (−re ) − 2 log rre e−re e−r − + (3.36) ΨR (r) = 2πr 2πre 2π où re = R [1 + u2 /R] avec la constante de Milne u2 ' 1 − 2/π 2 [23].

3.7.2

Généralisation aux processus branchants de la formule de Cauchy pour les longueurs de corde

La formule de Cauchy est connue en physique des réacteurs et en neutronique car elle permet de relier la longueur moyenne hliS parcourue dans un volume V , par des particules voyageant en ligne droite émises E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 42 aléatoirement sur la surface S du volume V , uniquement au volume V lui-même et à la surface S via la formule [28] hliS = 4V /S. (3.37) Cette formule, valable en dimension 3, peut être généralisée à une dimension d quelconque hliS = ηd V /S,

√ où ηd = π(d − 1)Γ((d − 1)/2)/Γ(d/2) est la constante de normalisation associée à un espace de dimension d. Des travaux plus récents ont permis de montrer que cette dernière loi était également valable pour les marches aléatoires de Pearson démarrant aléatoirement du bord du volume [24, 25, 29]. Comme la prise en compte du branchement s’avère cruciale en neutronique, on peut chercher à caractériser cette longueur moyenne de corde dans le cadre plus général des marches aléatoires branchantes. Pour simplifier les développements, on considérera des vols exponentiels réorientés de manière isotrope à chaque collision. A chacune de ces collisions, la probabilité d’avoir k fils est bk , et la génératrice associée à cette loi de probabilité est notée G(z) = b0 + b1 z + b2 z 2 + ... (on reprend les notations définies en 3.4). On peut alors utiliser l’équation 3.21 : cette équation se simplifie dans le régime stationnaire (que l’on peut définir uniquement si les longueurs parcourues par les particules sont finies, ce qui est assuré si ν1 ≤ 1) pour prendre la forme L† lm (r~0 , ω~0 ) + m1V (r~0 , ω~0 )lm−1 (r~0 , ω~0 ) + Σt

m X

νj Bm,j

 i  l ω (r~0 ) = 0,

(3.38)

j=2

où la fonction indicatrice porte sur le volume V , et où l’on a utilisé la notation qui consiste à associer à toute fonction h son opérateur de moyenne sur les directions ˆ dω~0 h(r~0 , ω~0 ). hhiω (r~0 ) = ωd Afin d’intégrer l’équation 3.38 sur les positions de départ on définit deux autres opérateurs : l’opérateur de moyenne surfacique ˆ ˆ dω~0 dS hhiS = ηd (ω~0 .~n)h(r~0 , ω~0 ), S ωd et l’opérateur de moyenne volumique ˆ hhiV =

dr~0 V

ˆ

dω~0 h(r~0 , ω~0 ), ωd

avec ~n la normale entrante définie sur la surface S au point considéré. La mesure de probabilité sur la moyenne d~ ω de surface prend la forme particulière ηd dS ω .~n) afin d’assurer que le flux incident isotrope soit uniformément S ωd (~ distribué sur S [26, 27]. En intégrant l’équation 3.38 sur toutes les positions de départ possibles et toutes les directions possibles, et en appliquant le théorème de la divergence de Gauss on a   m X





 V hlm iS = ηd m lm−1 V + Σt (ν1 − 1) hlm iV + Σt νj Bm,j li ω (r~0 ) V  . (3.39) S j=2 Si l’on s’intéresse à la moyenne des longueurs de corde, on a donc hliS = ηd

V [1 + Σt (ν1 − 1) hliV ] . S

(3.40)

Cette relation de récurrence sur les moments de la distribution des longueurs de corde est donc une généralisation de la formule de Cauchy aux processus de transport branchants. Elle relie la moyenne de la longueur parcourue par des marches aléatoires branchantes démarrant de points aléatoirement distribués sur la surface à la longueur parcourue par les mêmes marches démarrant d’un point aléatoirement choisi dans le volume. On remarque en particulier que si ν1 = 1, alors la formule de Cauchy est exactement retrouvée. Cela signifie que pour des processus branchants ayant un fils en moyenne, la longueur moyenne parcourue dans le volume par une marche démarrant au bord ne dépend toujours que du volume V et de la surface S via l’équation 3.37.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 43

3.7.3

La loi de l’arcsinus de Lévy avec absorption

La loi de l’arcsinus, découverte par Paul Lévy vers 1940, donne la distribution du temps de résidence sur le demi-axe des réels positifs (ou négatifs) d’un mouvement Brownien uni-dimensionel démarrant en 0. Cette √ loi admet comme densité de probabilité asymptotique en t la fonction Pt (tV ) = 1/π t(1−tV ), où le volume V d’observation est le demi-axe des réels positifs, t est le temps d’observation, et tV est le temps de résidence du mouvement Brownien dans V observé jusqu’au temps t. La forme caractéristique en U de cette loi indique ainsi que le mouvement Brownien a tendance à passer son temps d’un côté de son point de départ. Cette loi est également valable pour des processus discrets, comme par exemple des sauts de taille constante, pour lesquels on recense le nombre de “collisions” nV d’un côté donné du point de départ. Si l’on observe cette marche aléatoire √ jusqu’au n-ième saut, la densité de probabilité de nV est alors donnée par Pn (nV ) = 1/π n(1−nV ). Elle a par la suite été généralisée à de nombreux processus, Markoviens et non-Markoviens [30, 31, 32, 34, 35, 36, 37] et a trouvé des applications dans différents domaines, comme par exemple pour la compréhension du résultat d’élections : si deux candidats A et B s’affrontent dans une élection, et que le choix des électeurs pour l’un de ces deux candidats est aléatoire et de probabilité 1/2 pour chacun des candidats, on peut imaginer que le processus de dépouillement est une marche aléatoire sur l’axe des réels (ou des entiers) dont les sauts sont de taille constante (+1 pour le candidat A et -1 pour le candidat B). La loi de l’arcsinus de Lévy indique alors que si l’on observe le résultat à un temps t élevé, les chances sont fortes pour que l’un des deux candidats ait creusé l’écart et ait été en tête pendant l’essentiel du dépouillement. On cherche dans cette section à retrouver la loi de l’arcsinus de Lévy avec le formalisme introduit en section 3.6, puis à généraliser cette loi aux marches aléatoires avec absorption [5]. Considérons une marche aléatoire de type Pearson sur l’axe des réels se déplacant de +1 avec probabilité 1/2 et de −1 avec probabilité 1/2. Le “volume” d’observation est le demi-axe des réels positifs, c’est-à-dire V (x) = H(x), où H(x) est la fonction de Heaviside (dite aussi fonction échelon). Le noyau de transport de cette marche peut s’écrire T (x0 → x) =

1 1 δ(x − x0 − 1) + δ(x − x0 + 1), 2 2

où δ(x) est la fonction de Dirac qui permet d’assurer une probabilité égale aux événements x = x0 + 1 et x = x0 − 1. Comme ce processus discret ne comporte que des événéments de type “diffusifs” (pas d’absorption ou de branchement), l’équation 3.25 se simplifie considérablement pour prendre la forme ˆ Qn+1 (u|x0 ) = dx1 T † (x1 → x0 ) exp {−u1V (x1 )} Qn (u|x1 ). Comme la forme de l’opérateur de transport a été explicitée, cette équation permet le calcul analytique, par récurrence sur n et jusqu’à une valeur arbitraire de n, des probabilités discrètes associées aux différentes valeurs possibles de n et de nV . Le résultat des calculs est reporté dans le tableau 3.1, généré avec x0 = 0. n\nV 0 1 2 3 4 5 6 7

0 1 1/2 1/4 2/8 3/16 6/32 10/64 20/128

1

2

3

4

5

6

7

1/2 1/4 1/8 2/16 3/32 6/64 10/128

2/4 2/8 2/16 4/32 6/64 12/128

3/8 3/16 3/32 6/64 9/128

6/16 6/32 6/64 12/128

10/32 10/64 10/128

20/64 20/128

35/128

Table 3.1 Probabilité d’avoir fait nV collisions dans V à la n-ième collision Pn (nV |0) Comme la formule de récurrence ne permet pas de calculer directement une formule analytique pour Pn (nV |0), on peut chercher à deviner cette dernière en observant les valeurs particulières reportées dans le tableau (l’usage de l’encyclopédie en ligne des nombres premiers [33] peut également s’avérer être utile). On remarque que la colonne nV = 1 est une fonction combinatoire de n   n − 2  Pn (nV = 1|0) = 2−n , n−2 2

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 44 où d.e désigne la partie entière par excès. On remarque également que les colonnes se regroupent deux par deux et qu’au sein de ces binômes, une colonne et la suivante sont liées par un décalage d’une unité en n. Enfin, la colonne nV = 0 est obtenue par normalisation. Cela permet donc d’obtenir une expression générale de la forme  Pn (nV |0) =

− 2 nn−2



V n n



V

2

2−n .

2

La limite de cette expression pour n et nV grands est Pn (nV |0) =

1 π

p

nV (n − nV )

,

et l’on a bien retrouvé la loi de l’arcsinus de Lévy. On peut alors s’intéresser à un processus plus général pour lequel, à chaque collision, on a une probabilité ps d’être diffusé, et une probabilité pa d’être absorbé. En cas de diffusion, la loi du saut est la même que précédemment, en cas d’absorption le processus prend fin. Dans ce cas, l’équation 3.25 devient ˆ Qn+1 (u|x0 ) =

dx1 T † (x1 → x0 ) exp {−u1V (x1 )} (pa + ps Qn ),

et permet de générer par récurrence, comme précédemment, un tableau de valeurs 3.2 où les entrées sont des fonctions de ps . n\nV 0 1 2 3 4

0 1 1/2

2−ps 4 4−2ps 8 8−4ps −p3s 16

1 1/2

2−ps 4 4−2ps −p2s 8 8−4ps −2p3s 16

2

3

4

2ps 4 2ps (2−ps ) 8 2ps (4−2ps −p2s ) 16

3p2s 8 3p2s (2−ps ) 16

6p3s 16

Table 3.2 Probabilité d’avoir fait nV collisions dans V à la n-ième collision Pn (nV |0) dans le cas avec absorption La structure polynomiale rend l’identification de l’expression analytique de Pn (nV |0) plus délicate, mais les étapes du raisonnement sont les mêmes que dans le cas purement diffusif. En effet, la première colonne peut s’écrire p   3 2 1 − ps + 1 − p2s  ps 4+2y 2 + 2y 2 F1 (1, 2 + y, 3 + y, ps ) Pn (nV = 1|0) = + , 1+y 4 2 2(2 + y) pour n ≥ 2, où on a utilisé la notation y =

 n−3  . Aussi, les colonnes sont liées entre elles par 2

  p nV −1  n s V  Pn−nV +1 (1|0), Pn (nV |0) = nV 2 2 pour nV ≥ 2. La forme analytique explicite de Pn (nV |0) est donc assez complexe, mais son développement asymptotique pour n et nV grands s’écrit p   p nV −1  n 1 − ps + 1 − p2s s V  P∞ (nV |0) = . nV 2 4 2 La figure 3.6 compare la loi de l’arcsinus à cette formule qui la généralise, dans le cas où ps = 0.95.

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 45

Figure 3.6 Géneralisation de la loi de l’arcsinus, dans le cas de sauts discrets. Ronds bleus : ps = 1 et pa = 0. Triangles rouges : ps = 0.95 et pa = 0.05. Dans les deux cas, la limite de diffusion apparaît en pointillés.

3.7.4

Caractérisation de l’enveloppe convexe du déclenchement d’une épidémie

Introduction : le modèle SIR En 1925, Mc Kendrick propose un modèle mathématique permettant de décrire l’évolution temporelle d’une épidémie au sein d’une population animale ou humaine [39]. Pour ce faire, il subdivise cette population constante de N individus en trois sous-groupes : le groupe I(t) des individus infectés au temps t, celui des individus susceptibles de le devenir S(t), et le groupe R(t) des gens qui “recouvrent” de la maladie, pour l’avoir déjà contracté et en être immunisé, ou par décès. Ce modèle, dit SIR [38], est gouverné par les équations d’évolution suivantes S(t) + I(t) + R(t) = N, dS/dt = −βIS,

dI/dt = βIS − γI, dR/dt = γI, où β est le taux de transmission de l’épidémie (βSdt nouveaux infectés en dt, par infecté, et donc βSIdt nouveaux infectés en dt au total) et γ est le taux de “récupération” (probabilité γdt de récupérer en dt). C’est un modèle de champ moyen (on ne considère par les fluctuations) complètement connecté (n’importe quel infecté est à même d’entrer en contact avec n’importe quel susceptible). Or, la phase de déclenchement d’une épidémie est souvent sujette à des fluctuations fortes dues au nombre faible d’individus infectés au démarrage, et par conséquent plusieurs travaux [40, 41, 42] ont permis de prendre en compte ces fluctuations, à l’aide de processus stochastiques de type Galton-Watson (processus branchant binaire, sans dépendance spatiale). Afin de décrire au mieux cette phase du déclenchement de l’épidémie, il est important de pouvoir prendre en compte ces fluctuations temporelles, mais également de construire un modèle spatial d’évolution de l’épidémie. Cette sous-section décrit ainsi les grandes lignes du développement d’un tel modèle permettant la caractérisation de l’étendue spatiale d’une épidémie animale dans sa phase de déclenchement [9]. La caractérisation de cette étendue spatiale est importante en ce qu’elle permet, par exemple, de délimiter une zone spatiale de mise en quarantaine pour lutter contre la propagation de l’épidémie. L’idée consiste à calculer l’enveloppe convexe d’un mouvement Brownien branchant. En effet, le mouvement Brownien est le paradigme adéquat pour la description

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 46 de la migration animale, et le processus de branchement est à même de décrire précisément la transmission d’une maladie infectieuse d’un individu infecté à un individu susceptible. Mouvement Brownien branchant Si l’on considère le déclenchement de l’épidémie (un individu infecté), on vérifie que S = N − 1 ' N , et par conséquent la population d’infectés est donnée par dI/dt = (R0 − 1)γI,

βN γ

où on a défini le taux de reproduction R0 = . Si le taux de reproduction est inférieur à 1 l’épidémie s’éteint, s’il est supérieur à 1 la population des infectés envahit celle des susceptibles, et si R0 = 1 on est dans le cas critique (il existe des similitudes fortes entre la propagation des épidémies et le transport des neutrons en milieu fissile, en neutronique R0 est le rayon critique d’un milieu fissile qui permet de préserver la taille d’une population de neutrons qui s’accroît par fissions et qui disparaît par fuites et par absorptions). Par conséquent un individu infecté peut être sujet à trois événements distincts dans un intervalle de temps dt : • il peut recouvrir avec probabilité γdt. Par analogie avec la section 3.4, cette probabilité est celle d’une absorption (événement b0 ). • il peut infecter un susceptible avec probabilité γR0 dt. Cette probabilité est celle d’un branchement à deux fils (événement b2 ).

2 • il peut se déplacer de ∆x : cette diffusion est alors caractérisée par son coefficient D tel que ∆x = 2Ddt. Une diffusion peut être vue comme un événement à un fils (b1 ). Une fois le modèle défini, on cherche à caractériser la mesure que l’on souhaite faire. En l’occurrence, celle de l’étendue spatiale de l’épidémie. Une méthode courante pour ce faire consiste à enregistrer la position de tous les individus du début de l’épidémie à un temps t donné, et à en calculer l’enveloppe convexe (i.e. le polygone minimum permettant de contenir l’ensemble des positions). Ainsi, en termes mathématiques, le problème consiste à calculer l’enveloppe convexe du maximum d’un mouvement Brownien branchant. Enveloppe convexe A l’aide d’arguments fondés sur la géométrie intégrale, Cauchy a proposé une méthode permettant le calcul du périmètre et de l’aire de courbes convexes en 2 dimensions [43]. Cette méthode repose sur l’utilisation de la fonction support M (θ), qui consiste à calculer pour chaque valeur de l’angle θ (angle polaire entre l’axe x et le point considéré) la distance minimale d’approche des perpendiculaires venant de l’infini à la droite portée par M (θ) permettant de ne pas intersecter la courbe convexe (voir la figure 3.7).

Figure 3.7 Construction de l’enveloppe convexe du processus en dimension 2, avec la fonction M (θ) qui représente la distance entre l’origine et l’enveloppe selon la direction θ. On définit cette fonction support par M (θ) = max (xτ cos(θ) + yτ sin(θ)) , O≤τ ≤t

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 47 où l’on a considéré un processus stochastique [x(τ ), y(τ )] en dimension 2, avec 0 ≤ t ≤ τ . Le périmètre L et l’aire A de la courbe convexe sont alors donnés par ˆ L=



(3.41)

M (θ)dθ, 0

et 1 A= 2

ˆ 0



(3.42)

 M (θ)2 − M 0 (θ)2 dθ.

Réduction dimensionnelle Des travaux récents ont permis d’adapter cette formule à l’enveloppe convexe d’un mouvement Brownien, en remarquant que les processus de diffusion sont dotés d’une symétrie rotationelle [44, 45]. Ainsi, les équations pour le périmètre et l’aire restent-elles valables pour θ = 0. Cette même réflexion s’applique au mouvement Brownien branchant, et permet de réécrire le périmètre et l’aire du maximum des positions comme des fonctions de M (0). Or, en s’appuyant sur la figure , on voit que M (0) prend la valeur M (0) = xτ =tm = xm (t), où xm (t) est la valeur maximum en x prise par la processus considéré de 0 à t, et M 0 (0) est donné par M 0 (θ = 0) = −xtm sin(θ)|θ=0 + ytm cos(θ)|θ=0 = ytm (t), où ytm (t) est la valeur de y au temps tm c’est-à-dire au temps auquel x = xm (t) (voir figure 3.8).

Figure 3.8 Marche aléatoire branchante composée de 5 individus, représentés par des couleurs différentes, et issus d’un unique individu ayant démarré la marche au point O. Les individus infectés au temps d’observation final t sont marqués par un point rouge. Les individus ayant recouvré sont marqués par un point noir. La trajectoire rouge est celle de l’individu ayant atteint un maximum en xm au temps tm . L’ordonnée correspondante est y(tm ). La figure de gauche donne les trajectoires dans le plan (x, y), la figure du centre donne la position en x en fonction de t et celle de droite donne y en fonction de t. Si l’on considère la moyenne des équations 3.41 et 3.42 on a alors hL(t)i = 2π hM (0)i = 2π hxm (t)i , et





 

 hA(t)i = π M (0)2 − M 0 (0)2 = π x2m (t) − y 2 (tm ) = π x2m (t) − 2D htm (t)i . Pa conséquent, les propriétés moyennes de l’enveloppe convexe ne nécessitent que l’étude d’un mouvement Brownien branchant en une dimension : la connaissance de la valeur maximum moyenne xm (t) de ce processus permet alors d’accéder à tm puis à y(tm ).

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 48 Equation pour le maximum d’un mouvement Brownien branchant en une dimension Notons Qt (xm ) la probabilité que le maximum en x du processus observé jusqu’au temps t soit inférieur à xm , pour un mouvement Brownien branchant en une dimension. L’équation pour cette grandeur est en fait identique à l’équation 3.17, où l’opérateur de transport est remplacé par un opérateur de diffusion. Cela s’interprète en écrivant Qt+dt (xm ) Qt+dt (xm ) = γdt + (1 − γ(R0 + 1)dt) hQt (xm − ∆x)i + γR0 dtQ2t (xm ),

(3.43)

car si l’on considère la probabilité qu’en t + dt le maximum en x soit inférieur à xm , alors trois cas de figure se présentent : • avec probablité γdt l’individu disparaît de 0 à dt (passe du statut I au statut R) et par conséquent la probabilité que son maximum en x reste inférieur à xm est de 1. • avec probabilité 1 − γ(R0 + 1)dt l’individu se déplace de ∆x entre 0 et dt et par conséquent la probabilité que x + ∆x < xm au temps t s’écrit Qt (x + ∆x < xm ) = Qt (xm − ∆x) (on prend garde par la suite à moyenner cette probabilité sur tous les ∆x possibles). • avec probabilité γR0 dt l’individu contamine un individu sain entre 0 et dt et par conséquent la probabilité que le maximum de la marche au temps t soit inférieur à xm vaut Q2t (xm ) (car les deux individus deviennent “indépendants” et ont la même probabilité d’avoir un maximum inférieur à xm au temps t. Le développement de l’équation 3.43 est alors identique au développement effectué dans les sections 3.2 et 3.4 et permet d’obtenir ∂2Q ∂ Q = D 2 −γ(R0 + 1)Q + γR0 Q2 + γ. (3.44) ∂t ∂xm Or Qt (x < xm ) est la cumulative du processus qt (x = xm ), qui donne la probabilité que le maximum du processus soit en xm au temps t. On peut donc obtenir le périmètre moyen de l’enveloppe convexe car ˆ +∞ hL(t)i = 2π hxm (t)i = 2π qt (xm )xm dxm , (3.45) 0

et par intégration par parties

ˆ hL(t)i = 2π

+∞

0

[1 − Qt (xm )] dxm .

Par des calculs similaires, on montre que l’aire de l’enveloppe convexe s’écrit ˆ +∞ hA(t)i = π [2xm (1 − Qt (xm )) − Tt (xm )] dxm ,

(3.46)

0

où Tt (xm ) est une grandeur qui obéit à la relation   ∂2 ∂ Tt − qt (xm ) = D 2 + 2γR0 Qt − γ(R0 + 1) Tt . ∂t ∂xm

(3.47)

Ainsi, le périmètre moyen donné par 3.45 et l’aire moyenne donnée par 3.47 peuvent être calculés en résolvant les équations 3.44 et 3.47 respectivement. Cette résolution est numérique, mais peut également être analytique dans la limite asymptotique en temps. Les figures donnent le périmètre moyen et l’aire moyenne calculées par une résolution numérique de ces équations pour différentes valeurs du taux de reproduction R0 . Dans la limite asymptotique en temps, et si l’on se place à la criticité, une résolution analytique donne alors accès aux formules suivantes s 6D hL(t → +∞)i = 2π + O(t−1/2 ), γ et hA(t → +∞)i =

24πD ln t + O(1). 5γ

Lorsque t devient grand, le périmètre moyen de l’enveloppe convexe de l’épidémie sature, mais son aire diverge. Ce résultat surprenant est lié aux queues des distributions de probabilités du périmètre et de l’aire, qui sont en loi de puissance : asymptotiquement la distribution de probabilité du périmètre évolue en L13 quand celle de l’aire varie en 24πD/(5γ) . Ce sont donc les fluctuations qui permettent d’avoir une aire moyenne divergente pour A2

E. Dumonteil

Université Paris-Sud

CHAPITRE 3. MODÉLISATION STOCHASTIQUE DU TRANSPORT DES PARTICULES NEUTRES 49 un périmètre moyen fini. Les distributions théoriques du périmètre et de l’aire ainsi que les valeurs du périmètre moyen et de l’aire moyenne sont comparées (figures 3.9 et 3.10 respectivement) à des simulations Monte-Carlo et indiquent un excellent accord.

Figure 3.9 Figure de gauche : périmètre moyen hL(t)i en fonction du temps t (en unités arbitraires), pour différentes valeurs du taux de reproduction R0 . Le résultat théorique (calculé par résolution numérique) apparaît en ligne solide et est comparé aux simulations Monte-Carlo d’un mouvement Brownien branchant notées par des points (105 individus simulés). Figure de droite : distribution de probabilité du périmètre L observé au temps t pour R0 = 1. Pour les deux figures, on a pris D = 1/2 et R0 γ = 0.01.

Figure 3.10 Figure de gauche : Aire moyenne hA(t)i en fonction du temps t (en unités arbitraires), pour différentes valeurs du taux de reproduction R0 . Le résultat théorique (calculé par résolution numérique) apparaît en ligne solide et est comparé aux simulations Monte-Carlo d’un mouvement Brownien branchant (105 individus simulés). Figure de droite : distribution de probabilité de l’aire A observée au temps t pour R0 = 1. Pour les deux figures, on a pris D = 1/2 et R0 γ = 0.01.

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

50

Références Articles signés ou cosignés [1] A. Zoia, E. Dumonteil, A. Mazzolo, "Residence time and collision statistics for exponential flights : the rod problem revisited", Physical Review E 84 021139 (2011) [2] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision number statistics for transport processes", Physical Review Letters 106, 220602 (2011) [3] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision densities and mean residence times for d-dimensional exponential flights", Physical Review E 83 041137 (2011) [4] A. Zoia, E. Dumonteil, A. Mazzolo, "Collision statistics for renewal processes with anisotropic scattering and absorption", Physical Review E 84 061130 (2011) [5] A. Zoia, E. Dumonteil, A. Mazzolo, "Counting statistics : a Feynman-Kac perspective", Physical Review E 85, 011132 (2012) [6] A. Zoia, E. Dumonteil, A. Mazzolo, "Discrete Feynman-Kac formulas for branching random walks", EPL 98, 40012 (2012) [7] A. Zoia, E. Dumonteil, A. Mazzolo, S. Mohamed, "Branching exponential flights : travelled lengths and collision statistics", J. Phys. A : Math. Theor. 45 425002 (2012) [8] A. Zoia, E. Dumonteil, A. Mazzolo, "Properties of branching exponential flights in bounded domains", EPL 100, 40002 (2012) [9] E. Dumonteil, S. Majumdar, A. Rosso, A. Zoia, "Spatial extent of an outbreak in animal epidemics", Proceedings of the National Academy of Sciences 110(11) 4239-4244 (2013)

Autres articles [10] K. Pearson, Nature 27, 294 (1905). [11] J. C. Kluyver, Proc. Koninklijke Akademie van Wetenschappen te Amsterdam 8, 341 (1906). [12] J. W. Strutt (Lord Rayleigh), Philos. Magazine 6 (37), 321 (1919). [13] M. Kac, Trans. Amer. Math. Soc. 65, 1 (1949). [14] M. Kac, in Proc. Second Berkeley Symp. on Math. Statist. and Prob. (UCP, 1951), pp. 189-215. [15] D. A. Darling and M. Kac, Trans. Amer. Math. Soc. 84, 444 (1957). [16] M. Kac, Probability and related topics in physical sciences (Lectures in applied mathematics, Wiley, 1957). [17] S. Sabhapandit, S.N. Majumdar and A. Comtet, "Statistical Properties of Functionals of the Paths of a Particle Diffusing in a One-Dimensional Random Potential", Phys. Rev. E v-73, 051102 (2006). [18] Bell and Glasstone, "Nuclear Reactor Theory (New York : Van Nostrand Reinhold Company)" (1970). [19] H. T. Hillen and G. Othmer, Siam J. Appl. Math 61, 751 (2000). [20] C. Jacoboni and P. Lugli, The Monte-Carlo method for semiconductor device simulation (Springer, 1989). [21] J. Dutka, Arch. Hist. Exact Sci. 32, 351 (1985). [22] M. Abramowitz and I. A. Stegun (Eds.), Handbook of mathematical functions with formulas, graphs, and mathematical tables, 9th edition (Dover, NY, 1972). [23] I. Freund, Phys. Rev. A 45, 8854 (1992). [24] S. Blanco and R. Fournier, Europhys. Lett. 61, 168 (2003). [25] A. Mazzolo, Europhys. Lett. 68, 350 (2004). [26] A. Mazzolo, J. Phys. A 37, 7095 (2004). [27] L.A. Santalo, Integral Geometry and Geometric Probability (Addision-Wesley, Reading, Mass) (1976). [28] P. Reuss, "Neutron Physics", EDP Sciences [29] O. Bénichou et al., Europhys. Lett. 70, 42 (2005).

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

51

[30] E. Barkai, J. Stat. Phys. 123, 883 (2006). [31] S.N. Majumdar and A. Comtet, "Local and the Occupation Time of a Particle Diffusing in a Random Medium", Phys. Rev. Lett. 89, 060601 (2002). [32] W. Feller, "An introduction to probability theory and its applications", 3rd Edition, Wiley, New-York (2002). [33] http ://oeis.org/ [34] J. Lamperti, Trans. Am. Math. Soc. 88, 380 (1958). [35] S. Sabhapandit, S. N. Majumdar, and A. Comtet, Phys. Rev. E 73, 051102 (2006). [36] C. Godreche and J. M. Luck, J. Stat. Phys. 104, 489 (2001). [37] A. Baldassarri, J. P. Bouchaud, I. Dornic, and C. Godreche, Phys. Rev. E 59, R20 (1999). [38] Bailey NTJ, "The Mathematical Theory of Infectious Diseases and Its Applications" (Griffin, London, 1975). [39] McKendrick AG, "Applications of mathematics to medical problems" Proc Edinburgh Math Soc 44(1) :98130 (1925). [40] P. Whittle, "The outcome of a stochastic epidemic : a note on Bailey’s paper", Biometrika 42(1/2) :116-122 (1955). [41] D.G. Kendall, "Deterministic and stochastic epidemics in closed populations", Proc 3rd Berkeley Symp Math Statist Prob 4 :149-165 (1956). [42] M.S. Bartlett, "An Introduction to Stochastic Processes", Cambridge Univ Press, Cambridge, United Kingdom (1956). [43] A. Cauchy, "Mémoire sur la rectification des courbes et la quadrature des surfaces courbes.", Mém Acad Sci Paris 22 :3-15 (1850). [44] J. Randon-Furling, S.N. Majumdar, A. Comtet, "Convex hull of N planar Brownian motions : Exact results and an application to ecology", Phys Rev Lett 103(14) : 140602 (2009). [45] S.N. Majumdar, A. Comtet, J. Randon-Furling, "Random convex hulls and extreme value statistics", J Stat Phys 138(6) :955-1009 (2010).

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

52

E. Dumonteil

Université Paris-Sud

Chapitre

4

Simulation Monte-Carlo du transport des particules neutres Sommaire 4.1

Principe de la simulation Monte-Carlo du transport des particules . . . . 4.1.1 Principe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.2 Algorithme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 TRIPOLI-4® . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Protection et méthodes de réduction de variance . . . . . . . . . . . . . . . 4.2.1 Principe de base de la réduction de variance . . . . . . . . . . . . . . . . . . . . 4.2.2 Réduction de variance pour les problèmes de transport . . . . . . . . . . . . . . 4.3 Criticité et itération sur la puissance . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Exemple illustratif : une cellule de réacteur à eau sous pression . . . . . . . . . 4.3.3 L’itération sur la puissance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Diagnostic de la convergence des sources à l’aide de l’entropie de Boltzmann . . 4.3.5 Diagnostic des corrélations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

54 54 57 57 58 58 59 65 65 66 68 69 71 75

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

54

L

e transport des particules par la méthode de Monte-Carlo constitue le pendant formel, du point de vue des méthodes numériques, de la théorie du transport stochastique exposée au Chapitre 3. En effet, le formalisme des intégrales de chemin met en évidence la recherche d’une mesure à un détecteur par la prise en compte de tous les chemins possibles pour les particules, de la même manière qu’un code Monte-Carlo de transport peut simuler numériquement toutes ces trajectoires. Si les hypothèses doivent néanmoins rester très fortes du côté théorique pour permettre des calculs analytiques, du côté numérique cette contrainte n’existe pas et il est possible de réaliser des simulations reproduisant des configurations physiques de manière extrêmement réaliste. Aussi peut-on dire, dans le cadre de la neutronique, que les simulations Monte-Carlo sont utilisées comme un outil quantitatif et réaliste, quand la théorie du transport se pose comme un outil qualitatif. L’objet de ce chapitre est de présenter de manière synthétique ma contribution au développement des méthodes et des codes Monte-Carlo. Dans une première section on présentera le principe de la simulation MonteCarlo du transport des particules neutres appliquée à la neutronique ainsi que TRIPOLI-4® [1, 2, 3, 5] , le code Monte-Carlo de neutronique développé au CEA. Une seconde section permettra d’introduire les méthodes de réduction de variance présentes dans le code TRIPOLI-4®. Ces méthodes sont utilisées dans la simulation des problèmes de protection (dimensionnement des infrastructures de radioprotection, instrumentation nucléaire, dispositifs médicaux etc.), pour lesquels les neutrons n’évoluent pas dans des milieux fissiles mais sont caractérisés par une atténuation très forte du flux des particules, rendant la mesure souvent très délicate. Une nouvelle méthode de réduction de variance utilisant des réseaux de neurones sera alors discutée [4]. Dans une troisième section on abordera les calculs de criticité, permettant de décrire le transport des neutrons dans des milieux fissiles (prise en compte du branchement). On verra que ces calculs posent des problèmes dans l’estimation des grandeurs physiques qui souffrent d’erreurs systématiques (aussi appelées biais). On présentera alors des travaux concernant le diagnostic de ces biais [5, 6, 7]. Des pistes préliminaires concernant le traitement numérique de ces biais seront données au chapitre suivant, section 5.1.

4.1

Principe de la simulation Monte-Carlo du transport des particules

4.1.1

Principe

4.1.1.1

Equation de Boltzmann linéaire en neutronique

Comme les densités caractéristiques de neutrons dans un cœur de réacteur nucléaire sont approximativement de 107 /cm3 (voir appendice A.5), il faudrait typiquement être capable de simuler environ 1014 neutrons pour avoir accès aux différents moments des distributions d’intérêt (mesure d’un flux dans un détecteur ex-core, distribution de la puissance dans le cœur d’un réacteur, etc.). Ceci est encore hors d’atteinte aujourd’hui. Par conséquent, les simulations Monte-Carlo pour le transport des neutrons visent à calculer des grandeurs moyennes. On se contente donc dans ce chapitre de discuter la résolution Monte-Carlo de l’équation de Boltzmann (en champ moyen), et pas de l’équation de Feynman-Kac pour la probabilité. En revanche, il est important de pouvoir associer des barres d’erreurs statistiques à ces grandeurs moyennes mesurées par un code, et à ce titre la section concernant la réduction de variance discute des moyens employés pour réduire la variance de la série moyenne d’une mesure Monte-Carlo (et non pas la dispersion physique réelle de la grandeur calculée). De plus, en neutronique, la probabilité qu’a un neutron de faire une interaction donnée (diffusion élastique, inélastique, fission, etc.) dépend fortement de l’énergie du neutron. Il convient donc à présent de reformuler l’équation de Boltzmann 3.10 afin de faire apparaître cette dépendance des sections efficaces et du flux neutronique à l’énergie du neutron. Ainsi, en remplaçant les dépendances en direction par des dépendances en vitesse, et en faisant apparaître explicitement les dépendances en énergie, cette équation à la stationnarité se réécrit ˆ ~ ω ~ .∇φ(~r, ~v ) + Σt (~r, ~v )φ(~r, ~v ) = d~v 0 Σt (~r, v~0 )φ(~r, v~0 )C(v~0 → ~v ; ~r) + Q(~r, ~v ), (4.1) où :

le flux neutronique ψ(~r, ω ~ ) a été remplacé par φ(~r, ~v ) Q(~r, ~v ) désigne les sources • Σt (~ r, ~v ) est la section efficace totale macroscopique du milieu ~0 → ~v ; ~r) plus usitée en neutronique remplace la notation C(~r, ~v |~r, v~0 ) • la notation C(v On s’assure bien que comme l’équation de Boltzmann présentée ici est linéaire par rapport aux sources, l’estimation d’une grandeur moyenne peut être réalisée de manière non biaisée à partir d’un échantillon de neutrons. •



E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES 4.1.1.2

55

Résolution de l’équation de Boltzmann par la méthode Monte-Carlo appliquée au transport des particules

D’un point de vue formel, la méthode de Monte-Carlo appliquée au transport des particules consiste à simuler la réalité microscopique du système. On échantillonne pour chaque neutron les différentes étapes de sa vie : lieu de la première collision, choix du nucléide sur lequel le neutron interagit, choix de la réaction, échantillonnage des paramètres cinématiques de la réaction, ... Ainsi, il n’est nul besoin de justifier cette démarche, qui possède une valeur intrinsèque. Dit autrement, il n’est nul besoin de se poser de questions pour “simuler” : on peut se contenter “d’imiter” le plus fidèlement possible. Ceci dit, le chapitre suivant montrera que l’étude des équations associées au processus stochastique simulé peut s’avérer fort utile dans certains cas. Il paraît donc important de pouvoir établir un lien entre le problème résolu par la méthode Monte-Carlo, et sa formulation mathématique. L’équation de Feynman-Kac établit ce lien entre les processus de diffusion (vus comme des intégrales de chemin) et certaines classes d’équations aux dérivées partielles (qui représentent des équations pour la probabilité du processus diffusif considéré). Dans le chapitre précédent, ce même lien a été discuté dans le cas des processus de transport, éventuellement branchants. On cherche par conséquent à présent à poursuivre cette similitude dans le cas simple de l’équation de Boltzmann linéaire (qui ne représente que le premier moment de la solution de l’équation plus générale pour la probabilité). On cherche à réécrire l’équation de Boltzmann, mais sous forme intégrale. On introduit tout d’abord χe (~r, ~v ) = Σt (~r, ~v )φ(~r, ~v ), (4.2) 0 0 la densité de collision entrante, ainsi que K(r~ , v~ → ~r, ~v ) K(r~0 , v~0 → ~r, ~v ) = C(v~0 → ~v ; r~0 )T (r~0 → ~r; ~v ),

(4.3)

qui représente la composition respective des opérateurs de collision C et de transport T . De la même manière que pour la dérivation de l’équation 4.1, on établit une équation de bilan au point (~r, ~v ) de l’espace des phases qui traduit que les particules en ce point ont été transportées directement de la source ou ont subi une collision et un transport à partir d’un autre point (r~0 , v~0 ) de l’espace des phases ˆ ˆ ˆ 0 0 0 0 0 0 ~ ~ ~ ~ ~ ~ χe (~r, ~v ) = χe (r , v )K(r , v → ~r, ~v )dv dr + Q(r~0 , ~v )T (r~0 → ~r; ~v )dr~0 . (4.4) 1

On aimerait simplifier cette équation intégrale en faisant apparaître directement le terme source (et pas une ´ intégrale sur ce terme). On introduit alors la densité de collision sortante χ(~r, ~v ) = χe (~r, v~0 )C(v~0 → ~v ; ~r)dv~0 + Q(~r, ~v ) (le nombre de départs en ~r selon la vitesse ~v est la densité de collision entrante après une collision, ou directement la source) et on s’assure que ˆ χe (~r, ~v ) = χ(r~0 , ~v )T (r~0 → ~r; ~v )dr~0 . (4.5) En notant L(r~0 , v~0 → ~r, ~v ) la composition des opérateurs de transport puis de collision L(r~0 , v~0 → ~r, ~v ) = T (r~0 → ~r; v~0 )C(v~0 → ~v ; ~r),

(4.6)

l’équation 4.4 peut effectivement être simplifiée par l’utilisation de 4.5 puisque l’on obtient ˆ ˆ χ(~r, ~v ) = χ(r~0 , v~0 )L(r~0 , v~0 → ~r, ~v )dr~0 dv~0 + Q(~r, ~v ).

(4.7)

On peut alors alléger la notation à l’aide de l’opérateur intégral L χ = Lχ + Q.

´

(4.8)

Cette équation en fait une équation de Fredholm de seconde espèce (une équation du type ϕ(x) = λ α(x, y)ϕ(y) dy+ β(x)) dont on peut écrire la solution à l’aide d’un développement en série de Von Neumann puisque X X χ = (I − L)−1 Q = Ln Q = χn , (4.9) n≥0

n≥0

1. Dans ce contexte, les noyaux de collision C et de transport T ne représentent plus des fonctions mais des opérateurs. Ainsi, la notation conditionnelle T (~ r r~0 ) d’usage en probabilité est elle remplacée par la notation T (r~0 → ~ r) plus couramment associée aux opérateurs. Dans le cas présent, on notera également que les opérateurs agissent à droite.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

56

écriture qui fait bien apparaître la récurrence entre deux collisions successives χn = Lχn−1 (il faut cependant s’assurer que la norme de l’opérateur L est inférieure à 1, ce qui est prouvé sous de larges conditions [16]). Cette notation utilisant des opérateurs reproduisant des suites de transport et collision décrit ainsi au plus près le processus de la simulation. On détaille à présent la forme de ces opérateurs dans la simulation. 4.1.1.3

Expression des opérateurs de transport et de collision

L’équation 3.27 avait introduit le produit des opérateurs de transport et de collision noté π3 , dans un cas très simplifié (milieu homogène, pas de dépendance énergétique, ...). En neutronique, le code de transport Monte-Carlo utilise pour l’opérateur de transport l’expression suivante ( ˆ ) |~r−r~0 | 0 0 ~ ~ T (r → ~r; ~v ) = Σt (~r, ~v ) exp − Σt (r + s~ ω , ~v ) ds , (4.10) 0

où l’opérateur T (r~0 → ~r; ~v ) fait apparaître la probabilité de ne pas faire de collision de r~0 à ~r (terme exponentiel) ainsi que la probabilité de faire une collision en ~r précisement (proportionnelle à Σt (~r, ~v )). De même, l’opérateur de collision a pour forme détaillée C(E, ω ~ → E 0 , ω~0 ; ~r) =

X isotope j

  σj,s (E) σj,i (E) Σj,t (~r, E) X [νj,i ] fj,i (E, ω ~ → E 0 , ω~0 ), Σt (~r, E) reaction i σj,t (E) σj,s (E)

(4.11)

où : • E est l’énergie du neutron • Σj,t représente la section efficace macroscopique totale du nucléide j • σj,x est la section microscopique du nucléide j pour la réaction x (les sections efficaces microscopiques et macroscopiques sont liées par Σj = Nj σj , avec Nj la concentration en cm−3 du nucléide j) • νj,i est le nombre moyen de neutrons produits par la réaction i sur le nucléide j . La notation [.] indique que la modification est prise en compte, dans la simulation Monte-Carlo, sur le poids statistique de la particule défini sous-section 4.2.1 • fj,i , donne les densités de probabilité de renvoi en angle et en énergie Cette formule fait apparaître distinctement le tirage du noyau, de l’interaction, des caractéristiques après choc, ainsi que les corrections de poids. 4.1.1.4

Estimateurs usuels

Nous avons vu au chapitre 3 qu’il était possible d’estimer des grandeurs en recensant la distance parcourue par une particule dans un détecteur, ou en recensant le nombre discret de collisions réalisées par la particule dans le détecteur. Les estimateurs des grandeurs calculées par le code fonctionnent de la même manière : pour calculer un flux on peut choisir de dénombrer un nombre de collisions, ou de mesurer une distance parcourue par la particule. On cherche ici à définir plus précisement les différents types d’estimateur. Les deux premiers présentés sont des estimateurs de type volumique, le troisième est un estimateur surfacique et le dernier un estimateur en un point. Pour le calcul du flux dans le domaine V (de volume V ), l’estimateur “collision” recense simplement les collisions dans le domaine 1 X 1V (xk ) φV = wk , (4.12) V Σt (xk ) k

avec w le poids statistique de la particule, 1V (x) la fonction indicatrice du domaine V, et k un indice de roulement sur les collisions. L’estimateur “corde” consiste à encaisser directement la distance parcourue dans le domaine et s’écrit 1 X φV = wl 1V (l)l, V l

avec l la distance parcourue par la particule. On notera que cette longueur est bien homogène à Σ1t . Pour estimer un taux de réaction (réaction i) on multiplie le flux (indépendamment de l’estimateur) par Σi (x). Avec J(S, ~v ) le courant de neutrons à travers une surface S (de normale ~n(~r ∈ S)) on a ˆ ~ r, ~v ).~n(~r) dS, J(S, ~v ) = J(~ S

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

57

et il suffit alors d’encaisser les poids des neutrons traversant la surface S ou, si l’on veut estimer le flux sur la surface, il suffit alors d’encaisser (pour les particules traversant la surface) le poids multiplié par 1/~ ω .~n (car ~ J(~r, ~v ) = φ(~r, ~v )~ ω ). Concernant les estimateurs ponctuels, l’idée consiste à sommer, après chaque collision au point ~r, la probabilité que la collision suivante ait lieu au point de détection r~0 . Cette probabilité s’écrit ( ˆ ) |~ r0 −~ r| 1 exp − Σt (~r + s~ ω0 ) ds 2. |~r0 − ~r| 0

4.1.2

Algorithme

Après avoir échantillonné la position, l’énergie et la direction de la particule suivant des distributions sources fournies par l’utilisateur, le code va d’abord échantillonner le prochain lieu de collision suivant l’équation 4.10. Après s’être assuré que la particule ne fuit pas de la géométrie, ou ne franchit pas une interface, il échantillonne le nucléide sur lequel a lieu la réaction ainsi que la nature de celle-ci, suivant l’équation 4.11. Si la particule n’est pas absorbée par le noyau, alors ce processus reprend de manière itérative jusqu’à absorption ou fuite de la géométrie (ou du domaine d’étude si l’utilisateur en a défini un, comme une coupure en énergie par exemple). La simulation de l’histoire d’un neutron ou d’un photon est résumée sur la figure 4.1. L’utilisateur ayant précisé au code des scores et des régions d’intérêt, ce dernier estime lors de la simulation l’espérance et la variance du score d’intérêt associées à la contribution de chacune des particules simulées.

Figure 4.1 Algorithme de principe d’une simulation Monte-Carlo de l’histoire d’un neutron. Cet algorithme ne fait pas apparaître la gestion des réactions de fission.

4.1.3

TRIPOLI-4®

Le CEA développe depuis les années 1960 des codes Monte-Carlo de transport des particules. TRIPOLI4® [1, 2, 3] représente la 4ème génération de ces codes, et est financé par un accord tripartite entre le CEA,

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

58

EDF et AREVA. Il est disponible sur demande à l’Agence à l’Energie Nucléaire de l’OCDE [11]. Il simule essentiellement le transport des neutrons, des photons, le transport couplé neutron-photon ainsi que le transport de certaines particules chargées (e+ , e− ). Il est utilisé principalement par l’industrie du nucléaire dans quatre types d’applications : les études de protection (où l’on s’intéresse à des systèmes très fortement atténuants, voir 4.2), les études de criticité (où la simulation des systèmes critiques suppose de considérer les processus de branchement, voir 4.3), les études de physique des cœurs des réacteurs nucléaires et les études d’instrumentation nucléaire. Le code peut calculer toutes les grandeurs physiques pertinentes à ces études, comme les flux, les courants, les doses, le dépôt d’énergie, le facteur de multiplication effectif (i.e. le rapport du nombre de neutrons d’une génération à celui de la génération précédente), ou les perturbations dues à des changements en densité ou en concentration. Du point de vue de l’architecture informatique, il repose sur cinq librairies : un module de géométrie qui assure la poursuite des particules (écrit en C), une librairie de communications (C++) principalement utilisée par la librairie de parallélisation (C++), un module de gestion de la mémoire (C++), un outil de lecture des sections efficaces au format ENDF/B et ENDL (fortran) et la librairie de simulation (C++). Le code est ainsi parallélisé et porté sur des stations Linux/Unix/PC et sur des machines massivement parallèles, comme le supercalculateur TERA. Dans toute la suite de ce travail, les méthodes développées ont été implémentées dans des versions de TRIPOLI-4.4 à TRIPOLI-4.8, versions utilisées pour la simulation des cas-tests permettant de qualifier ces méthodes.

4.2

Protection et méthodes de réduction de variance

Dans les problèmes de radioprotection comme, par exemple, le dimensionnement des matériaux permettant de se protéger des rayonnements d’une installation nucléaire, le flux des particules est par définition atténué de plusieurs ordres de grandeur. Ainsi, typiquement, dans un réacteur nucléaire le flux des neutrons thermiques est d’approximativement 1013 neutrons/cm2 /seconde (cf annexe A.5) quand la cuve du réacteur permet d’avoir un flux sortant de seulement quelques neutrons par cm2 par seconde. Par conséquent, la simulation Monte-Carlo des études de protection peut paraître paradoxale puisque l’on cherche à effectuer une mesure précise (et la précision est proportionnelle au nombre de particules) dans une zone radioprotégée, c’est-à-dire avec très peu de particules. Le théorème de la limite centrale permet d’emblée de savoir que dans ces régions, la variance (ou l’erreur statistique) associée à la mesure de l’espérance du score va être très élevée. D’où l’intérêt des méthodes de réduction de variance qui permettent, à parité de particules simulées, d’avoir de meilleurs précisions.

4.2.1

Principe de base de la réduction de variance

De manière tout à fait générale, on cherche à calculer l’intégrale d’une fonction f (x) ˆ I=

b

f (x)dx. a

La méthode stochastique la plus simple consiste à échantillonner uniformément une variable X, soit xn = 1, · · · , N sur l’intervalle [a, b] puis à calculer N 1 X (b − a)f (xn ). N n=1

La loi des grands nombres assure la convergence pour N grand de cette somme vers I ˆ b N 1 X 1 (b − a)f (xn ) = E((b − a)f (X)) = (b − a)f (x) dx = I, N →∞ N b − a a n=1 lim

La variance de (b − a)f (X) , quant à elle, vaut V((b − a)f (X)) = (b − a)

E. Dumonteil

ˆ a

b

f 2 (x)dx − I 2 ,

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

59

et est d’autant plus importante que la fonction f (x) devient inhomogène. On constate que la variance est faible dans les régions où f (x) varie peu (trop d’échantillonnages) et est trop élevée (pas assez d’échantillonnages) dans les régions où f (x) est varie beaucoup. Dans la méthode de l’échantillonnage suivant l’importance, les variables aléatoires ne sont plus uniformément réparties sur [a, b] mais générées à partir d’une distribution non uniforme. L’idée est de rechercher la forme de cette distribution permettant de mieux échantillonner les régions où l’intégrale est importante. Si l’on note g(x) la densité de probabilité de cette distribution, I s’écrit :   ˆ b f (x) f (x) I= g(x)dx = Eg , (4.13) g(x) a g(x) et la variance vaut :

 Vg

f (x) g(x)

ˆ

 =

a

b



f (x) −I g(x)

2

(4.14)

g(x)dx.

Cette variance est nulle si l’on choisit g(x) = f (x)/I, on parle alors de “jeu à variance nulle”. Evidemment ceci n’est pas réalisable puisque I est justement la quantité inconnue que nous cherchons à calculer. Mais une conclusion intéressante est que l’on minimise la variance si l’on trouve une forme de g assez proche de f . On définit également ici le poids statistique d’un événement : échantillonner suivant f (x) est équivalent à échantillonner suivant g(x) des événements auxquels on attribue une poids statistique f (x)/g(x) (le lecteur intéressé par le formalisme décrivant l’estimation de I à l’aide d’un système de poids se référera à l’annexe A.3). Evidemment, lors de la simulation d’histoires de particules, le poids statistique idéal des particules est plus complexe à construire puisque celles-ci réalisent des mesures après plusieurs collisions et transports successifs. Ces histoires représentent donc des chaînes de Markov (après une collision le système de la particule “perd” la mémoire et n’est décrit que par les caractéristiques cinématiques en sortie de la collision), et la construction d’un schéma idéal de biaisage ne peut s’établir de manière si directe. C’est l’objet de la sous-section suivante que de répondre à cette question.

4.2.2

Réduction de variance pour les problèmes de transport

Après avoir introduit un outil de mesure de la qualité d’une simulation, nous présenterons deux méthodes de réduction de variance mises en œuvre dans TRIPOLI-4®. 4.2.2.1

Quantification

Pour quantifier la qualité d’une simulation, on utilise deux grandeurs : la variance sur le résultat σ 2 et le temps de simulation t. Or, l’erreur statistique d’une simulation Monte-Carlo (σ) varie en √1n , où n est le nombre de particules simulées tandis que, pour n suffisamment grand, le temps de calcul t est proportionnel à n. Ainsi, on a choisi de mesurer la qualité d’une simulation Monte-Carlo par la relation suivante, dite figure de mérite (F OM ) 1 F OM = 2 , σ t qui ne dépend ni du temps de simulation, ni du nombre de particules simulées. 4.2.2.2

Une méthode générique : la capture implicite

Elle consiste à remplacer une capture stérile par une diffusion afin de diminuer le nombre de particules « inutiles » à la simulation (i.e. qui disparaissent avant d’avoir potentiellement contribué à la mesure). Evidemment la mise en œuvre de cette technique s’accompage d’une modification du poids statistique des particules, suivant ω= 4.2.2.3

σti − σai σai = 1 − . σti σti

Le schéma de la réduction de variance dans TRIPOLI-4®

L’annexe A.2 montre que la résolution de 4.1 est équivalente à la résolution de l’équation ˆ ~ ∗ (~r, ~v ) + Σ∗t (~r, ~v )φ∗ (~r, ~v ) = Σt (~r, ~v 0 )φ∗ (~r, ~v 0 )C(~v 0 → ~v ; ~r) d~v 0 + Q∗ (~r, ~v ), ω ~ .∇φ E. Dumonteil

(4.15)

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

60

− où le flux φ(~r, ~v ) de l’équation de Boltzmann a été remplacé par un flux quelconque φ∗ (~r, ~v ) = φ(~r, ~v )= (→ r , E) → − → − → − → − (la fonction = ( r , E) réalise le mapping de φ vers φ∗ ), et où Q∗ ( r , E) = = ( r , E) S ( r , E) est la densité de − − − − − → (→ − source biaisée et Σ∗t (→ r , E) = Σt (→ r , E) − K (→ r , E) → ω (→ r , E) .− ω 0 r , E) est la section efficace totale biaisée que l’on a exprimé en fonction de → − − r ,E ) ∇=(→ → − • du paramètre de biaisage K ( r , E) = − r ,E ) =(→ → − → − r ,E ∇= ) ( − → r, E) = • de la direction d’intérêt ω0 (~ − → → − ∇= r ,E ( ) Cette remarque ne présume cependant en rien de la prévalence d’une équation par rapport à l’autre (en terme de figure de mérite des résolutions numériques Monte-Carlo associées), d’autant plus que l’équation 4.15 est définie au champ de jauge = près, a priori quelconque. Cependant, nous cherchons à améliorer à présent la variance de la simulation Monte-Carlo, et dans ce cadre on montre (voir annexe A.4) que le biaisage idéal (pouvant théoriquement permettre une mesure à variance nulle) consiste à prendre l’importance avant choc χ†e (r~0 , ~v ) ou après choc χ† (~r, ~v ) comme fonction d’importance, avec χ† (~r, ~v ) qui s’interprète comme l’espérance de la contribution au résultat d’une trajectoire démarrant en (~r, ~v ). Ainsi, conformément au biaisage suivant − l’importance tel que défini en section 4.2.1 on prouve en annexes A.3 et A.4 que si l’on prend = (→ r , E) égal à † ~0 † ∗ l’importance Σt χe (r , ~v ) on a bien la relation χe χe = =φ = φ qui doit permettre d’obtenir une variance nulle. La conclusion de ce développement est la suivante : si l’on connaît, même approximativement, l’importance − − χ† (~r, ~v ), alors la modification du noyau de transport (en prenant Σ∗t (→ r , E) à la place de Σt (→ r , E)) doit permettre d’espérer un gain sur la F OM . L’explication plus physique de cette transformation est la suivante : − − → (→ − ω (→ r , E) = − ω l’expression de Σ∗t montre que le libre parcours moyen des neutrons est plus grand lorsque → 0 r , E), c’est-à-dire lorsque les neutrons se déplacent dans la direction d’intérêt. Au contraire, s’ils se dirigent dans la direction opposée, la distance moyenne à la collision suivante est raccourcie. 4.2.2.4

Roulette russe, fractionnement et échantillonnage stratifié

Nous venons de voir qu’idéalement, on a en tout point de l’espace des phases de la simulation la relation : 1 1 ≈ χ† =

(4.16)

En fait cette grandeur représente le poids statistique idéal de la particule. Malheureusement l’importance après collision n’est toujours connue que de manière imparfaite, et il existe donc toujours un écart entre le poids statistique idéal χ1† , le poids statistique de référence =1 et le poids réel de la particule w. Or le calcul d’une trajectoire étant coûteux en terme de temps de calcul, on ne veut par conséquent pas simuler des particules de poids trop faible. On définit alors un poids de seuil, ou un poids minimal pour la simulation wmin et on tire un nombre aléatoire ξ uniformément entre [0, 1] et on applique le jeu suivant aux particules de poids w < wmin  w  , on garde la particule et on lui affecte le nouveau poids wmin ,  si ξ < wmin w   si ξ ≥ , on abandonne la particule. wmin On a ainsi défini un poids aléatoire wξ

wξ =

   wmin ,  

0,

w , wmin w si ξ > . wmin si ξ
wmax le nombre n est défini comme la partie entière du rapport   w n= . wmax E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

61

Le tirage de ξ uniformément entre [0, 1] permet de remplacer cette particule par un nombre de particules (identiques entre elles et de poids wmax ) donné par le jeu suivant  w  − n,  n + 1, si ξ < wmax w   − n. n, si ξ ≥ wmax L’augmentation du nombre de trajectoires permet de réduire la variance, mais peut également conduire à augmenter le temps de calcul. Il faut donc un bon réglage de wmax pour améliorer la F OM . La combinaison entre roulette russe et fractionnement oblige le poids des particules à rester toujours dans une intervalle [wmin , wmax ] que l’on appelle fenêtre de poids. Après le fractionnement, les trajectoires de n ou (n + 1) particules suivent la même loi pour la suite de leurs parcours. Au lieu de les générer de manière indépendante, la méthode d’échantillonnage stratifié essaie de mieux répartir ces particules dans la direction du premier transport. Cette méthode disperse les particules identiques de manière homogène sur la distance parcourue jusqu’à la collision suivante. 4.2.2.5

Utilisation de l’équation de Placzek

L’objet de cette sous-section, ainsi que de la suivante, consiste à développer des méthodes qui permettent en →. Comme − → est la direction d’intérêt, celle-ci pointe vers le détecteur sur lequel pratique le calcul de K et de − ω ω 0 0 on souhaite concentrer la mesure (attention on parle de direction d’intérêt dans l’espace des phases, c’est-à-dire que l’on tient compte de l’énergie). Pour établir K il est possible d’avoir recours à une modélisation puis une résolution du problème de manière déterministe. Ainsi, l’annexe A.4.3 détaille la manière dont on obtient la relation   Σt + K Σs ln =1 2K Σt − K dite équation de Placzek monocinétique, pour peu que l’on fasse l’hypothèse d’un problème isotrope, de type diffusif, à symétrie sphérique. Le calcul de K est alors réalisé pour chacun des matériaux par le code, et le schéma de réduction de variance repose alors sur ce pré-solveur déterministe. On notera également qu’il existe une version polycinétique pour la résolution de ces équations, permettant de relaxer les hypothèses faites et d’avoir une meilleur approximation du flux adjoint. Dans les deux cas, l’espace et les énergies sont discrétisés, afin de permettre le calcul de l’importance. On construit donc des cartes d’importances. 4.2.2.6

Réduction de variance par apprentissage

Une autre classe de méthode de réduction de variance proposée [4] peut reposer sur l’utilisation, à un moment donné de la simulation, des histoires passées de cette simulation. En effet cette information n’est pas a priori exploitée, et on peut par conséquent chercher à doter la simulation d’une capacité d’apprentissage. L’idée consiste à utiliser dans un premier temps toutes les traces/histoires des particules simulées afin de tenter d’estimer la fonction d’importance, pour utiliser cette information dans un second temps à l’aide du schéma de réduction de variance de TRIPOLI-4® présenté ci-dessus. Plus spécifiquement, on peut imaginer alimenter un réseau de neurones dans sa phase d’apprentissage par les coordonnées dans l’espace des phases (~r, ~v ) de toutes les collisions d’une particule en leur associant la contribution au détecteur de la particule. Cette méthode d’apprentissage supervisé permet à la simulation de “comprendre” les positions et les directions des particules qui contribuent à la mesure. Elle repose sur le fait que χ† (~r, ~v ) ou I s’interprètent comme l’espérance de la contribution au résultat d’une trajectoire démarrant en (~r, ~v ). En utilisant un réseau de neurones, il est possible d’apprendre pour chacune des collisions de chacune des particules l’espérance moyenne de leur contribution à la mesure en fonction des paramètres cinématiques. Après avoir introduit la classe de réseaux de neurones utilisée, et avoir discuté du mode “classification” qui nous permettra d’approcher la fonction d’importance, nous présenterons les résultats de cette méthode sur deux cas tests concrets nécessitant la mise en œuvre d’une réduction de variance. Introduction : le perceptron multicouche Les réseaux de neurones sont des outils numériques de modélisation non-linéaire par apprentissage. Les neurones “informatiques” sont calqués sur ceux des êtres humains - et de bien d’autres espèces - dans leur comportement : ils permettent d’intégrer un signal d’entrée reçu sur leurs multiples dendrites (les entrées du neurone) afin de produire une réponse donnée sur leur axone de sortie. D’un point de vue mathématique, un neurone (cf figure 4.2) est donc une fonction non-linéaire, paramétrée, à valeurs bornées. Les variables sur lesquelles il opère sont les entrées xi (i = 1, . . . , n) (reproduisant les dendrites),

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

62

et la valeur calculée par le réseau de neurones est la sortie y. Cette sortie est une fonction non-linéaire ϕ, appelée fonction d’activation d’une combinaison linéaire ν nommée potentiel des entrées. Les xi sont pondérés par les poids synaptiques wi , auxquels on ajoute un biais w0 . Ainsi, un neurone est décrit par l’équation ! n X y = ϕ(ν) = ϕ w0 + wi xi . (4.17) i=1

Figure 4.2 Représentation d’un neurone. Dans le cas de l’apprentissage réalisé pour cette étude, nous avons utilisé un réseau de neurones de type perceptron multicouche. Celui-ci est doté d’une fonction d’activation de type sigmoïde à forme logistique, c’està-dire s’écrivant 1 ϕ(ν) = , 1 + exp(−ν) ainsi que d’un réseau de neurones non-bouclé à n entrées, NH neurones cachés et NS neurones de sortie (voir la figure 4.3).

Figure 4.3 Réseau de neurones à n entrées, une couche cachée de NH neurones et NS neurones de sortie. L’intérêt des réseaux de neurones est lié aux propriétés conférées par l’association en réseau de chacun des neurones décrits ci-dessus, et comme on utilisera plus spécifiquement par la suite un réseau non bouclé à

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

63

une couche cachée à fonction d’activation logistique ϕ et avec un unique neurone de sortie linéaire, celui-ci se modélise à l’aide de la fonction g du vecteur ~x des paramètres d’entrée xi et du vecteur w ~ des poids synaptiques    NH n X X wS,i × ϕ wi0 + g(~x, w) ~ = wS,0 + wij xj  , i=1

j=1

où l’indice S signifie “sortie”. L’apprentissage sera réalisé de manière supervisée (on parle “d’entraînement”), c’està-dire que l’on cherchera à reproduire une fonction non-linéaire inconnue analytiquement, mais pour laquelle on dispose de valeurs en nombre fini qui sont entachées de bruit (modélisation statique ou de régression). La propriété importante et fondamentale des réseaux de neurones non bouclés à apprentissage supervisé que nous exploitons dans cette étude est celle de l’approximation parcimonieuse : toute fonction bornée suffisamment régulière peut être approchée uniformément, avec une précision arbitraire, dans un domaine fini de l’espace de ses variables, par un réseau de neurones comportant une couche de neurones cachés en nombre fini, possédant tous la même fonction d’activation, et un neurone de sortie linéaire. Pour un perceptron multicouche, l’algorithme d’entraînement utilise la méthode de rétro-propagation du gradient pour ajuster les poids w ~ à l’aide du gradient de la fonction de coût donnée par rapport à chaque poids. Ce coût est souvent défini comme la moyenne des différences (sur tous les neurones) entre la valeur désirée et la valeur calculée par le réseau de neurones à partir de la base d’apprentissage. La classification et sa mise en œuvre dans le cadre de la détermination de l’importance Chercher à résoudre un problème de classification signifie que l’on souhaite trouver une application de l’ensemble des objets à classer dans un ensemble de classes définies a priori et auxquelles cet objet peut appartenir. Les réseaux de neurones chargés de cette tâche s’appellent des classifieurs. Les méthodes de classification probabiliste reposent sur l’idée que l’on peut traiter les descripteurs (les entrées du réseau de neurones, qui représentent les variables potentiellement d’intérêt) et les classes comme s’il s’agissait de variables aléatoires. Le problème de la classification consiste alors à déterminer la probabilité P(Ci |x) qu’un objet (indiqué par un descripteur vectoriel x) appartienne à une classe Ci (i = 1, . . . , N ). Etant donné la probabilité a priori de chaque classe P(Ci ), que l’on peut supposer être tout simplement la proportion d’objets appartenant à la classe i dans les exemples, la formule de Bayes fournit la solution du problème de classification P(x|Ci )P(Ci ) . P(Ci |x) = PN j=1 P(x|Cj )P(Cj )

(4.18)

Cette estimation n’est correcte que si les descripteurs obéissent aux mêmes densités de probabilité conditionnelles que celles des échantillons qui ont servi à les estimer (c’est-à-dire la base d’apprentissage). On considère maintenant un classificateur à deux classes Ci (i = 0, 1). Ce classificateur va permettre de construire une sortie binaire. En effet, dans les problèmes de protection, la quantité d’intérêt est le flux de particules après atténuation par un matériau protecteur. Ce flux défini sur le domaine spatial V du détecteur peut s’écrire à l’aide de la fonction indicatrice f (~r, ~v ) = 1V (~r). (4.19) Après une simulation de N particules, on enregistre la trajectoire de chacune d’entre elles. Un traitement a posteriori va noter les trajectoires qui passent par le détecteur V dans la classe C1 et les autres dans la classe C0 . Comme la suite des positions d’une trajectoire notée (xi )i=1,...,n,... dans l’espace des phases (des positionsvitesses) après chaque transition (une collision et un tranport) est un processus Markovien, on peut considérer les points spatiaux de collision comme des sources d’émission et les xi comme des points de départ dans l’espace des phases. Donc nous possédons une base de données ayant pour entrée les xi de chacune des collisions de chacune des N particules et pour sortie la classe C0 ou C1 indiquant si la trajectoire a contribué à la mesure réalisée par le détecteur ou non. La formule de Bayes (4.18) donnée précédemment s’explique comme suit : 1. P(Ci ) (i = 0, 1) s’interprête comme la proportion de données dont la sortie est i ; 2. P(x|Ci ) (i = 0, 1) est la densité de distribution des positions de départ dans l’espace des phases pour une classe Ci ; 3. P(C1 |x) est la quantité que l’apprentissage essaie d’approcher, c’est-à-dire la probabilité pour une particule qui part d’une position x d’arriver au détecteur.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

64

Dans le problème adjoint après choc (A.10), nous avons ˆ I=

Q(x)χ† (x)dx,

où Q(x) est la source. Si I est le flux que l’on cherche au détecteur dans notre problème, nous avons vu que χ† (x) peut s’interprêter aussi comme la probabilité qu’une trajectoire dont le point de départ est x arrive au détecteur. De ce fait, P(C1 |x) est égale à l’importance après choc χ† (x) et si l’on note l’approximation de P(C1 |x) par le réseau de neurones P1 (x) on a P1 (x) ≈ χ† (x). (4.20) Cette dernière équation montre qu’il est possible d’utiliser le schéma de réduction de variance de TRIPOLI-4® qui s’appuie sur la connaissance de l’importance après choc, en redéfinissant les variables de l’équation 4.15 de la sorte ∇P ~ 1 (x) → − κ ( r , E) = P1 (x) (4.21) ~ 1 (x) ∇P − → ω0 (P ) = P1 (x)KP1 (x) Ainsi, le réseau de neurones utilise un échantillon de trajectoires de neutrons en début de simulation pour construire P1 (x) qui est directement utilisé dans le module de réduction de variance de TRIPOLI-4® comme approximation de l’importance. Comme le réseau de neurones produit une réponse continue sur son support, on peut parler de champ d’importance, à la différence de la carte d’importances discrètes en espace et en énergie (issues d’un précalcul déterministe) utilisée habituellement par TRIPOLI-4®. Résultat sur un cas test simple avec TRIPOLI-4® Un module de post-traitement des données issues de TRIPOLI-4® et fondé sur la librairie ROOT [54] a été développé. La librairie ROOT est développée par le CERN pour la physique des particules (en C++) et offre par conséquent de nombreuses possibilités aux codes Monte-Carlo de transport des particules. Elle contient notamment la classe TMLP (multi layer perceptron) qui implémente un perceptron multicouche tel que décrit ci-dessus. Ce module de traitement des données de TRIPOLI-4® porte le nom de t4ROOTtools (outils ROOT pour TRIPOLI-4®) et peut être utilisé en ligne (pendant la simulation Monte-Carlo) ou hors ligne (une fois la simulation terminée). Dans les deux cas, il utilise une structure de traces 2 permettant de décrire et de sauvegarder l’histoire des particules. C’est donc ce module qui a été paramétré de manière à pouvoir, dans une première phase de la simulation, utiliser toutes les histoires de neutrons comme base d’apprentissage du perceptron à une couche cachée. Afin de définir la fin de cette phase d’apprentissage, certaines histoires sont utilisées pour permettre le test d’hypothèses qui juge si la précision du réseau de neurones est suffisante. Si tel est le cas, la fonction définie par l’ajustement des poids du réseau de neurones est alors utilisée comme champ d’importance et doit permettre de réduire la variance du calcul. Afin de s’en assurer, deux cas tests ont été définis et simulés dans ces conditions, dont les résultats sont présentés figure 4.4. Le premier est un problème d’atténuation pure, par un bloc de fer (partie supérieure gauche de la figure 4.4). La difficulté dans l’étude de cette configuration ne réside pas dans la dépendance spatiale qui est une atténuation exponentielle en ligne droite, mais plutôt dans la prise en compte de la dégradation en énergie. Le second est un problème de contournement, dans lequel une source est placée d’un côté d’un réservoir d’eau et une barre de détection est placée sur le côté adjacent. L’ensemble est à l’air libre, qui diffuse quelques neutrons, qui peuvent alors contourner la cuve d’eau absorbante pour être détectés. La complexité de ce problème est liée à la dépendance spatiale du champ d’importance. Les résultats obtenus après apprentissage sont donnés par l’étude de la figure de mérite en fonction d’un paramètre libre du problème, noté µ. L’accélération obtenue dans le cas de l’atténuation est extrêmement probante, d’un facteur 16 environ pour µ = 3. Les résultats sont moins nets mais toujours positifs pour le problème de contournement puisque, pour une valeur unitaire de µ, la simulation par apprentissage neuronal est environ 2 fois plus rapide. Ces résultats très encourageants incitent à essayer de mieux qualifier cette méthode, afin de comprendre plus précisement ses potentialités et ses limitations et de pouvoir la rendre accessible dans les codes de production, utilisés par les industriels. 2. la “trace” d’une particule représente l’ensemble des événements permettant de décrire l’histoire de cette particule (lieux de naissance, de collision, de franchissement de volumes, énergies etc.). Le mot “trace” fait historiquement allusion aux traces laissées par les particules dans les premières chambres à bulles permettant de détecter leur passage.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

65

Figure 4.4 (a) Cas test d’atténuation. (b) Cas test de contournement.

4.3 4.3.1

Criticité et itération sur la puissance Introduction

Nous avons vu au chapitre 3 qu’il est possible de modéliser le branchement caractérisant, par exemple, le transport de neutrons provoquant des fissions. Ce domaine de la neutronique étudiant les systèmes fissiles s’appelle la criticité et permet de décrire des systèmes neutroniques dans différentes configurations critiques (des systèmes sous-critiques comme par exemple un château de transport d’assemblages combustibles usés, des systèmes critiques comme un cœur de réacteur nucléaire en fonctionnement, et sur-critiques comme des armes nucléaires ou certaines situations accidentelles). Les simulations Monte-Carlo de transport des particules en mode criticité reposent sur un principe relativement différent de celui des simulations de protection : au sein d’un cycle, les particules sont transportées normalement, mais à chaque événement de fission, les descendants sont stockés dans un tampon informatique (leurs caractéristiques cinématiques ainsi que leurs positions sont enregistrées dans ce tampon dit “buffer”). Ils sont alors utilisés comme source du cycle suivant. Ce processus itératif porte le nom d’itération sur la puissance [13, 14, 15]. Dans ces conditions, les hypothèses du théorème de la limite centrale sont alors remises en cause, puisque les cycles ne sont plus indépendants les uns des autres. Cela peut s’avérer extrêmement problématique, dans la mesure où ce théorème permet d’évaluer les intervalles de confiance des simulations de criticité utilisées dans le domaine de la sûreté (voir A.1). Heureusement, pour nombre de configurations physiques, la corrélation inter-cycles n’est forte que pendant la phase de convergence des sources (préfigurant la stationnarité, et pendant laquelle les particules explorent la géométrie). Mais il existe une classe de problèmes pour lesquels la corrélation ne disparaît pas ; ces problèmes sont dits faiblement

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

66

couplés. Cette section présente mes principaux résultats dans le domaine de la simulation des systèmes critiques faiblement couplés. Ces résultats sont majoritairement liés à la construction d’outils de diagnostic des problèmes potentiels de la simulation. Les outils de diagnostic des corrélations sont très importants pour plusieurs raisons : • d’une part la mesure d’une longueur de corrélation (en nombre de cycles) permet de regrouper les cycles en “paquets” qui deviennent ipso facto décorrélés (voir par exemple [25, 26]). On est alors reconduit à des distributions Gaussiennes et au cadre statistique habituel concernant la distribution des scores • d’autre part les corrélations induisent des biais dans l’estimation des barres d’erreurs sur les scores (voir les travaux de Gelbard par exemple [22, 23, 24]), biais qui peuvent être corrigés par une mesure directe de l’auto-corrélation (il existe différentes méthodes permettant l’estimation de la variance d’une série de valeurs corrélées) • enfin la connaissance des corrélations intéresse l’utilisateur des codes de criticité qui peut ajuster le nombre de particules à simuler en connaissance de cause (les systèmes faiblement couplés nécessitent plus de particules à parité de précision sur les résultats) Après avoir donné un exemple simple et illustratif de ce problème (une cellule de réacteur à eau pressurisé), on introduira le formalisme décrivant l’itération sur la puissance puis on présentera quatres méthodes de diagnostic : l’utilisation de l’entropie de Boltzmann pour diagnostiquer la convergence des sources (voir la référence [5, 6] de l’auteur, mais l’utilisation d’un critère entropique dans ce cadre a été originellement développée par Taro Ueki et Forrest Brown [18]), la méthode d’ajustement de la décroissance du flux neutronique pour la mesure du rapport de dominance (référence [7, 8]), celle du calcul de la transformée de Fourier pour détecter la présence d’oscillations, et enfin celle de l’autocorrélation pour mesurer directement les corrélations [10]. Il est à noter qu’il existe également des méthodes d’accélération en criticité (utilisant une modification de l’itération sur la puissance [9] ou utilisant la matrice de fission [21]) mais celles-ci se sont à ce jour avérées infructeuses et ne sont par conséquent pas présentées dans ce travail. De même, on se restreint ici uniquement aux méthodes de diagnostic de la convergence des sources et des corrélations, mais la section 5.1 du chapitre 5 fournira un cadre formel complet pour la compréhension du mécanisme à l’origine des corrélations, et proposera à ce titre des perspectives pour améliorer les simulations critiques de systèmes faiblements couplés.

4.3.2

Exemple illustratif : une cellule de réacteur à eau sous pression

Un exemple caractéristique de problème faiblement couplé est donné par l’étude de systèmes physiques dont les dimensions caractéristiques sont très supérieures au libre parcours moyen des neutrons. Ainsi, l’étude radiale d’un cœur de réacteur à eau sous pression (quelques mètres de diamètre) ou l’étude axiale d’un crayon d’un assemblage (4 mètres environ) sont des configurations typiques aux dimensions caractéristiques très supérieures au libre parcours moyen des neutrons (de quelques centimètres). Par conséquent, à des fins d’illustration, nous étudierons dans la suite une cellule simplifiée de R.E.P. 3 représentée Fig.4.5, haute de 400 cm et dont la matériau combustible au centre est composé d’U O2 enrichi à 3.25% (le rayon des pastilles est de 0.407 cm) contenu dans une gaine en zircaloy (d’épaisseur 0.07 cm) et utilisant un modérateur H2 O. Les conditions aux limites spatiales de ce problème sont de type réflexion sur les bords de la cellule et de type fuite aux extrémités. La source de neutron initiale est par défaut un pic de Dirac placé au centre.

Figure 4.5 Vue en coupe de la cellule R.E.P. simplifiée. 3. La filière des réacteurs à eau sous pression est la filière la plus répandue dans le monde. La France possède actuellement 58 de ces réacteurs.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

67

40 35

10000 20000 30000 40000 50000 60000 70000 80000 Cycle nu mbe

r

10

m

(c

20 15

)

30 25

10

2.5 2 1.5 1 0.5 0 0

z/

Flux (a.u.)

La simulation de cette configuration (figure 4.6) avec beaucoup de cycles contenant chacun peu de neutrons (105 cycles de 102 neutrons) fait apparaître la solution cosinusoïdale de l’équation de Boltzmann avec sources critiques mais également indique la présence d’oscillations du flux dans le plan z − cycles très marquées (la cellule est portée par l’axe z). Cette démarche qualitative permet donc rapidement de constater qu’il existe des corrélations inter-cycles (se manifestant dans ce cas sous forme d’oscillations). Ces corrélations brisent les hypothèses du théorème de la limite centrale, et conduisent à l’apparition de biais dans l’estimation de la variance du flux : la cellule étant symétrique par rapport à son centre en z = 200 cm, le flux devrait l’être également, ce qui n’est pas observé sur la figure 4.7 (les petites barres d’erreur sont masquées par la taille des points). En effet, cette figure indique des différences de 15 sigmas entre les valeurs des flux de certaines régions symétriques, c’est-à-dire bien au-delà des 3 sigmas acceptables. Dans cette section nous formaliserons mathématiquement ce problème en le liant à sa cause, l’itération sur la puissance, puis nous proposerons différentes solutions permettant de diagnostiquer cette autocorrélation.

5 0

Figure 4.6 Flux moyen intégré sur x et y exprimé en fonction du nombre de cycles et de z, pour 105 cycles de 102 neutrons.

E. Dumonteil

Université Paris-Sud

Flux (a.u.)

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

68

1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0

5

10

15

20 |z/10-20|

Figure 4.7 Flux en fonction de |z/10 − 20| (carrés pleins : z négatifs ; cercles vides : z positifs) pour 105 cycles de 102 neutrons.

4.3.3

L’itération sur la puissance

Conformément à l’introduction ci-dessus, c’est le suivi des neutrons génération par génération qui induit de manière naturelle un lien entre ces générations. L’avantage est de permettre à la distribution spatiale des neutrons de fission de converger vers la source réelle (on simule le processus de manière identique à son existence physique), l’inconvénient étant que la corrélation entre ces générations interdit a priori d’utiliser les générations comme mesures indépendantes d’un jeu Monte-Carlo. On notera ainsi que même si les sources sont convergées (les neutrons ayant exploré tout l’espace des phases), l’équilibre “thermique” n’est pas forcément atteint. Or c’est cet équilibre qui permettait d’assurer un découplage suffisant entre les cycles. Un paramètre physique appelé rapport de dominance (le rapport des deux premières valeurs propres de l’équation de Boltzmann avec terme de source) permet justement d’effectuer une mesure de ce découplage. Nous montrons ici comment ce paramètre intervient dans l’estimation au cycle n quelconque du kef f et du flux. L’équation de Boltzmann avec terme source s’écrit de la manière suivante [19] φ=

1 F φ avec F = (L + T − S)−1 M, kef f

(4.22)



kef f est le facteur de multiplication, première valeur propre du problème L, T, S, et M sont respectivement les opérateurs de fuite, de collision, de diffusion et de multiplication respectivement associés au courant de divergence angulaire, au taux de réaction total, à la diffusion et à la production de neutrons par fission (n+1) La méthode de l’itération sur la puissance revient à évaluer φ(n+1) et kef f en appliquant l’opérateur F au • •

(n)

flux φ(n) (et en supposant kef f et φ(n) connus à l’itération n) ˆ φ

E. Dumonteil

(n+1)

=

1 (n)

kef f

(n)

·F φ

et

(n+1) kef f

=

(n) kef f

· ˆ

dV M φ(n+1) . dV M φ(n)

(4.23)

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

69

L’équation 4.22 fait apparaître l’opérateur F dont les vecteurs propres sont notés {ui } et les valeurs propres correspondantes {ki } 1 ui = F ui avec k0 = kef f > k1 > k2 > · · · (4.24) ki On peut alors construire une récurrence permettant d’exprimer ces deux dernières grandeurs en fonction de φ(0) qui peut être lui-même exprimé sur sa base de vecteurs propres φ(0) =

∞ X

ˆ ai ui avec ai =

(4.25)

dV φ(0) ui ,

i=0

où dV représente un volume élémentaire d’espace des phases d~rd~ ω dE. Si l’on note le rapport de dominance k1 ≤ 1, k0

(4.26)

  a1 n+1 ≈ C1 u0 + DR u1 , a0

(4.27)

  a1 ≈ k0 1 + C2 DRn (DR − 1) , a0

(4.28)

DR = on obtient finalement [19] φ

(n+1)

et (n+1)

kef f

avec C1 et C2 des constantes. Cette dernière équation montre ainsi que le comportement du flux et du kef f sont très différents dans le cas de rapports de dominance élevés : en effet si le kef f tend rapidement vers sa première valeur propre, le flux est affecté par un terme en DRn . C’est ce terme qui induit un couplage éventuellement fort, dans le cas des rapports de dominance proches de 1, d’une génération à la suivante, même lorsque les sources sont convergées. Ces problèmes apparaissent principalement dans deux cas de figure : les systèmes de dimension importante devant le libre parcours moyen des neutrons (comme l’étude axiale d’un crayon combustible ou radiale d’un cœur) ou les sytèmes fortement découplés (comme le “three thick slab”, 3ème benchmark international décrit référence [29]) sont représentatifs des configurations à haut rapport de dominance. La simulation délicate de ces configurations physiques suppose de pouvoir s’assurer que les sources du problème sont convergées et que le calcul des barres d’erreur associé aux scores simulés tient bien compte de la corrélation entre les différents cycles. Ces deux étapes de diagnostic sont discutées dans les deux sections suivantes.

4.3.4

Diagnostic de la convergence des sources à l’aide de l’entropie de Boltzmann

Lors de la simulation, avant même d’établir le degré de corrélation des cycles successifs, il convient de s’assurer que les sources sont bien convergées et que les neutrons ont bien exploré tout l’espace des phases qui leur est accessible. Une méthode originale implémentée dans le code MCNP [33] a été proposée par Taro Ueki et Forrest Brown en 2003 [18], fondée sur l’utilisation de l’entropie de Shannon. Il s’agit de mailler l’espace physique à l’aide d’un réseau judicieusement choisi afin d’effectuer, pour chaque cycle, une mesure de “l’entropie des neutrons” sur ce réseau. Une variante à ce critère entropique fondé sur l’utilisation de l’entropie de Boltzmann a été proposée [5, 6] et implémentée dans TRIPOLI-4®. Cette variante présente des résultats sensiblement différents à faible nombre de particules. L’entropie de Boltzmann se présente alors comme une correction à l’entropie de Shannon, premier terme de l’équation suivante SBoltzmann = −

B X i=1

pi ln pi −

ln((2N π)B−1 2N

QB

i=1

pi )

où B est le nombre total de mailles spatiales, N le nombre total de neutronsé et pi le nombre de neutrons dans la maille spatiale i. Afin de comprendre l’intérêt de l’utilisation de ce type de diagnostic, on peut étudier par exemple les résultats de la simulation d’un château d’entreposage d’assemblages, illustré figure 4.8. Le stockage est fait de telle sorte que les éléments combustibles sont alternés avec des structures en béton de même forme. La structure périphérique est, quant à elle, faite d’eau qui réfléchit les neutrons. Comme le coin supérieur gauche de ce château abrite un assemblage radioactif, et le coin supérieur droit une structure en béton, la configuration n’est pas symétrique et la réactivité doit se concentrer dans la partie supérieure gauche car deux côtés de

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

70

l’assemblage radioactif sont en contact avec de l’eau. La figure 4.9 présente les résultats de la simulation de ce problème lorsque celle-ci est initialisée avec une source de type Dirac spatial dans le coin supérieur droit : le kef f semble converger entre les cycles 2000 et 4000 mais re-augmente de manière assez soudaine pour se stabiliser à une valeur supérieure. Cette augmentation survient quand les neutrons sont “captés” par l’assemblage le plus réactif (ayant deux côtés au contact de l’eau) : comme le nombre de neutrons est identique d’un cycle à l’autre (le code ajuste la taille de la population de particules, afin de prévenir les divergences rapides, et modifie les poids statistiques des neutrons en conséquence), les parties les plus réactives concentrent en effet toute la mesure (le nombre de neutrons). Le regroupement spatial a alors pour effet de faire chuter l’entropie qui était en constante augmentation pendant la phase d’exploration (partie inférieure de la figure 4.9). Ainsi, l’entropie s’avère-t-elle être une information supplémentaire importante pour diagnostiquer la convergence des sources spatiales : l’augmentation de l’entropie jusqu’au cycle 4000, avant sa chute rapide, permet d’éviter un jugement erroné de convergence basé uniquement sur l’analyse du kef f .

Figure 4.8 Château d’entreposage d’assemblages (benchmark OECD/NEA [32]).

Figure 4.9 kef f et entropie en fonction du nombre de cycles de la simulation du château d’entreposage.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

71

Figure 4.10 Distribution spatiale du taux de réaction de fission des neutrons dans la simulation du château d’entreposage des assemblages, en début de simulation (figure de gauche), en milieu de simulation (figure au centre) et en fin de simulation (figure de droite).

4.3.5

Diagnostic des corrélations

Une fois les sources convergées se pose la question des corrélations. Comme nous l’avons vu dans la soussection 4.3.2, ces dernières peuvent d’une part ralentir le calcul (l’information fournie par chaque cycle est moindre puisque celui-ci est fortement corrélé au précédent) et peuvent d’autre part être la source d’erreurs systématiques dans l’évaluation des incertitudes des résultats qui doivent alors être calculées en prenant l’autocorrélation des scores de la simulation en compte [22, 23]. De plus, la mesure des longueurs typiques de corrélation (en nombre de cycles) permet de mettre en œuvre la méthode des “paquets”, qui permet alors d’avoir une série de scores décorrélés. Il convient donc de quantifier le plus précisement possible ces corrélations. Pour ce faire, trois stratégies différentes ont été mises en place et sont détaillées dans cette sous-section.

4.3.5.1

Diagnostic des corrélations par l’ajustement de la relaxation du champ neutronique

La première méthode consiste à exploiter l’équation 4.27 : en effet le second membre de l’équation donne la structure du terme à l’origine des corrélations sur le flux (et donc sur les autres grandeurs dérivées), corrélations de forme DRn . Pour les rapports de dominance proches de l’unité, ce terme induit un couplage présent sur des périodes de temps (comptées en nombre de cycles) très grandes. On peut donc imaginer d’ajuster le flux (éventuellement intégré en angle et en énergie) par une fonctionnelle dépendant du nombre de cycles n. Afin d’éviter de noyer la dépendance en DR dans le bruit statistique du flux φ, on peut initialiser la simulation en excitant le champ neutronique (on place une source de Dirac spatiale, par exemple, dans un point z0 arbitraire appartenant à un matériau fissile). Ainsi, c’est l’ajustement de la décroissance de cette excitation qui permet d’extraire le DR via l’équation 4.27. Si l’on introduit le flux intégré en angle à la position z de la manière suivante ˆ ϕ(z) =

dE d~ ω φ(z, ω, E),

(4.29)

alors en tout point z = z0 on a, au cycle n n

ϕ(z0 , n) ≈ C1 (DR) + C2 = C1 exp (n ln DR) + C2 ,

(4.30)

où C1 et C2 sont des constantes fonctions de z0 . Cette expression doit permettre d’extraire le rapport de dominance d’un ajustement de la décroissance du flux. Afin de s’en assurer, un modèle analytique de la cellule R.E.P. présentée sous-section 4.3.2 a été développé [5] et permet, sous certaines hypothèses (théorie monocinétique avec géométrie simplifiée de type cylindre uniforme), de calculer une valeur approximative de DR = 0.97 pour une longueur de cellule de 500 cm. La simulation par TRIPOLI-4® de cette configuration permet, à l’aide de la méthode présentée ci-dessus, de mesurer une valeur de 0.971 ± 0.01 (voir Fig.4.11), valeur compatible avec la prédiction théorique.

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

72

Fit : n in [5,1000] DR=0.971 1.2

Flux Fit

1

Flux

0.8 0.6 0.4 0.2 0 100 200 300 400 500 600 700 800 900 1000 # cycle Figure 4.11 Calcul du DR de la cellule R.E.P. par un ajustement de la décroissance du flux. 4.3.5.2

Diagnostic des corrélations par la mesure de l’autocorrélation de rang k

Si la mesure du rapport de dominance permet de diagnostiquer rapidement d’éventuels problèmes de corrélation, il est possible de mesurer directement les corrélations afin, par exemple, de corriger l’estimation de la variance du score, ou de définir une taille des “paquets” de cycle définis en sous-section 4.3.1. Le cadre le plus général pour la mesure de ces corrélations consiste à calculer l’auto-corrélation d’une série de scores (chaque score est le résultat d’une mesure Monte-Carlo lors de la simulation d’un cycle (batch) de particules, comme par exemple le calcul du flux dans un volume donné dans un groupe en énergie donné). On utilise alors la formule pour le calcul de l’auto-corrélation de rang k de cette série de flux ϕ, c’est-à-dire de la moyenne sur tous les cycles de la corrélation entre un cycle j et un cycle j + k, donnée par Rk =

(ϕj − µ)(ϕj+k − µ) , σ2

(4.31)

où µ et σ 2 représentent respectivement l’espérance et la variance de la variable aléatoire ϕ, approchées par la moyenne et la variance estimée de la série des ϕj . Typiquement, dans les simulations de criticité, un score est corrélé sur un nombre donné de cycles (d’autant plus grand que le DR est élevé) puis devient décorrélé par la suite. Ce nombre de cycles est donc la taille du “paquet”. Le seuil d’auto-corrélation définissant la taille des paquets est arbitrairement fixé à 0.1. La mesure de l’auto-corrélation en fonction du rang k dans le cas de la cellule R.E.P. est donnée figure 4.12. Cette figure met en évidence une décroissance exponentielle (en fonction de k) de l’auto-corrélation et permet, pour le système simulé, de mesurer une taille de paquet approximativement égale à 100. Cette décroissance exponentielle peut être interprétée dans le cadre des processus dits de Gauss-Markov. En effet, le score associé à un cycle est issu d’une moyenne sur l’ensemble des particules simulées dans le cycle, et présente donc une distribution Gaussienne (si suffisamment de particules ont contribué à la mesure et si l’on se place après la convergence des sources). D’autre part, dans les simulations de criticité, un cycle de dépend que du cycle précédent (les sources du cycle n + 1 sont les neutrons de fission du cyle n) et par conséquent le score est Markovien dans la variable “cycle”. Les processus de Gauss-Markov présentant ces deux caractéristiques combinées ont une auto-corrélation de rang k à la stationnarité donnée par Rk = σ 2 e−β|k| ,

E. Dumonteil

(4.32)

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

73

où σ est la variance des ϕ2j donnée par σ 2 =E(ϕ2j ) et β est une constante de temps associée au processus étudié [34, 35]. Un ajustement de la figure 4.12 par une exponentielle indique qu’effectivement l’auto-correlation effective suit une loi de Gauss-Markov.

Figure 4.12 Autocorrélation de rang k pour la cellule R.E.P. de 400 cm simulée avec 105 cycles de 102 neutrons.

4.3.5.3

Diagnostic des oscillations à l’aide de la transformée de Fourier rapide

Dans de nombreux cas de figure, et particulièrement quand les sources sont convergées (au sens où les neutrons ont exploré tout le système et sont à l’équilibre thermique), les corrélations prennent la forme d’oscillations. Il peut alors sembler judicieux de mesurer ces oscillations par une méthode spectrale comme par exemple la transformée de Fourier rapide (F.F.T.). Des développements ad-hoc ont par conséquent été réalisés dans le module de post-traitement des données t4ROOTtools. On simule avec TRIPOLI-4® une cellule R.E.P. de 400 cm à l’aide de 105 cycles de 102 neutrons, et on calcule le flux intégré en énergie sur un réseau de 40 mailles adjacentes recouvrant toute la cellule. La F.F.T. est alors calculée pour chacun des flux dans les 40 mailles en z en utilisant la transformation usuelle ϕ∗k =

n−1 X

ϕj e−2π



−1jk/n

,

(4.33)

j=0

où les cycles j sont numérotés de 0 à n − 1 (n vaut 105 dans notre cas), et où k est la variable transformée de j. Cette transformée pour le flux calculé dans la maille correspondant au centre de la cellule (z = 200 cm) est donnée figure 4.13. L’analyse de cette figure permet de mesurer une période caractéristique d’oscillations de 100 cycles (présence d’un “pic” relatif associé à cette valeur), qui confirme l’analyse plus qualitative de la figure 4.6 (les oscillations ont une période approximativement identique). Elle met également en évidence une décroissance très rapide de l’amplitude du spectre de puissance pour les périodes plus élevées (l’amplitude est divisée par 10 en moins de 100 cycles). Conformément à l’équation 4.27 de la sous-section 4.3.3, des simulations supplémentaires ont pu mettre en évidence que cette période caractéristique ne dépend ni du nombre de particules simulées, ni de la position à laquelle la mesure est effectuée [7], et ne dépend que du rapport de dominance. On vérifie

E. Dumonteil

Université Paris-Sud

CHAPITRE 4. SIMULATION MONTE-CARLO DU TRANSPORT DES PARTICULES NEUTRES

74

de plus que cette fréquence de 100 cycles est compatible avec la taille de paquet de 100 cycles mesurée par l’autocorrélation de rang k.

2000 1800 1600 1400 1200 1000 800 600 400 200 0 10-1

1

10

102

3

10

5

104 10 Frequency

Figure 4.13 Spectre de la transformée de fourier rapide (en échelle log) du flux au centre d’une cellule R.E.P. de 400 cm simulée avec 105 cycles de 102 neutrons.

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

75

Références Articles signés ou cosignés [1] C.M.B. Diop et al., "TRIPOLI-4 : A 3D Continuous energy Monte-Carlo transport code", Transactions of the American Nuclear Society vol95 p661 (2006) [2] E. Dumonteil et al., "An Overview on the Monte-Carlo Particle Transport Code TRIPOLI-4", Transactions of the American Nuclear Society vol97 p694-695 (2007) [3] C.M.B. Diop et al., "TRIPOLI-4 : A 3D Continuous energy Monte-Carlo transport code", 1st International Conference on Physics and Technology of Reactors and Applications, Marrakech, Morroco (2007) [4] E. Dumonteil, "On a new variance reduction technique : neural network biasing. Applications on two test cases in the Monte-Carlo code TRIPOLI-4", Nuclear Technology (2009) [5] E. Dumonteil et al., "Source convergence diagnostics using Boltzmann entropy criterion : application to different OCDE-NEA criticality benchmarks with the 3-D Monte-Carlo code TRIPOLI-4" , Physics of Reactors 2006 : Advances in Nuclear Analysis and Simulations, Vancouver, Canada (2006) [6] A. L’Abbate, T. Courau and E. Dumonteil, "Monte-Carlo criticality calculations : source convergence and dominance ratio in an infinite lattice using TRIPOLI-4 and MCNP", 1st International Conference on Physics and Technology of Reactors and Applications, Marrakech, Morroco (2007) [7] E. Dumonteil and T. Courau, "Dominance ratio assessment and Monte-Carlo criticality calculations : dealing with high dominance ratio systems", Nuclear Technology, Vol172, 2, P120-131 (2009) [8] E. Dumonteil and T. Courau, "A New Method to Compute the Dominance Ratio in Monte-Carlo Simulations Application to a Simple Pin Cell with The 3-D Monte-Carlo Code TRIPOLI-4", Proceedings of ICAPP 2008, Anaheim, CA, USA (2008) [9] B. Nease and E. Dumonteil, "A decorrelation technique for iterated Monte-Carlo source calculation", Joint International Conference of the 7th Supercomputing in Nuclear Application and the 3rd Monte-Carlo, Tokyo, Japan (2010) [10] E. Dumonteil and F. Malvagi, "Variance estimation in Monte-Carlo criticality calculations", 2d International Conference on Physics and Technology of Reactors and Applications, Fez, Morroco (2011)

Autres articles [11] http ://www.oecd-nea.org/tools/abstract/detail/nea-1716/ [12] http ://root.cern.ch/drupal/ [13] Duderstadt, J.J., Hamilton, L.J., "Nuclear reactor analysis", Wiley, New York (1976). [14] Bell, G.I., Glasstone, S., "Nuclear reactor theory", Van Nostrand Reinhold Company, New York (1970). [15] Rief, H., Kschwendt, H., "Reactor analysis by Monte-Carlo ", Nuc. Sci. Eng. 30, 395-418 (1967). [16] J. Spanier, E.M. Gelbard, "Monte-Carlo Principles and Neutron Transport Problems", Addison-Wesley (1969). [17] P. Reuss, "Neutron Physics", EDP Sciences [18] T. Ueki, and F. Brown, "Stationarity and Source Convergence Diagnostics in Monte-Carlo Criticality Calculations" Proceedings of M&C 2003, Nuclear Mathematical and Computational Science : A Century in Review, Gatlimburg Tennessee, April 6-11 (2003) [19] B. Nease, F. Brown, and T. Ueki, "Dominance Ratio Calculations with MCNP" Proceedings of International Conference on the Physics of Reactors "Nuclear Power : A Sustainable Resource", Interlaken, Switzerland, September 14-19 (2008) [20] T. Ueki, F. Brown, D. Kent Parsons, and J. S. Warsa, "Time Series Analysis of Monte-Carlo Fission Sources I : Dominance Ratio Computation" Nuclear Science and Engineering, Volume 149, (2005) [21] J. Dufek and W. Gudowski, "Fission matrix based Monte-Carlo criticality calculations" Annals of Nuclear Energy, Volume 36, p1270 (2009).

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

76

[22] E. Gelbard, and R. Prael, "Computation of Standard Deviations in Eigenvalue Calculations" Progress in Nuclear Energy, Volume 24, p237 (1990). [23] O. Jacquet, R. Chajari, X. Bay, A. Nourri, and L. Carraro, "Eigenvalue Uncertainty Evaluation in MC Calculations, Using Time Series Methodologies" Proceedings of Int. Conf. Advanced Monte-Carlo for Radiation Physics, Particles Transport Simulation and Applications (Monte-Carlo 2000), Lisbon, Portugal, October 23-26 (2000). [24] T. Ueki, F. Brown, D. Kent Parsons, and D. Kornreich, "Autocorrelation and Dominance Ratio in MonteCarlo Criticality Calculations" Nuclear Science and Engineering, Volume 145, (2003). [25] F. X. Hugot, Y. K. Lee, F. Malvagi, "Recent RD around the Monte-Carlo code TRIPOLI-4 for criticality calculation" Proceedings of International Conference on the Physics of Reactors "Nuclear Power : A Sustainable Resource", Interlaken, Switzerland, September 14-19 (2008). [26] N. Petrov, N.P. Kolev, G. Todorova, F.-X. Hugot, T. Visonneau, "TRIPOLI-4 Solutions to VVER-1000 Assembly and Core Benchmarks" Proceedings of International Conference on Mathematics, Computational Methods & Reactor Physics M&C-2009, Saratoga Springs, New York, May 3-7, (2009). [27] J. P. Both, A. Mazzolo, Y. Peneliau, O. Petit, B. Roesslinger, Y. K. Lee, and M. Soldevila, "TRIPOLI-4, a Monte-Carlo Particles Transport Code. Main Features and Large Scale Applications in Reactor Physics" Proceedings of International Conference on Supercomputing in Nuclear Application, SNA 2003, Paris, France, September (2003). [28] F. Brown, "Fundamentals of Monte Carlo Particle Transport" LA-UR-05-498,LANL (2005). [29] R.N. Blomquist, A. Nouri (OECD NEA-France), "The OECD/NEA Source Convergence Benchmark Program", TANSAO 87 1-586 (2002). [30] http ://root.cern.ch/drupal/ [31] O. Jacquet, "OECD/NEA Source Convergence Benchmark 4 ; Array of interacting spheres. Revision 1 : 12/12/01", http ://www.nea.fr/html/science/wpncs/convergence/specifications/index.html. [32] N. Smith , "OECD/NEA Source Convergence Benchmark 1 ; Checkboard storage of assemblies", http ://www.nea.fr/html/science/wpncs/convergence/specifications/index.html. [33] T.E. Booth, F.B. Brown, J.S. Bull, et al., "MCNP5 1.50 Release Notes", LA-UR-08-2300, LANL (2008). [34] J.L. Doob. "The Brownian movement and stochastic equations.", Annals of Math., 43 (1942). [35] I. Karatzas and S. E. Shreve. Brownian motion and stochastic calculus. Springer-Verlag (1988).

E. Dumonteil

Université Paris-Sud

Chapitre

5

Apports de la modélisation à la simulation Sommaire 5.1

Le “clustering” des neutrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 Phénoménologie du clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.2 Modélisation et calcul des corrélations . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.3 Développement d’un outil de diagnostic pour les simulations Monte-Carlo de criticité . 5.2 Accélération des simulations de Monte-Carlo évoluant par la méthode des échantillons corrélés . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Principe de la méthode, TRIPOLI-4-D . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.2 Mise en œuvre de l’échantillonnage corrélé pour accélérer les simulations . . . . . . . . 5.2.3 Calcul des perturbations acceptables par intégrales de chemins . . . . . . . . . . . . . 5.2.4 Perspectives technologiques, big data . . . . . . . . . . . . . . . . . . . . . . . . . . . . Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

78 78 79 87 90 90 91 92 95 97

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

78

O

n a cherché, dans les deux chapitres précédents, à présenter respectivement le développement d’outils théoriques et numériques permettant la modélisation et la simulation de la marche aléatoire des neutrons dans les milieux fissiles. On a ainsi pu voir que l’équation de Feynman-Kac branchante pour les processus discrets était l’outil mathématiques ad hoc pour la modélisation de ce processus physique. Dans ce chapitre, on cherche à présent à donner deux exemples concrets pour lesquels l’utilisation de ce formalisme d’intégrales de chemin s’avère indispensable à la compréhension qualitative des résultats des simulations, ainsi qu’à la définition de leur périmètre de validité. Dans le premier exemple, on reprend la phénoménologie du problème des corrélations abordé en sous-section 4.3.5 du chapitre 4 en cherchant non pas à diagnostiquer le problème, mais à en comprendre la cause. On montrera que celle-ci est liée à un effet neutronique découvert et décrit récemment [1, 2] : le clustering des neutrons, préfiguré par des travaux amorcés en 2005 [3, 4, 5, 6, 7]. Ce dernier décrit le regroupement en amas des neutrons dans certaines conditions, et est particulièrement marqué lorsque la puissance volumique du système considéré/étudié est faible. Comme les simulations Monte-Carlo du transport des neutrons sont toujours réalisées avec des statistiques, en nombre de neutrons, inférieures sur plusieurs ordres de grandeurs à celles des réacteurs nucléaires, ces simulations mettent le phénomène en évidence très clairement. On verra que l’équation de Feynman-Kac branchante dans sa forme dégénérée de la diffusion permet aisément de décrire ce phénomène qui n’aurait pas pu l’être, par ailleurs, avec une équation de champ moyen comme l’équation de Boltzmann. Dans le second exemple, on discute des codes Monte-Carlo dits évoluants, permettant de prendre en compte les changements de composition des milieux dûs à l’irradiation sous flux. Ces méthodes ayant connu un essor récent, TRIPOLI-4® a été pourvu de cette fonctionnalité d’évolution [8] qui a depuis été testée [9] et discutée [10, 11, 12, 13]. En particulier, on présentera une méthode d’accélération des simulations Monte-Carlo des neutrons dans les milieux évoluants (sous l’effet de l’irradiation sous flux des différents nucléides constituant ces milieux) reposant sur la technique dite de l’échantillonnage corrélé. On montrera que le succès et la validité de cette méthode sont conditionnés par la variance qu’elle induit sur les résultats du calcul, qui peut être elle-même modélisée par des calculs simples d’intégrales de chemins. Ce second exemple reprend les résultats principaux [14, 15] obtenus par M. Cyril Dieudonné, dont j’ai encadré la thèse [16].

5.1

Le “clustering” des neutrons

La simulation réaliste de réacteurs nucléaires de puissance requiert l’utilisation de codes industriels à même de prendre en compte une très grande variété d’effets physiques liés au transport des neutrons, comme le ralentissement des neutrons par le modérateur, et plus généralement la prise en compte de l’énergie des neutrons, l’utilisation des sections efficaces et des données d’anisotropie contenues dans les librairies internationales d’évaluation de données nucléaires telles que ENDF/B ou JEFF [17], la description précise de géométries tridimensionnelles, etc. [18, 19]. Le code TRIPOLI-4® présenté au chapitre 4 permet de mener à bien de telles études réalistes, et est utilisé de manière routinère pour des calculs de cellules, d’assemblages ou de cœurs. Dans les années 2000, la simulation des cœurs de type N4 ou celle de réacteurs aux dimensions encore plus grandes, comme l’EPR, a cependant montré des limites : ces configurations à symétrie radiale faisaient apparaître, lors de leur simulation, des dissymétries radiales dans le calcul de la nappe de puissance ou de la distribution spatiale du flux. Ce même problème apparaît lorsque l’on cherche à calculer le flux ou la puissance axiale d’une simple cellule d’un de ces cœurs, car la longueur axiale d’une cellule est comparable aux dimensions radiales du cœur entier. Cette difficulté a été reportée par de nombreux autres codes Monte-Carlo de criticité, mais n’a jamais trouvé d’explications satisfaisantes, jusqu’à récemment. C’est ce problème qui est discuté dans la présente section.

5.1.1

Phénoménologie du clustering

Afin de mettre ce problème en exergue, on considère à présent la cellule REP présentée en sous-section 4.3.2 du chapitre précédent, décrite Fig.4.5. Cette cellule est simulée cette fois avec des conditions spéculaires de réflexion des neutrons attachées à tous ses côtés (afin d’avoir, a priori, un flux plat dans la dimension axiale), avec une source initiale uniforme en espace (ce qui permet d’accélérer l’étape de la convergence des sources) et à l’aide de 1000 cycles de 10000 neutrons. Sa hauteur L est cependant variable, et on étudie trois configurations : L = 10 cm, L = 100 cm et L = 400 cm. On s’intéresse alors à la variation en “temps” (cycle n) et à la distribution axiale (l’axe est noté x) du flux ϕ (flux intégré en énergie et en angle). Pour ce faire, on utilise un

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

79

maillage de 40 tranches en x (on désignera ces tranches par le mot anglais associé, “bin”), et on reporte dans un histrogramme les valeurs pour chaque bin i du flux ϕ à la position xi et au cycle n. Les résultats, donnés Fig.5.1, indiquent que dans le cas de la cellule caractérisée par L = 10 cm, le flux semble plat en espace et constant d’un cycle au suivant, et les fluctuations paraissent bien poissonniennes (ce qui est vérifié par un test de type χ2 ). Quand L = 100 cm, les neutrons semblent s’agréger dans certaines zones spatiales, menant à des fluctuations temporelles non-poissonniennes. Les agrégats - ou clusters - paraissent se déplacer spatialement, d’un cycle à l’autre. Ce phénomène est encore plus marqué lorsque L = 400 cm : les clusters de neutrons sont clairement distinguables, ce qui indique que les fluctuations spatiales du flux sont du même ordre de grandeur que sa moyenne. Ainsi, on retrouve bien les observations faites en sous-section 4.3.5 du chapitre 4, mais on peut ajouter à ces observations deux remarques d’importance : • les corrélations temporelles évoquées dans cette sous-section semblent trouver leur origine dans des corrélations spatiales prenant la forme de clusters de neutrons • ce mécanisme de clustering semble d’autant plus prégnant que les dimensions du système étudié sont grandes, ou que la densité de neutrons est faible (le nombre de neutrons dans chaque cycle étant fixé, la densité de neutrons et la taille du système sont inversement proportionnelles) Est-il possible d’établir un modèle à même de reproduire cet effet de clustering ? C’est l’objet de la sous-section suivante.

Figure 5.1 Simulation d’une cellule R.E.P. de taille L variable, avec 1000 cycles de 104 neutrons par cycle. Le flux de neutrons Φ(i, n) sur un maillage axial (maille i) est représenté en fonction de l’indice de la maille i et du numéro du cycle n. sur la figure de gauche, L = 10 cm, sur la figure du centre L = 100 cm, et sur la figure de droite L = 400 cm.

5.1.2

Modélisation et calcul des corrélations

Comme nous l’avons vu au chapitre précédent c’est la compréhension des mécanismes de corrélations temporelles qui a longtemps monopolisé l’attention des théoriciens du transport neutronique, et a empêché le développement de travaux sur les corrélations spatiales. Pourtant, les premières équations permettant d’établir un tel modèle de corrélations spatiales pour les neutrons dans les milieux fissiles existaient déjà il y a de cela quatre décennies. Notamment, les travaux d’Osborn et de Yip en 1966 [24], d’Athreya et de Ney en 1972 [34], de Williams en 1974 [33] puis de Pázsit et Pál jusqu’en 2008 [25] ont ouvert la voie à l’écriture d’une équation modélisant les vols exponentiels branchants, dont la forme la plus générale est établie au second chapitre de ce manuscrit et dans quelques articles récents [6, 7], et dont certaines versions sont adaptées au calcul de la variance lors de simulations Monte-Carlo [18, 26]. On cherche donc, en s’appuyant sur le chapitre 3 de ce document, à établir un modèle du transport des neutrons dans les milieux fissiles permettant de capturer la physique de la phénoménologie exposée dans la sous-section précédente, c’est-à-dire celle du clustering. Un travail bibliographique a permis d’établir qu’un tel mécanisme caractérisant les systèmes branchants existait déjà dans la littérature mathématique. En effet, les travaux de Dawson en 1972 [36] ainsi que ceux de Cox et de Griffeath en 1985 [35] ont permis de montrer qu’en dimension d = 1 et d = 2, des particules initialement dispersées aléatoirement dans un milieu infini et effectuant un mouvement Brownien branchant mènent à une concentration de la masse dans des lieux aléatoires dont les dimensions sont de mesure nulle sur le support R considéré. Ces travaux ont trouvé une continuité dans une série d’articles publiés par Houchmandzadeh dans les années 2000 [37, 38, 39] visant à décrire le clustering de bactéries dans le plan, à l’aide d’une méthode directe (à opposer à la méthode backward présentée dans ce manuscrit) ainsi que dans un article publié par Young en 2011 dans la revue Nature [40] permettant

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

80

de caractériser certaines structures spatiales en biologie des écosystèmes de phytoplancton (dans ce travail, le modèle de clustering prend le nom de Brownian bugs). La démarche permettant de modéliser les corrélations spatiales que nous allons présenter dans cette section se démarque des travaux précédemment cités de deux manières. D’une part, elle est originale dans le domaine de la neutronique où le clustering des neutrons n’avait encore jamais été reporté. D’autre part, elle s’appuie sur le formalisme backward développé au chapitre 3, ce qui doit permettre d’obtenir à terme de nouveaux résultats concernant, par exemple, la prise en compte des conditions aux limites. Afin de bien l’expliciter, différentes hypothèses sont faites quant au transport des neutrons dans les milieux fissiles : d • les neutrons évoluent dans une boite de taille L , faite d’un matériau homogène et uniforme caractérisé par sa section efficace macroscopique Σt • leur transport est de type diffusion branchante (Branching Brownian Motion défini au chapitre 3), diffusion caractérisée par le coefficient de diffusion D et par le paramètre λ = Σt v (taux auquel se produit un événement). La vitesse v des neutrons est considérée constante P∞ k • la fonction génératrice du branchement est notée comme précédemment G(z) = k=0 bk z , où les bk donnent la probabilité d’avoir k fils lors d’un événement • on s’intéressera particulièrement au cas d = 1 (dont on peut présumer qu’il caractérise une cellule axiale) et au cas d = 3 (étude d’un cœur), ainsi qu’à des milieux infinis (L → ∞). Le modèle le plus simple de branchement sera le branchement dit binaire pour lequel b0 = 12 et b2 = 12 • afin de s’assurer que ce modèle simple reproduit bien, également, la phénoménologie du clustering, un code de transport simplifié (reposant sur ces hypothèses) a été développé. L’observation des résultats de simulations utilisant ce code fait effectivement apparaître des clusters de neutrons d’autant plus marqués que le temps d’observation est grand (cf Fig.5.2) Equation pour la moyenne du nombre de particules instantané avec une particule source On cherche dans un premier temps à caractériser la concentration de particules se trouvant à une position donnée et à un temps donné, sachant que la source initiale est faite d’une seule particule. Pour ce faire, on considère l’équation de Feynman-Kac branchante 3.17 que l’on rappelle ici  1 ∂ ~ r~ Qt − s1V (r~0 , ω~0 )Qt − Σt Qt + Σt G C † {Qt } , Qt = ω~0 .∇ 0 v ∂t et qui se réécrit

 ∂ Qt = D4r~0 Qt − s1V (r~0 , ω~0 )vQt − λQt + λG C † {Qt } ∂t en utilisant l’opérateur de diffusion D précédemment introduit. En reprenant une démarche similaire à [6], on remarque que ce qui nous intéresse à présent n’est pas le comptage cumulé des distances parcourues dans un détecteur quelconque, mais uniquement le comptage instantané du nombre de particules dans ce détecteur. L’équation précédente est alors modifiée en supprimant le terme de comptage cumulé (terme reposant

sur l’utilisation de la fonction indicatrice). On obtient alors une équation pour la transformée Wt (u) = umV (t) du nombre mV de particules dans le volume V au temps t, qui est l’équation backward d’un mouvement Brownien branchant ∂ Wt = D4x0 Wt − λWt + λG (hWt iΩ ) , (5.1) ∂t où la position spatiale est à présent notée x (et la position initiale de la particule source est x0 à t = 0), et où l’opérateur de moyenne sur les directions de sorties est noté h.iΩ (ces deux changements de notation visent à alléger l’écriture). On remarque au passage que le changement de représentation pour la transformée

Qt (s) = e−slV (t) → Wt (u) = umV (t) est également associé à un changement d’opérateur pour le calcul des ∂ . Ainsi, différents moments du nombre de particules, qui s’obtiennent par simples dérivations successives ∂u u=1 si l’on cherche à calculer la moyenne du nombre mV de particules dans le volume V au temps t, notée hmV i (t), on obtient par dérivation de l’équation 5.1 ∞ X ∂ k−1 hmV i = D4x0 hmV i − λ hmV i + λ kbk hWt iΩ hmV i . ∂t u=1 k=0

Avec la condition initiale Wtk (u = 1) = 1, ∀k, et en utilisant le premier moment factoriel ν1 = hki on a finalement ∂ hmV i = D4x0 hmV i + λ(ν1 − 1) hmV i , ∂t E. Dumonteil

(5.2)

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION t=0

81

t=100

z y

z

x

y

x

t=300

t=200

y

z

z y

x

x

Figure 5.2 Evolution de 500 particules effectuant des marches aléatoires Browniennes branchantes (branchement binaire) dans un volume en 3 dimensions (pour lequel L est grand devant les distances caractéristiques sur lesquelles les particules sont absorbées). Les positions initiales sont distribuées uniformément. Quand t augmente, le mélange diffusif n’est pas assez important pour contrecarrer le “clustering” spatial dû au mécanisme de branchement, et des agrégats de neutrons se forment. équation à laquelle on doit adjoindre sa condition initiale hmV i ( t = 0| x0 ) = 1V (x0 ) et une éventuelle condition aux limites. Equation pour la concentration moyenne de particules avec une distribution initiale S de particules sources On considère à présent une distribution initiale S aléatoire de N0 particules sources sur le support de volume Ld , chaque particule se trouvant en x0 avec une probabilité p(x0 ). On peut dans ce cas relier la transformée Wt ( u| S) pour cette source initiale à la transformée Wt ( u| x0 ) pour une particule source via ˆ N0 Wt ( u| S) = p(x0 )Wt ( u| x0 )dx0 Ld

(les N0 particules étant indépendantes et identiquement distribuées selon p(x0 ), on multiplie directement les transformées à 1 particule), qui permet de calculer la moyenne du nombre mV de particules dans le volume V au temps t avec la distribution initiale S en fonction de la même grandeur pour une particule initiale ˆ ∂ hmV i ( t| S) = Wt ( u| S) = N0 p(x0 ) hmV i ( t| x0 )dx0 . (5.3) ∂u Ld u=1 Si l’on s’intéresse à la concentration volumique moyenne du nombre de particules en V au temps t pour la source S que l’on définit comme cV ( t| S) = hmV i( t|S)/Ld , et que l’on considère non plus le comptage dans un volume V E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

82

mais la concentration en x, via le passage à la limite de V vers x, on a alors c( x, t| S) = lim cV ( t| S), V→x

qui devient, à l’aide de l’équation 5.3 ˆ N0 hmV i ( t| S) = lim p(x0 ) hmV i ( t| x0 )dx0 V→x Ld Ld V→x Ld ˆ c( x, t| S) = N0 p(x0 )c( x, t| x0 )dx0

c( x, t| S) = lim

(5.4)

Ld

Cette dernière équation indique que l’on peut construire la grandeur c( x, t| S) à partir de c( x, t| x0 ). Or c( x, t| x0 ) est donnée par la solution de l’équation ∂ c( x, t| x0 ) = D4x0 c( x, t| x0 ) + λ(ν1 − 1)c( x, t| x0 ), ∂t

(5.5)

obtenue en prenant le passage à la limite de V vers x de l’équation 5.2 et en la divisant par le volume Ld . Solution de l’équation pour la concentration moyenne de particules avec une distribution initiale uniforme de particules sources Afin de donner une solution à cette dernière équation, il convient de lui adjoindre sa condition initiale c( x, t = 0| x0 ) = δ(x − x0 ), ainsi qu’une condition aux limites du système. Comme on s’intéresse à un système de grande dimension L, on souhaite en fait étudier un système sans bords (c’est-à-dire une boite infinie). La solution générale de 5.5 s’écrit alors [43] eλ(ν1 −1)t − (x−x0 )2 e 4Dt , c( x, t| x0 ) = √ 4πDt

(5.6)

en dimension d = 1 (les développements qui suivent sont fait en une dimension, et on généralisera les résultats à d quelconque au paragraphe 5.1.2). Dans un milieu de taille finie, la concentration initiale c0 est définie comme le rapport du nombre de particules initiales sur les dimensions du volume (c0 = N0 /Ld ). En milieu infini, cette définition se généralise, pour notre source initiale uniforme, à c0 = N0 p(x0 ), ce qui permet de trouver la solution pour la concentration avec une source initiale de particules S uniforme en écrivant ˆ +∞ c( x, t| S) = N0 p(x0 )c( x, t| x0 )dx0 , ˆ

= c0

−∞ +∞

c( x, t| x0 )dx0 , −∞ λ(ν1 −1)t

= c0 e

(5.7)

.

Cette solution pour la moyenne de la concentration en x au temps t traduit le fait qu’en milieu infini, la concentration moyenne de particules diverge à grands temps si ν1 > 1, reste constante si ν1 = 1 et tend vers 0 si ν1 < 1. Du point de vue du neutronicien, cela s’explique bien puisqu’en milieu infini, il n’y a pas de terme de fuite, et par conséquent le kef f du système est donné par le nombre moyen de neutrons, et on a donc kef f = k∞ = ν1 . Calcul des corrélations Afin d’avoir accès aux corrélations spatiales, l’équation de Boltzmann (obtenue par application de l’opérateur de moyenne sur l’équation de Feynman-Kac branchante) seule n’est plus suffisante. L’idée consiste donc à formuler l’équation pour les corrélations entre deux volumes notés Vu et Vv , puis à résoudre cette équation pour V s’exprimant comme

en extraire les corrélations. La transformée de Laplace pour un volume

Wt (u) = umV (t) , sa généralisation à deux volumes s’écrit trivialement Wt (u, v) = umVu (t) v mVv (t) , où u et v sont respectivement les variables transformées du nombre de particules en Vu et du nombre de particules en Vv à l’instant t. Cette fonction génératrice permet de construire les moments successifs des nombres de particules

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

83

en Vu et en Vv pour une particule source en x0 , ainsi que les corrélations entre ces grandeurs, puisque l’on vérifie que d’une part la moyenne de ces nombres de particules est donnée par ∂ hmVu i ( t| x0 ) = , Wt ( u, v| x0 ) ∂u u=1,v=1 ∂ hmVv i ( t| x0 ) = , Wt ( u, v| x0 ) ∂v u=1,v=1 et que d’autre part les corrélations à deux volumes sont données par ∂2 hmVu mVv i ( t| x0 ) = . Wt ( u, v| x0 ) ∂u∂v u=1,v=1 L’équation de Feynman-Kac branchante pour les corrélations d’un processus diffusif se généralise directement à partir de sa forme à un volume, et s’écrit ∂ Wt ( u, v| x0 ) = D4x0 Wt ( u, v| x0 ) − λWt ( u, v| x0 ) + λG (hWt ( u, v| x0 )iΩ ) . ∂t

(5.8)

La généralisation de cette équation à des processus de transport (où l’opérateur de diffusion D∆x0 est remplacé ~ r~ ) porte en neutronique le nom d’équation de Pál-Bell. Pour calculer les par l’opérateur de transport ω~0 .∇ 0 ∂2 sur l’équation 5.8 ce qui donne corrélations à deux volumes, on fait à présent agir l’opérateur ∂u∂v u=1,v=1

X ∂ ∂2 k hmVu mVv i = D4x0 hmVu mVv i − λ hmVu mVv i + λ bk hWt ( u, v| x0 )iΩ , ∂t ∂u∂v u=1,v=1 k

∂ hmVu mVv i = D4x0 hmVu mVv i − λ hmVu mVv i + ∂t  2   X ∂ k−1 + λ bk k hWt ( u, v| x0 )i hWt ( u, v| x0 )iΩ ∂u∂v u=1,v=1 k      X ∂ ∂ k−1 λ bk k (k − 1) hWt ( u, v| x0 )iΩ hWt ( u, v| x0 )iΩ hWt ( u, v| x0 )iΩ , ∂v ∂u u=1,v=1 k

et par application de la condition initiale Wtk (u = 1, v = 1) = 1, ∀k, X ∂ hmVu mVv i = D4x0 hmVu mVv i − λ hmVu mVv i + λ bk k [hmVu mVv i + (k − 1) hmVv i hmVu i] , ∂t k

∂ hmVu mVv i = D4x0 hmVu mVv i − λ hmVu mVv i + λν1 hmVu mVv i + λν2 hmVv i hmVu i , ∂t où l’on a utilisé la définition des deux premiers moments factoriels ascendants. La présence du second moment factoriel ν2 = hk(k − 1)i du processus de branchement traduit le fait que les corrélations entre volumes sont affectées justement par ce terme, le premier moment factoriel n’étant plus suffisant (on peut d’ailleurs s’étonner de ce que les moments d’ordre supérieur n’influent pas sur ces corrélations...). On a finalement l’équation pour les corrélations ∂ hmVu mVv i = D4x0 hmVu mVv i + (λ − ν1 ) hmVu mVv i + λν2 hmVv i hmVu i . ∂t

(5.9)

La démarche consiste alors à faire le lien entre ces corrélations à deux volumes pour une particule source et les corrélations à deux volumes pour une source distribuée via (cf paragraphe précédent) ˆ N0 Wt ( u, v| S) = p(x0 )Wt ( u, v| x0 )dx0 Ld

∂ hmVu i ( t| S) = Wt ( u, v| S) ∂u

E. Dumonteil

ˆ = N0

u=1,v=1

Ld

p(x0 ) hmVu i ( t| x0 )dx0 ,

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION ˆ ∂ hmVv i ( t| S) = = N0 Wt ( u, v| S) p(x0 ) hmVv i ( t| x0 )dx0 , ∂v Ld u=1,v=1

84

ˆ  ∂2 = N0 (N0 − 1) p(x0 ) hmVu i hmVv i ( t| x0 )dx0 + Wt ( u, v| S) hmVu mVv i ( t| S) = ∂u∂v Ld u=1,v=1 ˆ  N0 p(x0 ) hmVu mVv i ( t| x0 )dx0 . (5.10) Ld

Si l’on divise l’équation 5.9 par le volume, pour transformer les nombres de particules en concentrations volumiques, et que l’on applique le passage aux limites Vu → x et Vv → y, on obtient finalement l’équivalent de l’équation 5.5 pour les corrélations ∂ c( x, y, t| x0 ) = D4x0 c( x, y, t| x0 ) + λ(ν1 − 1)c( x, y, t| x0 ) + λν2 c( x, t| x0 )c( y, t| x0 ), ∂t

(5.11)

où c( x, y, t| x0 ) est la valeur moyenne du produit des concentrations en x et en y (quand c( x, t| x0 )c( y, t| x0 ) représente le produit des valeurs moyennes des concentrations). Comme précédemment, on adjoint à cette équation sa condition initiale c( x, y, t = 0| x0 ) = δ(x − x0 )δ(y − x0 ). Le passage à une source distribuée se fait quant à lui en appliquant la même opération à l’équation 5.10 afin d’obtenir l’équivalent de la première des équations de 5.7 ˆ  ˆ  c( x, y, t| S) = N0 (N0 − 1) p(x0 )c( x, t| x0 )c( y, t| x0 )dx0 + N0 p(x0 )c( x, y, t| x0 )dx0 , Ld

Ld

soit, pour N0 grand ˆ c( x, y, t| S) ' c( x, t| S)c( y, t| S) + N0

 p(x0 )c( x, y, t| x0 )dx0 .

Ld

Afin de tenir compte de la “condition aux limites” dans un volume infini, on impose L → ∞ à cette dernière équation. La solution de l’équation aux dérivées partielles Eq.5.11 est, comme pour le calcul de la moyenne, référencée dans la littérature mathématique [43] et permet de donner c( x, y, t| S) via c( x, y, t| S) = c( x, t| S)c( y, t| S) + δ(x − y)c( x, t| S)+ ˆ

ˆ



N0 λν2



0

dx0 p(x0 ) −∞

ˆ

t

dt 0

−∞

e dx00 c( x, t0 | x0 )c( y, t0 | x0 ) p



(x00 −x0 )2 4D(t−t0 )

0

4πD(t − t0 )

e−λ(ν1 −1)(t−t ) .

Comme la fonction de corrélation de paire g(x, y, t) d’un système ne contient par construction que les corrélations, elle est définie comme g(x, y, t) =

c( x, y, t| S) − c( x, t| S)c( y, t| S) − δ(x − y)c( x, t| S) c( x, t| S)c( y, t| S)

(valeur moyenne du produit des concentrations en x et en y, auquel on soustrait le produit de la moyenne des concentrations en x et en y ainsi que l’autocorrélation en x = y, le tout étant normalisé). Cette fonction vaut donc, pour notre système N0 λν2 g(x, y, t) = c( x, t| S)c( y, t| S)

ˆ

ˆ





0

dx0 p(x0 ) −∞

ˆ

t

dt 0

−∞

e dx00 c( x, t0 | x0 )c( y, t0 | x0 ) p



(x00 −x0 )2 4D(t−t0 )

4πD(t −

0

t0 )

e−λ(ν1 −1)(t−t ) .

En utilisant l’équation 5.6, en permutant les intégrales en espaces et en développant le calcul on montre finalement que cette fonction de corrélation peut être exprimée en fonction de c( x, t| S) via λν2 g(x, y, t) = c( x, t| S)c( y, t| S) E. Dumonteil

ˆ

t

0

ˆ



dt0 e2λ(ν1 −1)t 0

−∞

e− dx00 c( x00 , t − t0 | S) √

(x−x00 )2 4Dt0

4πDt0

e− √

(y−x00 )2 4Dt0

4πDt0

.

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

85

En injectant dans cette expression la solution pour la moyenne de la concentration donnée par la dernière équation de 5.7, et en développant le calcul on peut finalement exprimer la fonction de corrélation de paire en fonction de la distance entre x et y g(r = |x − y| , t) =

λν2 c0

ˆ

t

0

r2

e− 8Dt0 λ(ν1 −1)(t0 −t) dt0 √ e . 8πDt0

Généralisation à un espace de dimension quelconque Ld Par analyse dimensionnelle ou en reprenant les calculs en remplacant L par Ld , on montre que cette dernière expression se généralise en dimension d quelconque par ˆ r2 0 λν2 t 0 e− 8Dt0 dt eλ(ν1 −1)(t −t) . (5.12) g(r = |x − y| , t) = d/2 c0 0 (8πDt0 ) On distinguera par la suite trois “types” de clustering : • le “clustering mathématique” qui indique une divergence de cette fonction de corrélation de paire pour r → 0. Ce clustering est celui auquel fait référence la littérature mathématique. Il caractérise un système pour lequel la masse (l’ensemble des particules) se concentre en des points de mesure nulle sur le support. Cette caractéristique peu “physique” est liée au mécanisme de diffusion : les trajectoires Browniennes sont faites de collisions en tous points de la trajectoire, ce qui permet effectivement d’agréger la masse dans des “points” à grands temps. Une autre hypothèse forte requise pour l’émergence de ce mécanisme est le nombre infini de particules initiales : si la concentration c0 est finie, comme on se place en milieu infini, le nombre de particules initiales doit être également infini • le “clustering physique” qui est le clustering auquel on fait allusion au sens littéral du terme : il apparaît en effet des “trous” et des “agrégats/clusters” dès lors que g(r, t) ≈ 1. Cette dernière condition caractérise effectivement un système pour lequel les fluctuations spatiales sont du même ordre de grandeur que la moyenne • le “mécanisme de clustering” qui est la définition la plus lâche que l’on puisse donner : si des “clusters” ne se forment pas, les corrélations spatiales liées au mécanisme de clustering “déforment” la solution moyenne d’un problème donnée par l’équation de Boltzmann et doivent être par conséquent prises en compte Etude du branchement binaire (critique en milieu infini) pour d = 1 et d = 3 Si l’on considère le cas simple d’un branchement à deux fils critique (qui doit donc être binaire en milieu infini, c’est-à-dire tel que b0 = 1/2 et b2 = 1/2), on a ν1 = 1 et ν2 = 1. La dépendance en temps disparaît donc, ce qui traduit que le système est exactement critique et que la masse totale ne fluctue pas (en moyenne). L’équation 5.12 s’exprime dans ce cas analytiquement puisque l’on a λ g(r, t) = c0

ˆ

r2

t 0

dt 0

e− 8Dt0 (8πDt0 )

d/2

,

qui peut se réécrire sous la forme g(r, t) =

λ d r2 π −d/2 r2−d Γ( − 1, ), 8Dc0 2 8Dt

où Γ désigne la fonction gamma. cas d = 1 La partie gauche de la figure 5.3 représente la fonction de corrélation de paire ci-dessus, en dimension d = 1. Initialement plate, cette fonction croît rapidement en fonction de t, pour des petites valeurs de r. Particulièrement, en r = 0, la fonction de corrélation de paire diverge à grands temps. Cette dernière caractéristique représente le clustering, au sens mathématique du terme. Il traduit le fait que pour une distribution initiale uniforme de la masse sur le support, en dimension 1 (ce qui est également vrai en dimension 2), la masse tend à se regrouper dans certaines positions aléatoires. De manière encore plus surprenante, la divergence en temps de la fonction en r = 0 indique que l’échelle spatiale sur laquelle est distribuée cette masse tend vers 0 ! Les papiers de Cox évoqués en introduction montrent que le support sur lequel est distribuée la masse, à grands temps, est de mesure nulle. Qualitativement, en dimension 1, un ensemble de particules initialement distribuées aléatoirement et uniformément sur un support infini et mues par un mouvement Brownien branchant développe

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

86

toujours en temps ce phénomène d’agrégation des particules qui peut sembler très contre-intuitif puisque les particules n’interagissent pas entre elles. Ce phénomène traduit que les fluctuations en espace sur la masse créent des “trous”, et que dès lors que ces régions spatiales sont vidées de leur contenu en particules, le mécanisme de branchement et de transport ne parvient plus à les repeupler. cas d = 3 En dimension d = 3, les résultats présentés sur la partie droite de la figure 5.3 sont qualitativement très différents. En effet, la fonction de corrélation de paire croît rapidement pour les faibles valeurs de r, λ . mais sature à grands temps. La valeur limite atteinte par la fonction est ainsi contrôlée par le préfacteur 8Dc 0 Physiquement, cela indique qu’il n’y a pas, d’un point de vue mathématique, de clustering dans la mesure où la fonction de corrélation ne diverge pas (on utilisera tout de même ce terme par la suite puisqu’il est, au sens propre, exact en ce qu’il désigne une forme d’agrégation des neutrons entre eux). Cependant, ce phénomène invalide les résultats obtenus par l’étude de l’équation de Boltzmann (valable pour un nombre de particules infini), et est d’autant plus marqué que • λ, le taux auquel se produit un événement de branchement (absorption, diffusion, fission) est élevé. En effet, dans la limite où le mécanisme de branchement devient trop faible, le comportement des neutrons est simplement balistique. • D est faible, ce qui signifie que les particules ont tendance à peu diffuser, et donc à rester sur place. Dans ce cas, le transport des particules ne parvient plus à mélanger suffisamment les particules (il n’y pas de “brassage” spatial). • c0 est faible, puisque de fortes concentrations initiales des particules impliquent des distances moyennes inter-particules faibles au départ, ce qui prévient donc la formation d’agrégats. Ces trois paramètres contrôlent donc l’intensité du clustering pour d = 3, alors que pour d = 1 ils ne contrôlent que le temps caractéristique au bout duquel le phénomène apparaît (ce temps caractéristique d’apparition du clustering en une dimension est inversement proportionnel à λ, et proportionnel à D et c0 ).

Figure 5.3 Fonction de corrélation de paire g(r, t) en dimension d = 1 (à gauche) et d = 3 (à droite)

Discussion des résultats dans le contexte de la neutronique et des simulations Monte-Carlo L’étude du branchement binaire critique associé à un mouvement Brownien branchant a permis de fournir un cadre explicatif au mécanisme d’agrégation/de clustering des particules observé en sous-section 5.1.1. Dans le contexte de la neutronique, on ne peut plus parler de “clustering mathématique” mais uniquement de “clustering physique” ou faire allusion simplement à un “mécanisme de clustering” (voir la définition de ces termes donnée au paragraphe 5.1.2). En effet, en neutronique deux hypothèses fortes du modèle sont brisées, qui empêchent d’utiliser le terme de “clustering mathématique” • les neutrons ne “diffusent” pas mais sont “transportés” (distinguo entre équation de la diffusion et équation du transport). Les résultats du modèle sont donc qualitatifs en ce sens qu’ils ne sont valables que loin des interfaces (traitement du combustible fissile uniquement), pour des milieux isotropes et des milieux denses • les neutrons sont en nombre fini (éventuellement très grand) Concernant le “clustering physique”, celui-ci peut être discuté selon qu’on le considère en physique des cœurs de réacteurs nucléaires, à proprement parler, ou dans les simulations Monte-Carlo de la physique des cœurs. Dans les simulations Monte-Carlo de criticité pour la physique des cœurs, les systèmes étudiés sont simplifiés puisque l’on ne les considère que d’un point de vue neutronique (pas d’effets mécaniques ou thermo-hydrauliques

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

87

par exemple). Ces simulations font donc des hypothèses très proches de celles qui ont été proposées dans la sous-section précédente, et le modèle théorique doit permettre de modéliser assez finement les résultats des simulations (cela sera discuté dans la sous-section suivante), hormis pour un paramètre : c0 . En effet, la caractéristique essentielle des simulations Monte-Carlo, comparativement à des configurations réalistes, est qu’il est actuellement encore impossible de simuler autant de neutrons que dans la réalité puisque typiquement, un réacteur nucléaire nécessiterait de simuler 1015 neutrons par secondes (voir l’appendice A.5). Par conséquent le mécanisme de clustering est beaucoup plus fort dans les simulations Monte-Carlo des cœurs nucléaires qu’il ne l’est dans les cœurs réels, car les valeurs de c0 associées aux simulations Monte-Carlo sont comparativement très faibles. En physique des réacteurs nucléaires, il est très dur de se prononcer a priori sur l’intensité de ce mécanisme, et ce pour plusieurs raisons : • il existe de nombreuses contre-réactions tendant à stabiliser la puissance dans un cœur (contre-réactions thermo-hydrauliques, effets thermiques de type Doppler-broadening etc.) • il est très dur de qualifier la dimensionnalité des composants des cœurs : une cellule REP est-elle un système 1d ou 3d ? Dans le cas d’un cœur, le système est-il vraiment 3d ou correspond-il plutôt à un ensemble de cellules 1d communiquant faiblement ? • les effets de géométrie doivent être pris en compte (tel que la présence de matériau modérateur non-fissile, etc.) • les neutrons retardés ont sûrement un impact sur la dynamique du système De plus, que l’on considère la physique des cœurs ou les simulations Monte-Carlo de criticité, les systèmes étudiés sont complexes et par conséquent la fonction de corrélation dépend de l’espace : il n’est pas suffisant de considérer g(r, t), il faut calculer g(x, y, t). Toutes ces restrictions laissent penser qu’une modélisation purement théorique risque de s’avérer compliquée à mettre en œuvre. Cependant, la connaissance de cette fonction est très importante car même si un système neutronique ne fait pas apparaître de “clustering physique”, il n’en reste pas moins que le “mécanisme de clustering” tend à invalider la pertinence d’une approche de type Boltzmann uniquement, et ce particulièrement dans un régime pour lequel c0 est faible (démarrage du réacteur, extinction, fonctionnement en fraction du régime nominal, etc.). Il est donc important de définir une stratégie permettant le calcul numérique de cette fonction. En physique des cœurs et en neutronique, deux options sont ouvertes : soit on cherchera à écrire une équation pour la fonction de corrélation pour une configuration réaliste qui prendrait en compte tous les effets sus-cités, soit on simule cette configuration à l’aide d’un code Monte-Carlo de transport des neutrons, avec un nombre réaliste de neutrons. La sous-section suivante finalise les considérations sur le clustering en discutant ce point.

5.1.3

Développement d’un outil de diagnostic pour les simulations Monte-Carlo de criticité

Estimation de la fonction à deux points Les simulations Monte-Carlo de criticité réalisées à l’aide de TRIPOLI-4® permettent de sonder le comportement du système à nombre de particules fixé. Si ce nombre de particules est réaliste, on mesure des grandeurs a priori accessibles à la mesure. S’il ne l’est pas, on accède uniquement à des valeurs moyennes et les effets tels que les corrélations spatiales et le clustering font alors figure d’artefacts indésirables. Dans les deux cas, il est important de pouvoir avoir accès aux corrélations spatiales. Une stratégie simple pour estimer la fonction de corrélation à 2 points lors d’une simulation consisterait à enregistrer, lors d’un cycle donné, toutes les histoires des neutrons (on rappelle ici que le cycle n fait office de variable de temps t). On reporterait alors dans un histogramme toutes les distances inter-collisions r pour tous les neutrons. En pratique, cette stratégie s’avère extrêmement coûteuse en temps de calcul : par exemple, pour un cycle de 108 neutrons effectuant chacun 102 collisions, le nombre de distances à calculer serait égal au nombre de paires 2 15 possibles dans un ensemble de cardinal 108 , c’est-à-dire C10 ... On choisit donc une seconde option : 8 ' 5.10 la mesure de l’entropie des sources présentée en sous-section 4.3.4 du chapitre précédent repose sur l’utilisation d’un maillage en espace. On peut calculer, au cycle n, le flux de neutrons intégré en énergie et en angle φ(i, n) sur chaque bin i de ce maillage, et utiliser un estimateur g˜(r, n) de la fonction à 2 points que l’on écrit

  ¯ ¯ φ(i, n) − φ(n) φ(i + r, n) − φ(n) i , g˜(r, n) = σ 2 (n) ¯ où r est la distance entre deux bins, h.ii désigne la moyenne sur tous les bins i, φ(n) est le flux moyenné sur 2 tous les bins i, et σ (n) est la variance estimée du flux sur les positions spatiales des bins. On observera que cet estimateur repose sur le flux des neutrons, à la différence de la fonction à 2 points, mais le flux de neutrons

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

88

au bin i est directement proportionnel au nombre de neutrons de ce bin. Ainsi, si l’on fait l’hypothèse d’une certaine ergodicité du système (on moyenne sur des positions en espace plutôt que sur différentes réalisations du système), on peut montrer que g˜(r, n) permet d’estimer g(r, t). Afin de mettre en œuvre ces considérations, la fonction g˜(r, n) a été implémentée au sein du code TRIPOLI4® et estimée lors de la simulation de la cellule REP présentée en sous-section 4.3.2 et au début de ce chapitre. La figure 5.4 rappelle, sur la partie de gauche, la fonction à 2 points théorique pour un système infini en 1 dimension, et donne, sur la partie de droite, la fonction à 2 points estimée par la méthode décrite ci-dessus. On constate un très bon accord qualitatif entre ces deux fonctions, qui présentent toutes les deux un profil spatial initial uniforme, puis qui augmentent particulièrement aux faibles valeurs de r jusqu’à arriver à saturation. Sur la fonction à 2 points estimée par la simulation, on remarque cependant que certaines valeurs de ge sont négatives : cela est dû aux méthodes de renormalisation du nombre de neutrons d’un cycle au suivant qui permettent de stabiliser la population des neutrons, et qui induisent ipso facto des anti-corrélations.

Figure 5.4 Fonction de corrélation à 2-points gt (x, y) représentée en fonction de la coordonnée spatiale r = |x − y| pour un système 3-D non borné. La fonction fait apparaître un pic en 0 dont l’amplitude croît avec le temps puis sature. A gauche : fonction théorique. A droite : fonction empirique mesurée par la simulation TRIPOLI-4® de la cellule R.E.P. de 400 cm. Dans cette dernière fonction, le temps est mesuré en nombre de cycles n et l’espace en nombre de mailles spatiales i.

Recommandation pour les utilisateurs de codes Monte-Carlo de criticité Dans la pratique, les utilisateurs de codes Monte-Carlo de criticité peuvent ajuster le nombre de neutrons simulés N0 afin de prévenir la formation de clusters, et l’apparition ipso facto d’asymmétries affectant les résultats des simulations (attention, le nombre de neutrons simulés doit ici s’entendre comme le nombre de neutrons simulés par cycle). Afin de construire un ordre de grandeur pour connaitre a priori ce nombre de neutrons, on peut remarquer que la distance caractéristique l sur laquelle ceux-ci se déplacent est le libre parcours moyen avant absorption donné par p l = hra2 i,

où ra est la distance d’absorption [42, 41]. Ainsi, si le système a une dimension typique notée L, la condition sur N0 permettant de s’assurer qu’il n’y a aucun “trous” dans la distribution spatiale des neutrons est N0 ≈ (L/l)3 .

(5.13)

Dans l’exemple considéré Fig.5.1, le libre parcours moyen des neutrons est de l ≈ 6 cm, et par conséquent afin de ne pas voir le clustering se développer il faut • N0  10 pour un crayon de hauteur L = 10 cm. Cette condition était vérifiée (les simulations étaient réalisées avec 104 neutrons par cycles), et effectivement le clustering des neutrons n’a pas eu lieu 3 • N0  5.10 pour un crayon de hauteur L = 100 cm. Ce nombre étant du même ordre de grandeur que celui utilisé pour les simulations, on s’attend à ce qu’à cette valeur les premières corrélations soient apparues, ce qui était bien le cas 5 • N0  3.10 pour un crayon de hauteur L = 400 cm. Cette condition n’était pas satisfaite, et comme prévu des clusters de neutrons sont apparus

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

89

On retiendra donc par conséquent que la condition donnée par l’équation 5.13 fournit un critère quantitatif simple permettant à l’utilisateur de s’assurer que l’effet de clustering n’induiera pas, dans les simulations, de corrélations spatiales.

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

5.2

90

Accélération des simulations de Monte-Carlo évoluant par la méthode des échantillons corrélés

Cette dernière sous-section, très courte, porte sur une méthode originale d’accélération des simulations de Monte-Carlo évoluant par la méthode des échantillons corrélés. Cette méthode a été mise en œuvre dans la thèse de doctorat de M. Cyril Dieudonné, thèse dirigée par le professeur C. Diop et dont j’ai été l’encadrant, et a donné lieu à deux publications dans des proceedings de conférences, ainsi qu’à une publication dans la revue Annals of Nuclear Energy. Par conséquent on ne cherchera pas ici à être exhaustif dans la présentation des résultats qui sont donnés avec force détails dans le manuscrit de thèse de M. Dieudonné par exemple, mais plutôt à expliquer les grandes lignes de la méthode et à montrer l’utilité de l’approche par intégrales de chemins pour en encadrer l’utilisation. On verra finalement que cette méthode peut être généralisée à d’autres types de couplages multi-physique et se positionne naturellement dans la thématique très contemporaine du “big data”.

5.2.1

Principe de la méthode, TRIPOLI-4-D

Système d’équations couplées Les codes Monte-Carlo dits “évoluants” visent à résoudre l’équation du transport en prenant en compte l’évolution temporelle des concentrations isotopiques due à l’irradiation sous flux des compositions. L’idée consiste à résoudre un système d’équations couplées : l’équation de Boltzmann et les équations de Bateman. En effet, l’équation du transport, que l’on rappelle ici (cf Eq.4.1) ˆ ~ r, ~v ) + Σt (~r, ~v )φ(~r, ~v ) = d~v 0 Σt (~r, v~0 )φ(~r, v~0 )C(v~0 → ~v ; ~r) + Q(~r, ~v ), ω ~ .∇φ(~ permet de calculer le flux à un temps t donné en supposant connues les concentrations initiales isotopiques − N i (→ r , t) (qui apparaissent indirectement dans l’équation précédente via Σt , C et généralement Q, grandeurs qui dépendent toutes les trois des sections efficaces macroscopiques). Les équations de Bateman, dites équations − d’irradiation sous flux, permettent quant à elles de calculer l’évolution des concentrations isotopiques N i (→ r , t) sur un intervalle de temps d’irradiation dt, en supposant que le flux (ou les taux de réaction) est connu. Ces équations s’écrivent X X dN i → − − − − − − (− r , t) = τj→i (→ r , t) N j (→ r , t) + λj→i N j (→ r , t) − τdi (→ r , t) N i (→ r , t) − λi N i (→ r , t) , dt j6=i

(5.14)

j6=i

où − i → • N ( r , t) est la concentration du nucléide i au point ~ r et à l’instant t (en nombre d’atomes par 10−24 cm3 ) → − → − • τj→i ( r , t) est le taux de réaction microscopique d’alimentation du nucléide j vers le nucléide i : τj→i ( r , t) = ´ → − −1 σj→i (~v ) φ (~r, ~v , t) d v (en s ) ´ i → − − − i → i → • τ ( r , t) est le taux de réaction de disparition microscopique du nucléide i : τ ( r , t) = σd (− v ) φ (~r, ~v , t) d→ v d d −1 (en s ) → − • σj→i ( v ) est la section efficace microscopique d’alimentation du nucléide j vers le nucléide i (en b) − i → • σ ( v ) est la section efficace microscopique de disparition du nucléide i (en b) d −1 • λj→i est la constante de décroissance radioactive de j vers i (en s ) −1 • λi est la constante de décroissance radioactive de i (en s ) Elles se couplent à l’équation du transport via les taux microscopiques de réaction et de disparition, qui sont des fonctions du flux solution de l’équation du transport (les solveurs d’équation du transport permettent le plus souvent de calculer soit les flux soit les taux microscopiques directement). Si les équations d’évolution peuvent être résolues de manière analytique lorsque le flux est nul, en toutes généralités la résolution de ces équations se fait à l’aide de solveurs déterministes de type Runge-Kutta ou exponentielle de matrice [52]. Plusieurs codes existent qui permettent de résoudre ce système d’équations couplées, comme MOCUP [44], MVP[45], MCU[46], MCB[47], MONTEBURNS[48], MURE[49], ALEPH[51] ou VESTA. Nous discutons à présent de TRIPOLI-4®D, développé au CEA. TRIPOLI-4®-D TRIPOLI-4®-D est la fontion d’évolution (depletion en anglais) du code TRIPOLI4®, et permet de résoudre l’équation du transport à l’aide de TRIPOLI-4® et les équations d’évolution des compositions à l’aide du code MENDEL [52] qui est un solveur déterministe développé au SERMA. Le code TRIPOLI-4®-D fournit à l’utilisateur des interfaces C++ à TRIPOLI-4® et à MENDEL, ainsi que des scripts de couplage permettant de résoudre le système d’équations évoqué ci-dessus. Ces scripts peuvent être par

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

91

conséquent prédéfinis ou écrits de manière interactive à l’aide de l’interpréteur CINT [53] du logiciel ROOT [54] . Le code met en outre à profit les nombreuses fonctionnalités de ROOT pour le traitement des données et la visualisation des résultats. Si TRIPOLI-4®-D propose plusieurs schémas de résolution du système couplé d’équations. Nous nous contentons ici d’évoquer le plus simple d’entre eux : la méthode d’Euler explicite (ordre 1 en temps). Cette méthode correspond à l’algorithme 5.1 et consiste à itérer successivement sur le calcul des flux à l’aide de TRIPOLI-4® (les concentrations isotopiques sont données par MENDEL) puis sur le calcul des concentrations isotopiques (les flux sont donnés par TRIPOLI-4®). L’algorithme de couplage est initialisé avec des valeurs réalistes pour les concentrations isotopiques initiales. Un paramètre libre apparaissant dans cette résolution est donc l’intervalle de temps dt sur lequel on considère que les flux sont constants et qui est utilisé par MENDEL pour calculer le nouveau jeu de concentrations. Algorithm 5.1 Méthode d’euler explicite pour la résolution des équations couplées de Boltzmann et de Bateman  Require: N i (t=0) , T, δt . Etat initial, découpage temporel procedure Euler explicite(0, T, δt) for t = 0 → T do h i {φ, τ }(t) ← T RIP OLI N i (t) h i  i N (t+δt) ← M EN DEL N i (t) , {φ, τ }(t) , δt t ← t + δt end for

. Transport . Evolution

end procedure  Ensure: ∀t ∈ [0; T ], N i (t)

5.2.2

. Etat à chaque pas de temps

Mise en œuvre de l’échantillonnage corrélé pour accélérer les simulations

Les codes Monte-Carlo évoluants consistent donc à faire des appels successifs à un code Monte-Carlo résolvant l’équation du transport, et à un code d’évolution résolvant les équations d’irradiation sous flux. D’un point de vue technologique, l’utilisation de ces codes est entravée par leurs temps de calcul prohibitifs : en effet les solveurs Monte-Carlo de transport des particules sont eux-mêmes coûteux en temps de calcul (par comparaison aux solveurs déterministes), et par conséquent les nombreuses simulations Monte-Carlo successives utilisées par les codes Monte-Carlo évoluants mènent à des temps de simulation si longs qu’il est actuellement encore impossible de simuler de manière fine un système complexe comme un cœur de réacteur. Ce constat rend crucial le développement de méthodes à même d’accélérer les calculs Monte-Carlo évoluants. On présente ici la méthode utilisée dans le cadre de la thèse de M. Dieudonné. Echantillonnage corrélé pour le calcul des concentrations La méthode des échantillons corrélés est une méthode de type perturbative qui permet, pour un surcroît en temps de calcul a priori marginal, de calculer les réponses d’une simulation Monte-Carlo pour laquelle certains paramètres ont été perturbés, sachant que l’on a déjà réalisé la simulation Monte-Carlo “initiale” (avec le jeu de paramètres initiaux). Son principe consiste à enregistrer les traces/histoires (cf A.6) de la simulation initiale, puis à utiliser un algorithme de modification des poids statistiques des particules dans la simulation initiale qui prenne en compte la perturbation de certains paramètres, comme les densités des matériaux, les concentrations isotopiques ou les sections efficaces par exemple. Cela permet alors de calculer des scores associés à la situation perturbée, à partir d’histoires de particules générées dans la simulation initiale [55]. Dans la suite, on cherchera à évaluer l’impact des perturbations en concentrations et pour cela on considère le cas simple d’un transport monocinétique dans un milieu infini, homogène et isotrope composé d’un unique nucléide de section macroscopique totale Σt = N σt (la section microscopique totale du nucléide est notée σt et sa concentration N ). La modification des poids statistiques par échantillonnage corrélé au lieu de collision z~k d’une trace γˆm = {z~0 , z~1 , . . . , z~m } (où z~0 est la position de la source, z~m le lieu de l’absorption et les termes généraux z~k les lieux de collisions) décrivant un neutron ayant réalisé exactement m collisions avant absorption s’écrit [16]

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

k

κ =

collisions Y  06n6k

Σ∗t Σt



92



(5.15)

~ −z~n | e−(Σt −Σt )|zn+1 ,

où Σ∗t = N ∗ σt est la section macroscopique totale perturbée du nucléide (seule la valeur de sa concentration ~ −z~n | N ∗ a été modifiée). Cette expression découle naturellement de la probabilité Σt e−Σt |zn+1 d’échantillonner une collision en zn+1 ~ sachant que le neutron se trouve en z~n . Utilisation de l’échantillonnage corrélé pour accélérer les simulations Il est alors possible d’utiliser l’échantillonnage corrélé dans le cadre des simulations en évolution : lors du premier transport Monte-Carlo on enregistre les histoires des neutrons dans un fichier de traces, et l’on considère que les transports suivants ne sont que des perturbations en concentrations de ce premier transport initial (l’amplitude des perturbations est calculée par le code résolvant les équations de Bateman). Cela permet de mettre en œuvre l’échantillonnage corrélé sur des perturbations en concentrations dans le cadre des codes Monte-Carlo évoluant afin de “factoriser” les transports Monte-Carlo qui sont coûteux en temps de calcul. Cette idée avait été proposée dans une forme très contrainte par M. Moriwaki en 2002 [56, 57] et a été généralisée puis mise en œuvre dans la thèse de M. Dieudonné. L’algorithme associé à cette méthode s’insère très naturellement dans les différents schémas de résolution des équations couplées : pour le schéma de type Euler explicite par exemple, l’algorithme de principe 5.1 est remplacé par le nouvel algorithme 5.2. Algorithm 5.2 Remplacement des séquences successives de transport par des perturbations en concentration du jeu de traces initial  Require: N i (t=0) , T, δt h i  γ(t=0) ← T RIP OLI N i (t=0) . Création des traces  procedure Euler explicite perturbé(0, T, δt, γ(t=0) ) for t = 0 → T do h  i {φ, τ }(t) ← P ERT U RBAT ION γ(t=0) , N i (t) i h  i N (t+δt) ← M EN DEL N i (t) , {φ, τ }(t) , δt t ← t + δt end for

. Remplacement du transport

end procedure  Ensure: ∀t ∈ [0; T ], N i (t)

5.2.3

Calcul des perturbations acceptables par intégrales de chemins

Dans la pratique, la mise en œuvre de cette méthode s’est avérée délicate car les concentrations des différents nucléides peuvent varier fortement d’un pas de temps au suivant, et il devient difficile de parler de perturbations des concentrations... Afin d’appréhender l’impact de l’échantillonnage corrélé sur des variations en concentrations, un modèle simple a été développé, qui a permis d’établir des prédictions qualitatives solides sur l’effet de ces variations sur les scores calculés. On présente ici ce modèle dans sa formulation la plus simple (un nucléide en milieu infini, homogène et isotrope). Formalisme On considère une trace γˆm = {z~0 , z~1 , . . . , z~m } décrivant un neutron ayant parcouru, en m collisions, la distance cumulée γm =

m−1 X k=0

E. Dumonteil

|zk+1 ~ − z~k | =

m−1 X

lk .

k=0

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

93

La probabilité d’observer la trace γˆm (i.e. que le neutron ait effectué ses m collisions selon la distance cumulée γm ) est : m p (ˆ γm ) = (Σt ) e−Σt γm pm−1 (1 − p), | {z } {z } | probabilit´ e de probabilit´ e de parcourir γm f aire exactement en m chocs m chocs où p est la probabilité de faire une diffusion et 1 − p la probabilité d’être absorbé. Comme un score quelconque peut être calculé en fonction de l’estimateur du flux, la contribution au score R de cette trace s’écrit, à l’aide de l’estimateur collision (voir Eq.4.12) R (ˆ γm ) =

m X ω k (ˆ γm )

Σt

k=1

m 1 X k = ω , Σt k=1

où wk est le poids statistique du neutron au point de collision k, et où la dépendance de ce poids statistique à la trace a été rendue implicite. Cette expression permet de calculer le score moyen produit par un neutron via X R = R (ˆ γm ) p (ˆ γm ) ! ˆ ∞ m X 1 X k m = ω (Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p) , m Σ t m=1 R+ k=1

ainsi que le moment d’ordre 2 de ce score via X R2 = R2 (ˆ γm ) p (ˆ γm ) =

∞ ˆ X m=1

Rm +

m 1 X k ω Σt k=1

!2 m

(Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p) ,

et sa variance via

2

(5.16)

σ 2 = R2 − R . Pour simplifier les calculs, on décompose R et R2 en série R R2

= =

∞ X m=1 ∞ X

Rm , 2 , Rm

m=1

avec

  ´  Pm m k  Rm = Rm Σ1 ω (Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p) k=1 t +   . 2  R2 = ´ m 1 Pm ω k (Σ )m e−Σt γm dl . . . dl pm−1 (1 − p) t 1 m m k=1 Σt R +

2 avant d’analyser la convergence des séries de Cela permet de discuter de l’existence des intégrales Rm et Rm 2 terme général Rm et Rm . On remarque que les expressions pour la moyenne, le moment d’ordre 2 et la variance du score sont des intégrales sur tous les chemins possibles (on somme sur toutes les collisions, et à nombre de collisions donné, on intègre sur toutes les distances possibles entre chaque collision). Cette démarche est en fait analogue à celle proposée au chapitre 3, mais on reste dans l’espace direct (les équations ne s’appliquent pas à la transformée de Laplace).

Résolution du problème initial On considère que la simulation initiale (non-perturbée) est effectuée sans technique de réduction de variance et que par conséquent tous les poids statistiques valent 1. On a alors R (ˆ γm ) =

E. Dumonteil

m . Σt Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION Dans ce cas, on montre en annexe A.7 que R= et que σ2

1 , Σa

94

(5.17) 2

= pR .

(5.18)

Lors d’une simulation Monte-Carlo, comme on moyenne les contributions de N neutrons, la variance est donnée par σ 2 /N . Résolution du problème perturbé On considère maintenant une variation de la concentration δN = N ∗ − N de notre nucléide. Sans technique de réduction de variance, les poids statistiques prenant en compte la perturbation sont donnés, pour la kième collision, par κk ω k = κk . D’après l’équation 5.15 ces poids valent ∗

κk = αk e−(Σt −Σt )γk ,

(5.19)

avec α = = N /N . On montre alors en annexe A.7 que le score moyen et la variance associés à ce processus perturbé sont donnés par 1 cs ∗ R = ∗ =R , (5.20) Σa   2  ε(1−p ) − 1 R∗ 2 pour α ∈ D 1−pε σ cs 2 = , (5.21)  +∞ pour α ∈ /D Σ∗t /Σt



avec

ε=

α2 . 2α − 1

2 Cette expression de la variance du processus perturbé doit être associée à la condition d’existence de Rm P 2 cs de convergence de Rm suivante  α > 21 . pε < 1

cs

et

Ces deux conditions peuvent être réécrites sous une forme simple, à savoir un domaine de validité portant sur la variable α et fonction de la probabilité de diffusion p   √ √ 1− 1−p 1+ 1−p D= ; , p p et qui traduit une divergence de la variance lorsque α est en dehors de ce domaine. Analyse des résultats et domaine de validité Afin de s’assurer de l’exactitude des résultats obtenus, on réalise des simulations Monte-Carlo simplifiées mettant en œuvre l’échantillonnage corrélé, et reprenant les différentes hypothèses du modèle théorique (particules monocinétiques dans milieu infini et isotrope absorbant et diffusant). On fixe Σt = 1 et on fait varier α et p. La figure 5.5 compare ainsi les erreurs statistiques relatives ¯ cs ) théoriques (figure de gauche) et simulées (figure de droite). Le domaine de validité D conditionnant (σ cs /R l’existence de l’erreur statistique théorique est reporté sur le graphe de l’erreur statistique simulée. Les deux figures présentent un excellent accord. Bien sûr, à la différence de l’erreur théorique, l’erreur statistique estimée par les simulations ne diverge pas et est elle-même entachée d’une erreur statistique (que l’on pourrait calculer via le moment d’ordre 3 de la distribution du flux). Mais qualitativement, à l’intérieur du domaine de validité le comportement des deux fonctions est semblable, en fonction de p et de α. A l’extérieur du domaine de validité, l’erreur estimée croît rapidement mais reste finie. Cette observation doit être mise en regard de considérations sur le flux moyen : la figure 5.6 fait apparaître l’écart relatif entre le flux moyen théorique et le flux moyen estimé par les simulations en fonction de α et de p et montre qu’à l’intérieur du domaine de validité cet écart est très faible, et qu’à l’extérieur du domaine l’écart devient important. Ainsi, le domaine de validité dérivé théoriquement revêt une importance cruciale car il permet de distinguer les conditions sur p et sur α pour lesquelles les simulations donnent des résultats non-biaisés. Physiquement, le domaine D permet donc de délimiter les perturbations en concentrations (variable α), en fonction des propriétés du milieu, particulièrement de son pouvoir diffusant p et absorbant (1 − p). La forme de ce domaine montre que :

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION •







95

pour des milieux très diffusifs, l’amplitude des perturbations est très restreinte car D → Ø quand p → 1. En fait le flux et tous les moments de sa distribution divergent en milieu infini lorsqu’il n’y a pas d’absorptions... pour des milieux très  absorbants, l’amplitude des perturbations en concentration est importante puisque  l’on a D = 12 ; +∞ quand p → 0. Cette remarque est importante car l’évolution des combustibles en physique des réacteurs fait apparaître très rapidement des produits de fission et des noyaux absorbants (comme le xénon) dans tous les cas, on doit prendre garde à ce que les concentrations perturbées soient au moins égales à la moitié des concentrations non perturbées (condition α > 1/2) il faut être très prudent lorsque l’on met en œuvre l’échantillonnage corrélé pour accélérer les simulations Monte-Carlo en évolution car si l’amplitude des perturbations ne satisfait pas les conditions précédentes, l’estimation du flux par la méthode est biaisée

Figure 5.5 Erreur statistique relative théorique (à gauche) et simulée (à droite). On fait apparaître en courbe noir pleine le domaine de validité des résultats sur l’erreur simulée. On a fixé Σt = 1.

Figure 5.6 Ecart relatif entre le flux théorique moyen et le flux estimé par les simulations MonteCarlo en fonction de α et de p.

5.2.4

Perspectives technologiques, big data

Dans la pratique, la thèse de M. Dieudonné a montré que pour des applications réalistes (telles que l’étude des premiers cycles d’un assemblage REP) la méthode d’accélération se montrait performante et permettait de gagner jusqu’à un facteur 8 en temps de calcul ! En fait cette méthode peut être généralisée à d’autres types de couplage multiphysique, comme le couplage neutronique-thermohydraulique. Dans ce dernier cas, les

E. Dumonteil

Université Paris-Sud

CHAPITRE 5. APPORTS DE LA MODÉLISATION À LA SIMULATION

96

grandeurs perturbées sont les sections efficaces microscopiques qui sont modifiées par Doppler-broadening des températures. En cherchant à généraliser cette idée, on peut remarquer que l’échantillonnage corrélé s’appuie sur la persistence des données (enregistrement des traces d’une simulation) et peut à ce titre ouvrir des perspectives importantes, dans un contexte technologique où les capacités de stockage des ordinateurs ont été démultipliées. En effet, il est à présent presque possible de réaliser des expériences numériques consistant à simuler des configurations physiques avec un nombre réaliste de particules, et à enregistrer l’ensemble des données simulées sur des disques adhoc. Ces données de référence peuvent être fournies aux utilisateurs au même titre que les codes de simulation et, à l’aide de techniques dérivées de l’échantillonnage corrélé, permettre à ces utilisateurs de calculer des configurations différentes (perturbées) relativement à la configuration de référence, ou d’établir des schémas de calcul multi-physiques à partir de ces données. L’avantage est alors double : les temps de calcul deviennent très faible car la simulation à proprement parler est déjà faite (génération aléatoire des événements, tracking dans les géométries, etc.) et de plus les données fournies associées à une configuration physique sont validées par les équipes de développement à leur origine. Si les nombres de particules simulées sont réalistes, les simulations Monte-Carlo permettent de plus l’accès à tous les moments des distributions d’intérêt, et pas uniquement à la moyenne. Comme dans ce cas les volumes de données impliquées sont extrêmement importants, la mise en œuvre de ces méthodes devra être fondée sur les techniques très contemporaines dites de “big data” dont l’objet est justement de pouvoir proposer des outils de traitement et de manipulation de volumes massifs de données. Tous ces travaux doivent bien sûr être encadrés par des considérations théoriques sur les gammes de perturbations accessibles...

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

97

Références Articles signés ou cosignés [1] Dumonteil, E., Courau, T., "Dominance ratio assessment and Monte-Carlo criticality simulations : dealing with high dominance ratio systems ", Nuclear Technology 172, 120 (2010). [2] Dumonteil, E., et al., "Particle clustering in Monte-Carlo Criticality simulations", Annals of Nuclear Energy 63, 612-618 (2014). [3] Dumonteil, E., Le Peillet, A., Lee, Y.K., Petit, O., Jouanne, C., Mazzolo, A., "Source convergence diagnostics using Boltzmann entropy criterion", Proceedings of PHYSOR2006, Vancouver (2006). [4] L’Abbate, A., Courau, T., Dumonteil, E., "Monte-Carlo Criticality Calculations : Source Convergence and Dominance Ratio in an Infinite Lattice Using MCNP and TRIPOLI-4", Proceedings of PHYTRA1, Marrakech (2007). [5] Dumonteil, E., Malvagi, F., "Automatic treatment of the variance estimation bias in TRIPOLI-4 criticality calculations ", Proceedings of ICAP2012, Chicago (2012). [6] Zoia, A., Dumonteil, E., Mazzolo, A., Mohamed, S., J. Phys. A : Math. Theor. 45, 425002 (2012). [7] Zoia, A., Dumonteil, E., Mazzolo, A., Europhys. Lett. 100, 40002 (2012). [8] Brun, E., Dumonteil, E., "Etat d’avancement de TRIPOLI-4-D", Rapport DM2S (2010) [9] R. Reyes Ramirez et al., "Comparison of MCNPX-C90 and TRIPOLI-4-D for fuel depletion calculations of a Gas-cooled fast reactor", Annals of Nuclear Energy, Vol37, 8, P1101-1106 (2010) [10] E. Dumonteil and C.M.B. Diop, "Unbiased minimum variance of a matrix exponential function. Application to Boltzman/Bateman coupled equations solving", International Conference on Mathematics, Computational Methods and Reactor Physics 2009, Saratoga Springs, New York, USA (2009) [11] E. Dumonteil, C.M.B. Diop, "Biases and statistical error bars in Monte-Carlo burnup calculations : an unbiased stochastic scheme to solve boltzmann/Bateman coupled equations", Nuclear Science and Engineering, Vol 167, N2, P165-170 (2011) [12] E. Brun, E. Dumonteil, and F. Malvagi, "Uncertainties propagation in Monte-Carlo burnup codes. Implementation in TRIPOLI-4-D", Joint International Conference of the 7th Supercomputing in Nuclear Application and the 3rd Monte-Carlo, Tokyo, Japan (2010) [13] E. Brun, E. Dumonteil, and F. Malvagi, "Uncertainties propagation in Monte-Carlo burnup codes. Implementation in TRIPOLI-4-D", Progress in Nuclear Science and Technology, vol. 2, p. 879-885 (2011) [14] Dieudonné, C., et al., "Depletion calculations based on perturbations. Application to the study of a REPlike assembly at beginning of cycle with TRIPOLI-4", Proceedings of the International Conference on Supercomputing in Nuclear Applications and Monte-Carlo (2013) [15] Dieudonné, C., et al., "Monte-Carlo burnup code acceleration with the correlated sampling method. Preliminary test on an UOx cell with TRIPOLI-4", Proceedings of the International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering (2013) [16] Dieudonné, C., "Accélération de la simulation Monte-Carlo du transport des neutrons dans un milieu évoluant par la méthode des échantillons corrélés", Thèse de doctorat de l’Université Paris-Sud (2013)

Autres articles [17] https ://www.oecd-nea.org/dbdata/data/nds_eval_libs.htm [18] Lux, I., Koblinger, L., "Monte-Carlo particle transport methods : Neutron and photon calculations", CRC Press, Boca Raton (1991). [19] Brown, F., "LA-UR-05-4983 : Fundamentals of Monte-Carlo particle transport", Los Alamos National Laboratory (2005). [20] Brown, F., Proceedings of M&C2009, Saratoga Springs, N.Y. USA (2009). [21] Watson, H.W., Galton, F., "On the Probability of the Extinction of Families ", J. Anthropol. Instit. of Great Britain 4, 138 (1875).

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

98

[22] Gelbard E.M., Prael, R., "Computation of standard deviations in eigenvalue calculations ", Prog. Nucl. Energy 24, 237 (1990). [23] Jacquet, O., Chajari, R., Bay, X., Nourri, A., Carraro, L., "Eigenvalue uncertainty evaluation in MC calculations using time series methodologies ", Proceedings of ICRS2000, Lisbon (2000). [24] Osborn, R.K., Yip, S., "The foundations of neutron transport theory", Gordon and Breach, New York (1966). [25] Pázsit, I., Pál, L., "Neutron fluctuations : A treatise on the physics of branching processes", Elsevier, Oxford (2008). [26] Sjenitzer, B.L., Hoogenboom, J.E., "Variance reduction for fixed-source Monte-Carlo calculations in multiplying systems by improving chain-length statistics", Ann. Nucl. Eng. 38, 2195 (2011). [27] TRIPOLI-4 Project Team, "TRIPOLI-4 User guide", Rapport CEA-R-6169 (2008). [28] Brun, E., Dumonteil, E., Hugot, F.X., Huot, N., Jouanne, C., Lee, Y.K., Malvagi, F., Mazzolo, A., Petit, O., Trama, J.C., Zoia, A., "Overview of TRIPOLI-4 version 7 Continuous-energy Monte-Carlo Transport Code", Proceedings of ICAPP2011, Nice (2011). [29] Ueki, T., Brown, F., "Stationarity and Source Convergence Diagnostics in Monte-Carlo Criticality Calculations ", Proceedings of M&C2003, Gatlinburg, Tennessee (2003). [30] Ueki, T., Brown, F., Parsons, D.K., Kornreich, D.E., "Autocorrelation and dominance ratio in Monte-Carlo criticality calculations ", Nucl. Sci. Eng. 145, 279 (2003). [31] Ueki, T., Brown, F., Parsons, D.K., Warsa, J.S., "Time Series Analysis of Monte-Carlo fission sources : Dominance ration computation ", Nucl. Sci. Eng. 148, 374 (2004). [32] Ueki, T., J. Nucl. Sci. Tech. 49, 1134-1143 (2012). [33] Williams, M.M.R., "Random Processes in Nuclear Reactors", Pergamon Press, Oxford (1972). [34] Athreya, K.B., Ney, P., "Branching processes", Grundlehren Series, Springer-Verlag, New York (1972). [35] Cox, J.T., Griffeath, D., "Occupation times for critical branching brownian motions", Annals Prob. 13, 1108 (1985). [36] Dawson, D.A., "The critical measure diffusion process", Z. Wahrsch. Verw. Gebiete 40, 125 (1972). [37] Houchmandzadeh, B., "Clustering of diffusing organisms", Phys. Rev. E 66, 052902 (2002). [38] Houchmandzadeh, B., "Neutral clustering in a simple experimental ecological community ", Phys. Rev. Lett. 101, 078103 (2008). [39] Houchmandzadeh, B., "Theory of neutral clustering for growing populations ", Phys. Rev. E 80, 051920 (2009). [40] Young, W.R., Roberts, A.J., Stuhne, G., "Reproductive pair correlations and the clustering of organisms ", Nature, 412, 328 (2011). [41] Duderstadt, J.J., Hamilton, L.J., "Nuclear reactor analysis", Wiley, New York (1976). [42] Bell, G.I., Glasstone, S., "Nuclear reactor theory", Van Nostrand Reinhold Company, New York (1970). [43] http ://eqworld.ipmnet.ru/en/solutions/lpde/lpde103.pdf [44] RL Moore, BG Schnitzler, CA Wemple, RS Babcock, and DE Wessol. MOCUP : MCNP-ORIGEN2 coupled utility program. IdahoNational Engineering Laboratory Report INEL-95/0523.1995. [45] K. Okumura, M. Nakagawa, and K. Kaneko. Development of Burn-up Calculation Code System MVPBURN Based on Continuous Energy Monte-Carlo Method and Its Validation. In Proc. of Joint Int. Conf. on Mathematical Methods and Supercomputing for Nuclear Applications (M & C and SNA 97), volume 1, page 495 (1997). [46] EA Gomin and LV Maiorov. The MCU Monte-Carlo Code for 3-D Depletion Calculation. In Proc. of the International Conference on Mathematics and Computation, Reactor Physics and Environmental Analysis in Nuclear Applications, pages 16–2. [47] J. Cetnar, J. Wallenius, and W. Gudowski. MCB : A continuous energy Monte-Carlo Burnup simulation code. Actinide and Fission Product Partitioning and Transmutation", EUR (1998). [48] H.R. Trellue. Development of Monteburns : A Code That Links MCNP and ORIGEN2 in an Automated Fashion for Burnup Calculations. Technical report, LA-13514-T, Los Alamos National Laboratory, Los Alamos, NM, (1998).

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

99

[49] O. Meplan, A. Nuttin, O. Laulan, S. David, F. Michel-Sendis, and J. Wilson. MURE : MCNP Utility for Reactor Evolution. Description of Methods and First Applications, ENC, pages 11–14 (2005). [50] Z. Xu, P. Hejzlar, MJ Driscoll, and MS Kazimi. An Improved MCNP-ORIGEN Depletion Program (MCODE) and its Verification for High-Burnup Applications,’PHYSOR-2002, Seoul. Korea, October, pages 7–10 (2002). [51] Haeck, W., Verboomen, B., Abderrahim, H, and Wagemans, C., ALEPH : An Efficient Approach to MonteCarlo Burn Up. ALEPH : An Efficent Approach to Monte-Carlo Burn Up, Chattanooga, Tennessee, United States, pages 17–21 (2005). [52] A. Tsilanizara, S. Lahaye, E. Brun. Etat du solveur Evolution et les premières vérifications. SERMA/LLPR/RT/09-4640/A (2009). [53] Masaharu Goto. C++ interpreter - cint. CQ publishing (1997). [54] R. Brun, F. Rademakers. Root - an object oriented data analysis framework. Nucl. Inst. & Meth. in Phys. Res., A389 (1997). [55] Dejonghe, G., "Etudes d’effets différentiels par la méthode de Monte-Carlo dans le cadre de l’équation du transport. Applications aux calculs de protection et de neutronique", Thèse d’Orsay (1982) [56] Moriwaki, M., Ayoma, M., "Improvement of Monte-Carlo Lattice Burnup Calculation Performance with the Correlated Sampling Method", Journal of Nuclear Science and Technology (2002) [57] Moriwaki, M., "Multi-Assembly Analysis with an Advance Correlated Sampling Method", Journal of Nuclear Science and Technology (2003)

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

100

E. Dumonteil

Université Paris-Sud

Chapitre

6

Projet, perspectives L

a cohérence de la démarche présentée dans ce manuscrit se situe dans la modélisation et la simulation des processus de transport branchant au niveau discret. En effet, les 20 dernières années ont vu exploser la capacité de calcul et de stockage des ordinateurs, ce qui a autorisé la réalisation de véritables “expériences numériques” de type Monte-Carlo. Dans le domaine de la neutronique, ces simulations permettent alors d’avoir accès non plus uniquement aux valeurs moyennes, comme de coutume, mais également aux autres moments de la distribution de probabilité, comme la variance par exemple. Les méthodes de fouille de données rendent même possible l’analyse d’événements au cas par cas. Les méthodes numériques ainsi que la théorie du transport stochastique doivent accompagner ce changement profond offert par l’informatique. Afin de présenter mon projet de recherche à la lumière de ces considérations, je reprends dans la suite chacun des axes de travaux présentés au chapitre 2 (ingénierie, aspects numériques, aspects théoriques), en tentant d’apporter un éclairage sur les liens entre ces trois axes, ainsi que sur leur convergence. Ingénierie et développement des codes, transfert de technologie Les méthodes de simulation dites de Monte-Carlo ont été mises au point par des physiciens de Los Alamos travaillant sur le projet Manhattan pendant la seconde guerre mondiale. Issues de la neutronique, ces méthodes se sont généralisées à bien d’autres domaines et prennent une importance croissante avec l’augmentation de la puissance de calcul des ordinateurs (nombre de CPU par carte mère) et de leur capacité de stockage (mémoires RAM et ROM). Ainsi, l’enregistrement des traces des neutrons présenté ci-dessus et rendu possible par cette révolution technique entraîne, du point de vue de l’utilisateur, un changement de paradigme fort. En effet, depuis les années 1970 et jusqu’à présent, les principaux codes Monte-Carlo de la neutronique fonctionnaient sur le même principe : il s’agissait de simuler l’histoire des neutrons, une par une, pour n’enregistrer à la fin de leur vie que les grandeurs d’intérêt pour l’utilisateur (dans le langage du neutronicien, leur contribution au score). Le développement de la capacité de stockage d’informations (aujourd’hui de l’ordre du tera-octet sur des micro-ordinateurs commerciaux) permet maintenant de procéder à de véritables expériences numériques : par exemple, plutôt que de simuler un cœur de réacteur nucléaire chaque fois qu’il convient de calculer une grandeur physique d’intérêt (comme le flux ex-core vu par un instrument de dosimétrie), il est possible à présent d’imaginer de stocker l’histoire détaillée de tous les neutrons une seule fois sur un espace mémoire dédié, puis de faire des requêtes à cette maquette numérique avec des outils informatiques ad hoc chaque fois que l’on veut un résultat. Si un utilisateur souhaite faire un calcul pour des paramètres du cœur légèrement différents de ceux ayant permis de produire les histoires enregistrées, il est possible de proposer des méthodes perturbatives (comme l’échantillonnage corrélé) pour re-analyser les traces en tenant compte de la perturbation. Les axes de développement que je souhaite porter visent donc à améliorer les outils d’enregistrement des traces, de fouille de données, et d’études perturbatives pour rendre leur utilisation routinière et industrielle. Aspects numériques : simulation des processus de transport, développement de méthodes numériques L’enregistrement des traces permet d’imaginer de nouvelles méthodes numériques, dont deux sont présentées ici. D’une part, le développement de la méthode de réduction de variance s’appuyant sur un réseau de neurones s’est avéré très prometteur et doit être poursuivi. Cette méthode pourrait être aussi utilisée pour accélerer les 101

CHAPITRE 6. PROJET, PERSPECTIVES

102

simulations de criticité, en se substituant à l’itération sur la puissance, à l’origine des corrélations. Il s’agirait alors, à l’aide de réseaux de neurones ou d’algorithmes de sélection génétique, de trouver les sources MonteCarlo permettant de maximiser le kef f de la configuration physique. Aussi, la piste explorée pour réduire la variance consistait à évaluer la fonction d’importance. Des méthodes récentes utilisant la théorie des systèmes de particules en interaction [1, 2] pourraient compléter l’utilisation des méthodes d’apprentissage : ces méthodes supposent connue la fonction d’importance (dite fonction de coût ou fonction “potentiel” dans le contexte des mathématiques appliquées) et utilisent des systèmes de particules en interaction pour améliorer la variance. D’autre part, les travaux portant sur l’utilisation de l’échantillonnage corrélé pour accélérer les simulations Monte-Carlo évoluant (voir sous-section 5.2.2) peuvent eux-mêmes être prolongés en suivant deux axes différents. Le premier axe consisterait à proposer une formulation entièrement discrète des calculs d’évolution eux-mêmes : en effet, lors du calcul Monte-Carlo évoluant, la résolution des équations d’évolution de Bateman suppose une connaissance du flux dans l’ensemble de la géométrie, et impose de faire des hypothèses de modélisation physique (comme l’introduction d’un maillage spatial où le flux est supposé constant dans chacune des mailles). L’idée consisterait donc à poursuivre le principe de la formulation perturbative de ce problème de couplage non-linéaire afin de construire une formulation purement Monte-Carlo de la modélisation et de la simulation du transport évoluant. Cette formulation s’appuyerait sur l’utilisation de chacune des collisions de la simulation initiale pour évaluer le flux à un endroit donné, à l’aide de l’estimateur du flux en un point. Ce formalisme ferait l’économie des hypothèses fortes faites dans le cadre des schémas Monte-Carlo évoluant actuels, ce qui est extrêmement intéressant car les simulations Monte-Carlo sont utilisées bien souvent par l’industrie du nucléaire pour établir des calculs de référence. On notera également que l’étude qualitative de cette méthode reposant sur une vision discrète du transport (le flux est vu vraiment comme un ensemble de collisions et de parcours rectilignes, et non pas comme un champ scalaire) ne peut être faite qu’à l’aide d’outils théoriques adhoc, discutés à présent. Le deuxième axe évoqué en sous-section 5.2.4 viserait à généraliser l’utilisation de l’échantillonnage corrélé à d’autres couplages multi-physiques (couplage neutronique-thermohydraulique par exemple) et à utiliser alors les préceptes du “big data” pour élaborer des méthodes numériques à même de traiter les volumes très importants de traces de la simulation de référence. Aspects théoriques : modélisation des processus de transport à l’aide des marches aléatoires Concernant la théorie du transport de neutrons, il conviendrait d’en définir une formulation complètement stochastique et discrète, qui constituerait le pendant formel (et éventuellement analytique) aux simulations de transport Monte-Carlo (de la même manière que la théorie de champ moyen qu’est l’équation de Boltzmann représente le pendant formel des simulations déterministes capables de ne produire que des résultats moyennés). Nous avons vu que la mécanique statistique et plus particulièrement l’étude des marches aléatoires branchantes constituait un outil de choix pour parvenir à cette fin. Cependant, comme les simulations Monte-Carlo font très peu d’approximations physiques (on restitue le comportement des neutrons “comme dans la réalité” sans introduire de modélisation physique), les hypothèses utilisées dans les travaux précédents doivent être relaxées. Ainsi, on peut imaginer plusieurs axes de recherche se positionnant dans la continuité de ces développements : • étude des effets de spectre sur les moments d’ordre supérieur des distributions (beaucoup de solutions des équations de Feynman-Kac proposées dans ce manuscrit sont monocinétiques, or on sait que les effets de perte d’énergie des neutrons lors des réactions de diffusion et de ré-échantillonnage de leur énergie sur les spectres de fission lors des branchements sont potentiellement très importants) • généralisation des équations pour la prise en compte des effets cinétiques. En effet, lors des réactions de fissions, les interactions faibles mises en causes interviennent sur des échelles de temps différentes des réactions de diffusion : les neutrons issus de ces interactions faibles sont dits “retardés” et il convient de modéliser également cet effet sans lequel il serait impossible, par exemple, de piloter un cœur de réacteur nucléaire • poursuite de l’effort visant à décrire le mécanisme de clustering des neutrons. La démarche backward présentée dans ce manuscrit doit permettre d’étudier facilement, par exemple, différents types de conditions aux limites (fuite, réflexion) afin de positionner la théorie au plus près de la physique des cœurs. De même, il serait important d’évaluer l’impact des neutrons retardés sur le clustering.

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

103

Références [1] A. Guyader, N. Hengartner, E. Matzner-Lober, "Simulation and Estimation of Extreme Quantiles and Extreme Probabilities", Applied Mathematics and Optimization, Volume 64, Issue 2, p171-196 (2011) [2] P. Del Moral, J. Garnier, "Genealogical particle analysis of rare event", The Annals of Applied Probability, Vol. 15, No. 4, 2496-2534 (2005)

E. Dumonteil

Université Paris-Sud

RÉFÉRENCES

104

E. Dumonteil

Université Paris-Sud

Annexe

A

Démonstrations A.1

Rappel de la notion d’intervalle de confiance et le théorème de la limite centrale

On considère N variables aléatoires indépendantes et identiquement distribuées, notées XN , ou X par abus 2 de la variable aléatoire X finie. Le théorème de la limite centrale établit de langage. On suppose la variance σX la convergence en loi de la suite XN vers la loi normale. Ainsi, la distribution de la moyenne empirique de scores obtenus sur un grand nombre d’échantillons tend, d’après ce théorème, vers une distribution de Gauss σ2 2 de moyenne X et de variance σm = NX (avec N le nombre d’échantillons). Ce théorème permet de construire un intervalle de confiance - probabilité que la mesure soit exacte à l’erreur  près - noté C () sur m (la moyenne empirique) :  C () = P robabilit´ e X −6m6X + On montre alors que C (σm ) ' 68, 3% de confiance, C (2σm ) ' 95, 4% et C (3σm ) ' 99, 7%.

A.2

Réduction de variance dans les problèmes de transport

Nous cherchons à démontrer que :

→ − → − − ω . ∇φ∗ (P ) + Σ∗t (→ r , E) φ∗ (P ) =

¨∞ 0 4π

   → − → − − − − − Σs → r , E 0 → E, ω 0 → → ω φ∗ → r , E 0 , ω 0 d→ ω dE 0 + S ∗ (P ) ,

avec  ∗ φ (P ) = = (P ) φ (P )      −  Σ∗ (→ r , E) = Σ (P ) − t

t

→ − ∇=(P ) → − =(P )

→ −

∇=(P ) ω (P ) . → − ∇=(P )

  S ∗ (P ) = = (P ) S (P )   → − −  r ,E,→ ω)  K ∗ (P 0 → P ) = =(→ K (P 0 → P ) − =(− r ,E 0 ,→ ω 0) et où P représente un point dans l’espace des phase (~r, ω ~ , E). En effet, si on utilise le changement de variable φ∗ (P ) = = (P ) φ (P ) dans l’équation de Boltzmann → − → − − ω . ∇φ (P ) + Σt (→ r , E) φ (P ) = alors on a, pour le terme de gauche : 105

¨

− Σs φd→ ω dE 0 + S (P )

ANNEXE A. DÉMONSTRATIONS

→ − → − − ω . ∇φ (P ) + Σt (→ r , E) φ (P )

106

→ − φ∗ (P ) φ∗ (P ) − − = → ω .∇ + Σt (→ r , E) = (P ) = (P ) → − ∗ → − ∗ = (P ) ∇φ (P ) − φ (P ) ∇= (P ) φ∗ (P ) → − − = ω. + Σt (→ r , E) 2 = (P ) = (P ) →     − → − ∇= (P ) → − ∇= (P ) 1 → − → −  φ∗ (P ) ω . ∇φ∗ (P ) + Σt (P ) − ω (P ) . → = − = (P ) = (P ) ∇= (P ) i → − 1 h→ − − = ω . ∇φ∗ (P ) + Σ∗t (→ r , E) φ∗ (P ) = (P )

et pour le terme de droite : ¨

− Σs φd→ ω dE 0 + S (P )

¨∞ = 0 4π

   − → − → − → − − − Σs → r , E 0 → E, ω 0 → → ω φ → r , E 0 , ω 0 d ω 0 dE 0 + S (P )

 → −0  − ∗ → 0   φ r , E , ω → − → − − − = Σs → r , E 0 → E, ω 0 → → ω d ω 0 dE 0 + S (P ) = (P ) 0 4π  ∞  ¨     → → − → − − = (P ) 1  − − − 0 0 ∗ → 0 0  Σs → r , E 0 → E, ω 0 → → ω = → −  φ r , E , ω d ω dE  + S (P ) − = (P ) = → r , E 0 , ω0 ¨∞

0 4π



=

1  = (P )

¨∞

0 4π

     → − − → − → − − = (→ r , E, → ω) − − − Σs → r , E 0 → E, ω 0 → → ω φ∗ → r , E 0 , ω 0 d ω 0 dE 0 + S ∗ (P ) − − = (→ r , E0, → ω 0)

donc en modifiant l’opérateur K K ∗ (P 0 → P ) =

− − = (→ r , E, → ω) 0  → −0  K (P → P ) , → − 0 = r ,E ,ω

le deuxième membre de l’égalité devient  ∞  ¨ ¨     → → − → − − 1 − − − −  Σs → r , E 0 → E, ω 0 → → ω φ∗ → r , E 0 , ω 0 d ω 0 dE 0 + S ∗ (P ) , Σs φd→ ω dE 0 + S (P ) = = (P ) 0 4π

et on trouve alors → − → − − ω . ∇φ∗ (P ) + Σ∗t (→ r , E) φ∗ (P ) =

¨∞ 0 4π

   → − → − → − − − − Σs → r , E 0 → E, ω 0 → → ω φ∗ → r , E 0 , ω 0 d ω 0 dE 0 + S ∗ (P ) ,

d’où le résultat.

A.3

Système de poids pour l’estimation d’une grandeur I

De manière générique, le codes Monte-Carlo de transport de particules cherchent à mesurer une grandeur physique I liée au flux de particules et s’exprimant ˆ I= f (x)φ(x)dx, (A.1) P

où :

E. Dumonteil

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

107

x désigne un point de l’espace des phases (P) à 6 dimensions (3 coordonnées de position, deux d’angle, et l’énergie) • φ désigne le flux de particules • f est une fonction réponse qui, selon le contexte, peut prendre diverses formes : section efficace pour le calcul d’un taux de réaction, ou fonction indicatrice d’un domaine spatial pour celui d’un flux intégré dans ce domaine En prenant pour point de départ la décomposition en série d’opérateurs de transport et de collision donnée par 4.9, on calcule la grandeur d’intérêt via X   I= In ,   n≥0 ˆ    In = g(x)χn (x)dx, •

où g(x) est la fonction réponse associée à χ ˆ g(~r, ~v ) =

T (~r → r~0 ; ~v ) . dr~0 f (r~0 , ~v ) Σt (r~0 , ~v )

(A.2)

Or χ se décompose de la manière suivante : ˆ χn (xn ) = χn−1 (xn−1 )L(xn−1 → xn )dxn−1 , ce qui permet d’écrire In ˆ In = g(xn )χn (xn )dxn ˆ ˆ ˆ ˆ = dxn dxn−1 · · · dx1 dxQ(x0 )L(x0 → x1 ) · · · L(xn−1 → xn )g(xn ) ˆ e (n) )g(xn ), = dα(n) Q(x0 )L(α

(A.3)

avec α(n) = (x0 , x1 , . . . , xn ) la trajectoire d’une particule émise au point x0 dans l’espace des phases, et e (n) ) = Qn L(xi−1 → xi ) le noyau de l’opérateur intégral pour cette particule dont la trajectoire se L(α i=1 réalise à travers n étapes successives de transports et de collisions. Il est alors possible d’introduire un système de poids normalisateurs ˆ    w = Q(x0 )dx0 , 0  ˆ (A.4)    w(xi−1 ) = L(xi−1 → xi )dxi , ∀i ∈ {1, . . . , n}, et également les distributions permettant de normaliser les opérateurs de transition L  Q(x0 )   , p(x0 ) =  w0  L(xi−1 → xi )   p(xi−1 → xi ) = , ∀i ∈ {1, . . . , n}. w(xi−1 )

(A.5)

Donc le poids et la distribution associés à la trajectoire α(n) sont respectivement f (α(n) ) = w0 W

n Y

w(xi−1 ),

i=1

Pe(α(n) ) = p(x0 )

n Y i=1

(A.6) p(xi−1 → xi ).

Une réécriture d’un élément de la série de la quantité d’intérêt (A.3) est alors ˆ h i f (α(n) ) Pe(α(n) )dα(n) . In = g(xn )W

E. Dumonteil

(A.7)

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

108

Cette dernière formulation (A.7) explicite la mesure I associée à la fonction réponse g lors d’une simulation Monte-Carlo dans la mesure où nous utilisons les densités de probabilité définies par (A.5) pour échantillonner les trajectoires des particules. On voit ici l’avantage d’utiliser la densité de collision sortante : pour le premier point dans une trajectoire, on échantillonne la source directement. La simulation Monte-Carlo consiste à échantillonner N trajectoires suivant la loi Pe et à évaluer In avec l’estimateur : ξ In = Ainsi l’estimateur final I vaut : ξI =

A.4 A.4.1

N 1 X f (α(n) ). g(xn )(i) W (i) N i=1

∞ N 1 XX f (α(n) ). g(xn )(i) W (i) N n=0 i=1

(A.8)

Jeu à variance nulle Problème adjoint

Le calcul de I peut se mettre sous la forme d’un problème initial après choc ˆ    I = χ(x)g(x)dx,  ˆ    χ(x) = χ(x0 )L(x0 → x)dx0 + Q(x),

(A.9)

où l’on a utilisé la densité de collision sortante précédemment définie. Ce problème est équivalent au problème adjoint après choc ˆ    I = Q(x)χ† (x)dx,  ˆ (A.10)    χ† (x) = L(x → x0 )χ† (x0 )dx0 + g(x). En effet, similairement à l’équation 4.8, cette dernière équation est équivalente à : (A.11)

χ† = χ† L + g. qui se vérifie puisque :

ˆ I=

ˆ gχ =

ˆ †



(χ − χ L)χ =

ˆ †

χ (I − L)χ =

χ† Q = I.

La fonction χ† (x) est appelée l’importance après choc. Elle s’interprète comme l’espérance de la contribution au résultat d’une trajectoire démarrant en x. De la même manière, on peut considérer les problèmes initial et adjoint associés à la densité de collision entrante χe . Le problème initial avant choc est : ˆ    I = χe (x)h(x)dx,     ˆ  χe (x) = χe (x0 )K(x0 → x)dx0 + Q1 (x), (A.12)    ˆ     Q1 (~r, ~v ) = Q(~r0 , ~v )T (~r0 → ~r; ~v )d3~r0 ; problème équivalent au problème adjoint avant choc : ˆ    I = Q1 (x)χ†e (x)dx,  ˆ    χ†e (x) = K(x → x0 )χ†e (x0 )dx0 + h(x). E. Dumonteil

(A.13)

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

109

La fonction χ†e est appelée l’importance avant choc. En comparant les expressions de I dans (A.10) et (A.13) nous obtenons les relations réciproques entre les deux fonctions importances avant et après choc : ˆ   †  χ (~ r , ~ v ) = T (~r → ~r0 ; ~v )χ†e (~r0 , ~v )d3~r0 ,  ˆ (A.14)  †   χe (~r, ~v ) = C(~r; ~v → ~v 0 )χ† (~r, ~v 0 )d3~v 0 + h(~r, ~v ).

A.4.2

Jeu à variance nulle

Comme dans la méthode de l’échantillonnage suivant l’importance, on cherche ici la forme théorique de la fonction importance qui permet de minimiser la variance. Ainsi, si nous multiplions les deux côtés de l’équation 4.7 par χ† on obtient   ˆ L(x0 → x)χ† (x) dx0 χ(x)χ† (x) = Q(x)χ† (x) + χ(x0 )χ† (x0 ) χ† (x0 ) (A.15) ˆ † 0 † 0 ˆ 0 0 = Q(x)χ (x) + χ(x )χ (x )L(x → x)dx , où Lˆ est le noyau de transition modifié. On peut alors réécrire la quantité d’intérêt I comme  ˆ  g(x) χ(x)χ† (x)dx, I= χ† (x)

(A.16)

où χg(x) † (x) est la nouvelle fonction réponse à considérer. En reprenant les Eq.A.4 et Eq.A.5, le système de poids prend la forme  ˆ   w ˆ0 = Q(x0 )χ† (x0 )dx0        = I,   ˆ (A.17) L(xi−1 → xi )χ† (xi )  w(x ˆ i−1 ) = dxi , ∀i ∈ {1, . . . , n}  † (x  χ )  i−1    †  χ (x ) − g(x  i−1 i−1 )  , =  † χ (xi−1 ) et le système des distributions :     

Q(x0 )χ† (x0 ) , w ˆ0 ˆ i−1 → xi ) L(x → xi ) = , w(x ˆ i−1 ) pˆ(x0 ) =

    pˆ(xi−1

(A.18) ∀i ∈ {1, . . . , n}.

L’Eq.A.17 permet de calculer le poids d’une trajectoire α(n) n

Y χ† (xi−1 ) − g(xi−1 ) f ˆ (α(n) ) = I . W χ† (xi−1 ) i=1 L’estimateur ξI pour une trajectoire (N = 1) vaut alors ξI =

∞ n X g(xn ) Y χ† (xi−1 ) − g(xi−1 ) I = I. χ† (xn ) i=1 χ† (xi−1 ) n=0

(A.19)

Nous venons de mettre en évidence un jeu théorique permettant d’obtenir une variance nulle sur le résultat. Evidemment, pour connaître I il faut connaître χ† , ce qui suppose de connaître... I ! Si cela est a priori impossible, il reste toutefois possible d’intuiter la forme de la carte d’importance et de réduire par conséquent la variance du calcul de I.

E. Dumonteil

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

A.4.3

110

Equation de Placzek

En faisant les hypothèse d’un problème sphérique (ω~0 = −~r/r) et d’une diffusion isotrope, il est possible de trouver une relation pour K, en séparant les contributions spatiales =esp et directionnelles =dir de l’importance. En effet ~ r, ω ∇=(~ ~) .ω~0 , v(~r, ω ~) ~ esp (~r) ∇= = .ω~0 , =esp (~r) 1 d=esp (r) = − . =esp (r) dr

K

=

Après intégration, en notant r la distance au détecteur, = s’écrit =(~r, ω ~ ) = =dir (~ ω )e−Kr et donc χ†e (~r, ω ~) = −Kr =dir (~ ω )e † † † . En utilisant de plus χ (~r, ω ~ ) = χesp (~r)χdir (~ ω ), il vient compte tenu de l’hypothèse de diffusion Σt isotrope ˆ χ†e (~r, ω ~)

= = ∝

1 Σs † ~ 0 ) d2 ω ~ 0, χ (~r, ω 4π 4π Σt   ˆ 1 Σs † † 0 2 0 χesp (~r) χ (~ ω )d ω ~ , 4π Σt 4π dir χ†esp (~r).

Ainsi l’importance avant choc, sous nos hypothèses, ne dépend pas de la direction de la particule et on a une forme explicite pour l’importance avant choc : χ†e (~r) ∝ e−Kr /Σt . On peut alors calculer χ† ˆ †

χ (~r, ω ~)

= (~ r 0 =~ r +%~ ω)

=

=

ˆ

0

T (~r → ~r0 ; ω ~ )e−Kr d3~r0 , ∞

Σt e−Σt % eK(~r+%~ω).ω~0 d%, 0 ˆ ∞ −Kr e Σt e−(Σt −K~ω.ω~0 )% d%, 0

Σt e−Kr . Σt − K~ ω .ω~0

=

Or il faut également que χ†e vérifie l’équation d’importance suivante ˆ χ†e (~r)

= ˆ



= =

C(~ ω→ω ~ 0 ; ~r)χ† (~r, ω ~ 0 ) d2 ω ~ 0,

1 Σs Σt e−Kr d2 ω ~ 0, 0 .ω 4π Σ Σ − K~ ω ~ t t 0 4π   Σs Σt + K e−Kr . ln 2K Σt − K

Les deux équations portant sur χ†e imposent la relation suivante, appellée équation de Placzek Σs ln 2K

E. Dumonteil



Σt + K Σt − K

 = 1.

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

A.5

111

Ordres de grandeur pour la population neutronique dans un cœur de réacteur nucléaire

Un cœur de puissance fournit environ une puissance Pe = 900 MW électriques, c’est-à-dire Pth = 2700 MW thermiques. Comme une réaction de fission libère approximativement Ef = 200 MeV, soit Ef = 200.106 .1.6.10−19 J/s, le nombre de fissions par seconde dans le cœur est donné par Pth /Ef ' 1020 fissions/s (avec 1W= 1J/s). Le taux λ = Σt v auquel se produit un événement, pour un neutron, est le produit de la section macroscopique totale (environ 1/cm) et de la vitesse des neutrons (2200 m/s pour des neutrons thermiques). On a donc λ ' 2 ∗ 105 /s/neutrons. Le nombre de neutrons thermiques présents dans un réacteur de puissance est donc de 10 /2.105 ' 1015 neutrons (où on a fait l’approximation Σf = Σt ). Ce nombre de neutrons total dans le cœur est associé à une densité de 108 neutrons/cm3 (ce qui est réaliste), si l’on considère un cœur cubique de 4 mètres de côté. 20

Le flux thermique associé à ces densités est calculé via τf = Σf φ, et comme Σf ≈ 0.5 cm−1 (car σf ≈ 500.10−24 cm2 et NU 5 ≈ 1021 noyaux/cm3 ), il vient φ ≈ 1020 /4003 /0.5 ≈ 1013 n/cm2 /s. Si l’on souhaite calculer les différents moments des scores d’intérêt, et non pas uniquement des quantités moyennes, la simulation Monte-Carlo de ce type de configuration nécessite donc 105 milliards de neutrons par cycle, ce qui se situe à quelques ordres de grandeur des calculs les plus intensifs déjà réalisés. La simulation réaliste d’un crayon nécessiste quant à elle environ 1 milliard de neutrons par cycle (car un cœur de type REP est fait d’environ 200 assemblages de 17 × 17 crayons chacun).

A.6

Les histoires des neutrons : concept de traces

A l’instar des codes de simulation et de reconstruction en physique des hautes énergies, certains codes Monte-Carlo comme TRIPOLI-4® permettent d’enregistrer l’ensemble des données d’une simulation MonteCarlo. Cet ensemble de données est en fait constitué par toutes les informations associées aux histoires des particules simulées (l’ensemble des événements relatifs à l’histoire d’une particule) c’est-à-dire • • •

• • • • •



l’identifiant de la particule (neutron, photon etc) le type de l’événement (naissance, collision, franchissement d’interface, ...) le lieu de l’événement dans l’espace des phases (c’est-à-dire les coordonnées spatio-temporelles et l’énergie de la particule) les poids avant et après collision (si l’événement est une collision) le noyau choqué (si l’événement est une collision) les noms et indices des volumes et compositions où se produit l’événement un identifiant unique associé à l’histoire de la particule et au cycle/batch auquel elle appartient un identifiant unique associé à la particule parente, si la particule considérée est née d’une fission ou d’un événement de splitting les informations générales sur le cycle auquel appartient la particule (données de renormalisation du cycle, nombre de cycles, de particules par cycle, etc.) permettant de reconstruire des scores

Ces histoires de particules (qui portent aussi le nom de “traces”, par référence aux traces que laissaient les particules dans les premiers dispositifs de détection en physique des particules) sont alors enregistrées dans des fichiers de traces au format binaire, comme les fichiers ROOT. Les fichiers en question permettent de reconstruire a posteriori n’importe quel type de score pour une simulation donnée. Comme ordre de grandeur on peut retenir 1ko par histoire de neutron, ce qui signifie que l’enregistrement intégral d’une petite simulation d’un million de neutrons nécessite 1Go d’espace sur le disque dur. Dans TRIPOLI-4®, l’outil permettant de gérer ces fichiers de traces s’appelle t4ROOTtools. Il implémente des fonctionnalités de visualisation des histoires, de reconstruction des scores et d’analyse “microscopique”.

E. Dumonteil

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

A.7 A.7.1

112

Calcul par intégrales de chemins des scores Monte-Carlo de la simulation initiale et perturbée Simulation initiale

Moment d’ordre 1

Comme Σt > 0, l’intégrale Rm existe et vaut ! ˆ m 1 X i m Rm = ω (Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p) Σ t Rm + i=1 ˆ m m (Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p) = Σ t Rm + 1 mpm−1 (1 − p) Σt

= car ∀λ > 0 on a :

´ R+

  +∞ λe−λl dl = −e−λl 0 = 1. Or la série de terme général mq m−1 m∈N∗ (avec q ∈ [0; 1[)

est convergente, de somme

R=

1 , (1−q)2

∞ X

donc la série de terme général Rm l’est aussi, de somme ∞ 1−p X 1 1 1 1 1−p = = mpm−1 = Σt m=1 Σt (1 − p)2 Σt 1 − p Σa

Rm =

m=1

Moment d’ordre 2

2 existe et vaut De même, l’intégrale Rm

ˆ 2 Rm

m 1 X i ω Σt i=1

= Rm +

= =

m2 Σ2t

ˆ

!2 m

(Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p)

m

Rm +

(Σt ) e−Σt γm dl1 . . . dlm pm−1 (1 − p)

1 2 m−1 m p (1 − p) Σ2t

 Comme la série de terme général m2 q m−1 m∈N∗ (avec q ∈ [0; 1[) est convergente, de somme série de terme général

2 Rm1

R2

1+q , (1−q)3

donc la

l’est aussi, de somme

=

∞ X m=1

2 Rm

∞ 1 − p X 2 m−1 1 1+p 2 = m p = 2 = (1 + p) R Σ2t m=1 Σt (1 − p)2

Variance Cette dernière expression permet d’écrire la variance 2

σ 2 = R2 − R = pR

A.7.2

2

Simulation perturbée

Moment d’ordre 1

Pour une trace γˆm quelconque, on peut écrire ! m 1 X i,cs m cs R (ˆ γm ) p (ˆ γm ) = ω (Σt ) e−Σt γm pm−1 (1 − p) Σ∗t i=1 ! m 1 X i −(Σ∗t −Σt )γi m = αe (Σt ) e−Σt γm pm−1 (1 − p) Σ∗t i=1 =

E. Dumonteil

m ∗ 1 X i m α (Σt ) e−Σt γi e−Σt (γm −γi ) pm−1 (1 − p) ∗ Σt i=1

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

113 cs

Et comme Σt > 0 et Σ∗t > 0, l’intégrale Rm existe et vaut ˆ X m ∗ 1 cs m Rm = αi (Σt ) e−Σt γi e−Σt (γm −γi ) dl1 . . . dlm pm−1 (1 − p) Σ∗t Rm + i=1 ˆ m X ∗ 1 m i = α (Σt ) e−Σt γi e−Σt (γm −γi ) dl1 . . . dlm pm−1 (1 − p) m Σ∗t i=1 R+ =

m 1 X i m −i −m+i m−1 α (Σt ) (Σ∗t ) (Σt ) p (1 − p) Σ∗t i=1

=

m 1 X i −i m−1 αα p (1 − p) Σ∗t i=1

1 mpm−1 (1 − p) Σ∗t  Or la série de terme générale mq m−1 m∈N∗ (avec q ∈ [0; 1[) est convergente, de somme =

terme général Rm

cs

1 , (1−q)2

donc la série de

l’est aussi, de somme R

cs

=

∞ X

Rm

cs

=

m=1

∞ 1−p X 1 1−p 1 1 mpm−1 = ∗ = ∗ Σ∗t m=1 Σt (1 − p)2 Σt 1 − p

Comme il n’y a qu’un seul nucléide dans le milieu, la probabilité de survie à une interaction dans le milieu perturbé est la même que celle du milieu non perturbé p∗ =

Σ∗s N ∗ σs = ∗ =p ∗ Σt N σt

Le score perturbé converge donc effectivement cs

R1 =

1 1 1 ∗ = ∗ =R ∗ ∗ Σt 1 − p Σa

Moment d’ordre 2

Soit une trace γˆm quelconque, on a !2 m 1 X i,cs m cs 2 ω R (ˆ γm ) p (ˆ γm ) = (Σt ) e−Σt γm pm−1 (1 − p) Σ∗t i=1 1 !2 m X 1 m i −(Σ∗ −Σ )γ t i t = αe (Σt ) e−Σt γm pm−1 (1 − p) . 2 (Σ∗t ) i=1

La décomposition

!2

m X

ai

=

i=1

m X

a2i + 2

m−1 X

m X

ai aj

i=1 j=i+1

i=1

permet alors d’écrire 2 Rcs 2 (ˆ γm ) p (ˆ γm ) (Σ∗t )

=

m X

! α2i e

−2(Σ∗ t −Σt )γi

i=1



m−1 X

... + 2

m X

m

(Σt ) e−Σt γm pm−1 (1 − p) + . . .  αi+j e

−(Σ∗ t −Σt )(γi +γj )

i=1 j=i+1

=

m X i=1



m

α2i (Σt ) e−(2Σt −Σt )γi e−Σt (γm −γi ) pm−1 (1 − p) + . . .

... + 2

m−1 X

m X

i=1 j=i+1

E. Dumonteil

 (Σt )m e−Σt γm pm−1 (1 − p)

m





αi+j (Σt ) e−(2Σt −Σt )γi e−Σt (γj −γi ) e−Σt (γm −γj ) pm−1 (1 − p)

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

114 cs

2 Comme Σt > 0 et Σ∗t > 0, l’intégrale Rm existe si et seulement si 2Σ∗t − Σt > 0, c’est-à-dire si et seulement si α > 21 . Et dans ce cas, cette intégrale vaut :

2 Rm

cs

2 (Σ∗t )

ˆ

m X

= Rm +

i=1

ˆ

m−1 X

... + 2 Rm +

=

m X



m

α2i (Σt ) e−(2Σt −Σt )γi e−Σt (γm −γi ) dl1 . . . dlm pm−1 (1 − p) + . . .

α2i (Σt )

m

m−1 X

=

i=1



m X

α

i+j

(Σt )

m

α2i (Σt ) (2Σ∗t − Σt )

... + 2

m−1 X

m X

m

ˆ

=

i=1

=

m X i=1

=

−i



Rm +

−m+i

(Σt )

pm−1 (1 − p) + . . .

αi+j (Σt ) (2Σ∗t − Σt )

εi pm−1 (1 − p) + 2 εi pm−1 (1 − p) + 2

m−1 X

m X

i=1 j=i+1 m−1 X i=1



e−(2Σt −Σt )γi e−Σt (γj −γi ) e−Σt (γm −γj ) dl1 . . . dlm pm−1 (1 − p)

m

i=1 j=i+1 m X



e−(2Σt −Σt )γi e−Σt (γm −γi ) dl1 . . . dlm pm−1 (1 − p) + . . .

i=1 j=i+1 m X



m

αi+j (Σt ) e−(2Σt −Σt )γi e−Σt (γj −γi ) e−Σt (γm −γj ) dl1 . . . dlm pm−1 (1 − p)

i=1 j=i+1

ˆ

Rm +

i=1

... + 2

m X

−i

−j+i

(Σ∗t )

(Σt )

−m+j

pm−1 (1 − p)

εi pm−1 (1 − p)

(m − i) εi pm−1 (1 − p)

2ε ε (1 − εm ) pm−1 (1 − p) + 1−ε 1−ε



1 − εm m− 1−ε



pm−1 (1 − p)

soit   ε 1+ε m = (1 − ε ) pm−1 (1 − p) 2m − 2 1−ε (Σ∗t ) 1 − ε  Puisque la série de terme général mq m−1 m∈N∗ (avec q ∈ [0; 1[) est convergente, de somme 2 Rm

série de terme général

2 Rm

cs

cs

1

1 , (1−q)2

donc la

l’est aussi si et seulement si

pε < 1, 2 et dans ce cas, la série de terme général Rm

R2

cs

=

∞ X

2 Rm

cs

a pour somme

cs

m=1

= = = = =

E. Dumonteil

 ∞  1−p X 1+ε m 2m − (1 − ε ) pm−1 2 1−ε (Σ∗t ) 1 − ε m=1 "  # 2 1+ε 1 ε ε 1−p − 2 2 − 1−ε 1 − p 1 − pε (Σ∗t ) 1 − ε (1 − p)   1 ε 2 1+ε − 2 (Σ∗t ) 1 − ε 1 − p 1 − pε  ε 1 − p2 1 2 2 (Σ∗t ) (1 − p) 1 − pε  ε 1 − p2 ∗2 R 1 − pε ε

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

115

Variance Finalement, la variance s’écrit σ

E. Dumonteil

cs 2

=

R2

cs

−R

cs 2

"

#  ε 1 − p2 ∗2 = −1 R 1 − pε

Université Paris-Sud

ANNEXE A. DÉMONSTRATIONS

E. Dumonteil

116

Université Paris-Sud

Annexe

B

Publications principales B.1

liées au Chapitre 3

117

PRL 106, 220602 (2011)

week ending 3 JUNE 2011

PHYSICAL REVIEW LETTERS

Collision-Number Statistics for Transport Processes A. Zoia,* E. Dumonteil, and A. Mazzolo CEA/Saclay; DEN/DANS/DM2S/SERMA/LTSD; 91191 Gif-sur-Yvette, France (Received 4 March 2011; published 2 June 2011) Physical observables are often represented as walkers performing random displacements. When the number of collisions before leaving the explored domain is small, the diffusion approximation leads to incongruous results. In this Letter, we explicitly derive an explicit formula for the moments of the number of particle collisions in an arbitrary volume, for a broad class of transport processes. This approach is shown to generalize the celebrated Kac formula for the moments of residence times. Some applications are illustrated for bounded, unbounded and absorbing domains. DOI: 10.1103/PhysRevLett.106.220602

PACS numbers: 05.40.Fb, 02.50.r

Many practical problems, encompassing areas as diverse as research strategies, market evolution, percolation through porous media, and DNA translocation through nanopores, to name only a few [1–4], demand assessing the statistics of the random residence time tV spent by a walker inside a given domain V . Indeed, complex physical systems are often described in terms of ‘‘particles’’ undergoing random displacements, resulting either from the intrinsic stochastic nature of the underlying process, or from uncertainty [5,6]. As a particular case, when the particle is lost upon touching the boundary @V of V , the residence time is usually called first-passage time [7]. Fully characterizing tV is an awkward task, since its distribution generally depends on walker dynamics, geometry, boundaries and initial conditions, so that one has often to be content with the mean residence time [8]. This has motivated a large number of theoretical investigations over the last decade, covering both homogeneous and heterogeneous, scale-invariant media [7–13]. In the former case, the dynamics of the walker is usually modeled by regular Brownian motion, whereas in the latter one resorts to anomalous diffusion. A seminal work developed by Kac [14], based on a path integral approach, allows all the moments of the residence times of Brownian particles to be evaluated by resorting to convolutions over the ensemble equilibrium distribution of the walkers, for arbitrary boundary conditions on V [15,16]. However, in many realistic situations, the walker typically undergoes a limited number of collisions before leaving the explored domain, so that the diffusion limit is possibly not attained. Examples are widespread, and arise in, e.g., gas dynamics, neutronics and radiative transfer, electronics, and biology [17–20]. In all such systems, the stochastic path can be thought of as a series of straight-line flights, separated by random collisions, as in Fig. 1, and the dynamics is better described in terms of the Boltzmann equation, rather than the (anomalous) Fokker-Planck equation [17]. A natural variable for describing the walker evolution is therefore the number of collisions nV within the observed volume. Application of the diffusion 0031-9007=11=106(22)=220602(4)

approximation to the characterization of the counting statistics, which amounts to assuming a large number of collisions in V , might lead to inaccurate results [21]. In the present Letter, we address the issue of generalizing Kac approach to random walkers obeying the Boltzmann equation, i.e., not satisfying the diffusion regime, for arbitrary geometries and boundary conditions. We derive an explicit formula for the moments of nV , and illustrate its relation to the equilibrium distribution of the walkers. Knowledge of higher order moments allows estimating the uncertainty on the average, as well as reconstructing the full distribution of the collision-number. We show that when V is large as compared to the typical size of a flight, so that the diffusion limit is reached, Kac formula is recovered. Counting statistics.—Consider the random walk of a particle starting from a point-source located in r0 . At each collision, the particle can be either scattered, with probability p, or absorbed (in which case the trajectory terminates). For the sake of simplicity, we assume that scattering is isotropic. We denote by r the position of the walker entering a collision, as customary. The dynamics is characterized by ðr; r0 Þ, namely, the probability density of performing a displacement from r0 to r, between any two collisions. Define the transport operator ½fðrÞ Z ½fðrÞ ¼ ðr; r0 Þfðr0 Þdr0 ; (1) V

FIG. 1 (color online). A random walk starting from r0 and performing a limited number nV of collisions in a region V (with transparent boundaries), before being absorbed at rn .

220602-1

Ó 2011 American Physical Society

PRL 106, 220602 (2011)

over a d-dimensional volume V . We can then express the propagator ðr; njr0 Þ, i.e., the probability density of finding a particle in r at the nth collision (starting from r0 ), as ðr; njr0 Þ ¼ pn1 n ½ðr; r0 Þ, where n ½fðrÞ is the nth iterated operator Z Z ðr; rn Þ    ðr2 ; r1 Þfðr1 Þdr1    drn ; n ½fðrÞ ¼ V

V

(2) and  ¼ ðr  r0 Þ is a shorthand notation for the initial point-source condition. The probability of performing nV collisions in the volume V is related to the propagator by Z Z drðr;nV jr0 Þ drðr;nV þ1jr0 Þ: P ðnV jr0 Þ ¼ V

(3) The moments

nV ¼1

ðrjr0 Þ ¼ lim

N!1

nm P ðnV jr0 Þ V

(4)

N X

ðr; njr0 Þ;

(5)

n¼1

which intuitively represents the equilibrium (stationary) particle distribution [17]. It follows immediately that R hn1V iðr0 Þ ¼ V drðrjr0 Þ, i.e., the integral of the collision density over a volume V gives the mean number of collisions within that domain, hence the name given to ðrjr0 Þ. Higher order moments of nV can be obtained as follows. Define the operator Z ðrjr0 Þfðr0 Þdr0 : (6) ½fðrÞ ¼ V

By making use of the Neumann series

n¼1

hnm iðr0 Þ ¼ V

pn1 n ½fðrÞ ¼

 ½fðrÞ; 1  p

(7)

 ½fðrÞ; ½fðrÞ ¼ 1  p

V

(8)

(11)

(12)

which are defined as k-fold convolutions of the collision density ðrjr0 Þ with itself [14,15]. It follows the equivalence Z drk ½ðr; r0 Þ; (13) C k ðr0 Þ ¼ k! V

with Ckþ1 ðr0 Þ ¼ ðk þ 1Þ½Ck ðr0 Þ and C0 ðr0 Þ ¼ 1r0 2V , 1 being the characteristic function. The convergence of the integrals Ck ðr0 Þ depends on the features of the underlying stochastic process as well as on boundary conditions. For instance, the persistence property of walks in d  2 implies diverging Ck ðr0 Þ for transparent V in absence of absorption [15]. We finally obtain the central result of this Letter, i.e., the explicit formula for the moments of the collision number in V hnm iðr0 Þ ¼ V

m 1 X s pk Ck ðr0 Þ: p k¼1 m;k

(14)

When the underlying dynamics ðr; r0 Þ is known, Eq. (14) provides exact estimates of the collision statistics for transport-dominated processes. Thanks to linearity, Eq. (14) allows expressing hnm iðr0 Þ V as a combination of m Kac integrals, k ¼ 1; . . . ; m, each given from Eq. R (12). In particular, for m ¼ 1 we recover hn1V iðr0 Þ ¼ V drðrjr0 Þ, since s1;1 ¼ 1. Furthermore, for m ¼ 2, s2;1 ¼ s2;2 ¼ 1, so that Z Z dr2 dr1 ðr2 jr1 Þðr1 jr0 Þ þ hn1V i: (15) hn2V i ¼ 2p V

we have then

m X 1Z dr k!s pk k ½ðr; r0 Þ: p V k¼1 m;k

We introduce then the repeated Kac integrals Z Z Ck ðr0 Þ ¼ k! drk ... dr1 ðrk jrk1 Þ...ðr1 jr0 Þ; V

þ1 X

depend on the boundary conditions on @V , which affect the functional form of the propagator. The absence of boundary conditions corresponds to defining a fictitious (‘‘transparent’’) volume V , where particles can indefinitely cross @V back and forth. On the contrary, the use of leakage boundary conditions leads to the formulation of first-passage problems [7]. We introduce now the collision density

1 X

1Z drLim ðpÞð1  pÞ½ðr; r0 Þ; (9) p V P k s where Lis ðxÞ ¼ 1 k¼1 x =k is the polylogarithm function [22]. When m is a non-negative integer, the polylogarithm is a rational function, namely,   m X x kþ1 Li m ðxÞ ¼ k!smþ1;kþ1 ; (10) 1x k¼0 P where the coefficients sm;k ¼ k!1 ki¼0 ð1Þi ðkiÞðk  iÞm are the Stirling numbers of second kind [22]. Thanks to the recurrence properties of the Stirling numbers, Eq. (9) gives hnm iðr0 Þ ¼ V

V

hnm iðr0 Þ ¼ V

week ending 3 JUNE 2011

PHYSICAL REVIEW LETTERS

V

Let Gðzjr0 Þ be the moment generating function of P ðnV jr0 Þ. By definition we have the moment expansion

and in particular ðrjr0 Þ ¼ ½ðr; r0 Þ. Now, combining Eqs. (3) and (4), we get 220602-2

Gðzjr0 Þ ¼

1 X m¼0

hnm iðr0 Þ V

zm : m!

(16)

PRL 106, 220602 (2011)

10

The small-z expansion of Gðzjr0 Þ reads then Gðzjr0 Þ ’ 1 þ hn1V iðr0 Þz, which for the Tauberian theorems corresponds to the large-nV behavior. It follows the exponential tail 1

P ðnV jr0 Þ ’ enV =hnV iðr0 Þ ;

Cm ðr0 Þ ; Dm

5

0

(17)

provided that C1 ðr0 Þ is finite. Diffusion limit and Kac formula.—Suppose for the sake of simplicity that the walker moves at constant speed v, and p ¼ 1. Then, the time spent between any two collisions is ti ¼ jri  ri1 j=v. For isotropic walks, ðr; r0 Þ ¼ ð‘ ¼ jr  r0 jÞ thanks to the spherical symmetry. Then, flight times are identically distributed, and obey ti  wðti Þ, R where wðti Þ ¼ d ‘d1 ð‘Þðti  ‘=vÞd‘, d being the surface of the unit sphere. The diffusion limit of the transport process described above is obtained by letting the typical flight length  (the standard deviation of jump sizes, as customary) and the average intercollision time  ¼ hti i shrink to zero in such a way that the ratio D ¼ 2 = converges to a constant, namely, the diffusion coefficient. When  and  vanish, the collision PnV number in V diverges, whereas the quantity tV ¼ i¼1 ti converges to the residence time in the volume. Actually, tV should take into account also additional terms due to boundary conditions. However, as  ! 0 and nV ! 1, the trajectory will almost surely have a turning point touching the boundary, so that corrections can be safely neglected. Let now QðtV jr0 Þ be the distribution of the residence times. Under the previousR assumptions, in the Laplace space we have Qðsjr0 Þ ¼ expðstV ÞQðtV jr0 ÞdtV ¼ wðsÞnV . Any arbitrary wðsÞ with finite  has an expansion wðsÞ ’ 1  s when  ! 0. Then we have Qðsjr0 Þ ’ enV s , which implies QðtV jr0 Þ ’ ðtV  nV Þ for small . It follows that iðr0 Þ ’ m hnm iðr0 Þ, with hnm iðr0 Þ given by Eq. (14). If htm V V V we rescale the space variable r by , each of the terms of the sum in Eq. (14) carries a contribution 2k . In the diffusion limit, we therefore recover the celebrated Kac formula [14,15] for the moments of the residence times of Brownian motion htm iðr0 Þ ’ V

week ending 3 JUNE 2011

PHYSICAL REVIEW LETTERS

(18)

because all other terms in the sum vanish when  ! 0 and  ! 0, and sm;m ¼ 1. Moreover, we have the recursion mþ1 iðr0 Þ ¼ ðm þ 1Þ½htm iðr0 Þ, starting property DhtV V 0 from htV iðr0 Þ ¼ 1r0 2V . As for the residence time distribution, from Eq. (17) at long times we have QðtV jr0 Þ ’ expð  DtV =C1 ðr0 ÞÞ. Discussion and perspectives.—For simple geometries and ðrjr0 Þ, formula (14) is amenable to analytical solutions. Here we illustrate some significant cases, V being a d sphere of radius R centered in 0. Isotropic gamma flights with kernel ð‘Þ ¼ ‘d expð‘Þ=d ðÞ,  > 0, are a

0

0.5

1

1.5

2

0

1

2

3

4

1

0.5

0

FIG. 2 (color online). Moments hn1V iðr0 Þ (stars) and hn2V iðr0 Þ (circles) for 3d  ¼ 2 gamma flights with scattering probability p ¼ 1 in a volume V with transparent boundaries. Theoretical predictions from Eq. (14): solid lines; Monte Carlo simulations: symbols.

widespread transport process and describe, among others, search strategies in biology [23]. When  ¼ 2, the 3d scattering collision density assumes a simple form, ðrjr0 Þ ¼

1 : 4jr  r0 j

(19)

It follows 8 2 2 < 3R r0 ; r < R 0 hn1V iðr0 Þ ¼ R3 6 : ; r 0 R 3r0

(20)

and 8 4 < 25R 10R2 r20 þr40 þ hn1 iðr Þ; r < R 0 0 2 V : (21) hnV iðr0 Þ ¼ 4 R5 60 1 : þ hn iðr Þ; r 0 0 R 15 r0 V Comparisons with Monte Carlo simulations with 105 particles are shown in Fig. 2. Exponential flights ( ¼ 1) arise when the scattering centers are uniform, so that intercollision distances obey a Poisson distribution. Such a process is crucial for understanding, e.g., radiation propagation [17,18]. The transport kernel reads ð‘Þ ¼ ‘1d expð‘Þ=d . In 1d systems, the collision density for absorbing V and transparent boundaries reads pffiffiffiffiffiffiffi e 1pjrr0 j pffiffiffiffiffiffiffiffiffiffiffiffiffi : ðrjr0 Þ ¼ 2 1p

(22)

The moments hnm iðr0 Þ are again easily obtained from V Eq. (14). The resulting formulas are rather cumbersome and will not be presented here. Comparisons of exact

220602-3

PRL 106, 220602 (2011)

PHYSICAL REVIEW LETTERS

6

week ending 3 JUNE 2011

available, as an estimator to infer the equilibrium distribution of the underlying stochastic path, which is often not directly accessible.

4 2 0

0

0.5

1

1.5

2

2.5

3

3.5

4

2 1.5 1 0.5 0 0

0.5

1

1.5

2

2.5

3

3.5

4

FIG. 3 (color online). Moments hn1V iðr0 Þ (stars) and hn2V iðr0 Þ (circles) for 1d exponential flights with scattering probability p ¼ 0:5 in a volume V with transparent boundaries. Theoretical predictions Eq. (14): solid lines; Monte Carlo simulations: symbols.

formulas and Monte Carlo simulations with 105 particles are shown in Fig. 3. Analogous findings for exponential flights have been obtained also for leakage boundary conditions, where the collision density reads ðrjr0 Þ ¼

jr0 r=re  re r0 =r0 j  jr  r0 j ; 2

(23)

with re ¼ R þ 1 [24]. In principle, the integrals appearing in Eq. (14) can be carried out numerically for arbitrary complex volumes and ðrjr0 Þ. Moreover, the isotropy hypothesis can be possibly relaxed by replacing r with a state variable y ¼ fr; g accounting for the scattering angle. Actually, the approach presented in this Letter is fairly broad, in that it relies on a minimal number of hypotheses on the underlying displacement kernel ðr; r0 Þ. Though we have used here the language specific to transport phenomena, the same formalism could be rephrased in terms of a semi-Markov renewal process for an arbitrary state variable q evolving in the phase space according to ðq; q0 Þ. In this respect, Eq. (14) would provide the counting statistics for the events falling in V , i.e., satisfying a given ‘‘condition’’, when the diffusion regime is not attained, and the number of such events is small. Finally, observe that a nontrivial application of Eq. (14) would be to make use of the knowledge on the number of collisions in a given domain, when

*[email protected] [1] M. F. Shlesinger, Nature (London) 443, 281 (2006). [2] J. Ph. Bouchaud and A. Georges, Phys. Rep. 195, 127 (1990). [3] J. Ph. Bouchaud and M. Potters, The Theory of Financial Risk and Derivative Pricing (Cambridge University Press, Cambridge, England, 2003). [4] A. Zoia, A. Rosso, and S. N. Majumdar, Phys. Rev. Lett. 102, 120602 (2009). [5] B. D. Hughes, Random Walks and Random Environments Vol. I (Clarendon Press, Oxford, 1995). [6] G. H. Weiss, Aspects and Applications of the Random Walk (North Holland Press, Amsterdam, 1994). [7] S. Redner, A Guide to First-Passage Processes (Cambridge University Press, Cambridge, England, 2001). [8] S. Condamin et al., Nature (London) 450, 77 (2007). [9] D. ben Avraham and S. Havlin, Diffusion and Reactions in Fractals and Disordered Systems (CUP, UK, 2000). [10] R. Metzler and J. Klafter, Phys. Rep. 339, 1 (2000). [11] S. Condamin, O. Be´nichou, and M. Moreau, Phys. Rev. Lett. 95, 260601 (2005). [12] O. Be´nichou et al., J. Stat. Phys. 142, 657 (2011). [13] E. Barkai, J. Stat. Phys. 123, 883 (2006). [14] M. Kac, Probability and Related Topics in Physical Sciences, Lectures in Applied Mathematics (Wiley, New York, 1957). [15] A. M. Berezhkovskii, V. Zaloj, and N. Agmon, Phys. Rev. E 57, 3937 (1998). [16] D. S. Grebenkov, Phys. Rev. E 76, 041139 (2007). [17] C. Cercignani, The Boltzmann Equation and its Applications (Springer, New York, 1988). [18] M. Weinberg and E. P. Wigner, The Physical Theory of Neutron Chain Reactors (UCP, Chicago, 1958). [19] C. Jacoboni and P. Lugli, The Monte Carlo Method for Semiconductor Device Simulation (Springer, New York, 1989). [20] G. Le Cae¨r, J. Stat. Phys. 140, 728 (2010). [21] S. Blanco and R. Fournier, Phys. Rev. Lett. 97, 230604 (2006). [22] A. Erde´lyi et al., Higher Transcendental Functions (Krieger, NY, 1981). [23] F. Bartumeus et al., J. Theor. Biol. 252, 43 (2008). [24] A. Zoia, E. Dumonteil, and A. Mazzolo, Phys. Rev. E 83, 041137 (2011).

220602-4

PHYSICAL REVIEW E 85, 011132 (2012)

Counting statistics: A Feynman-Kac perspective A. Zoia,* E. Dumonteil, and A. Mazzolo CEA/Saclay, DEN/DANS/DM2S/SERMA/LTSD, F-91191 Gif-sur-Yvette, France (Received 14 December 2011; published 19 January 2012) By building upon a Feynman-Kac formalism, we assess the distribution of the number of collisions in a given region for a broad class of discrete-time random walks in absorbing and nonabsorbing media. We derive the evolution equation for the generating function of the number of collisions, and we complete our analysis by examining the moments of the distribution and their relation to the walker equilibrium density. Some significant applications are discussed in detail: in particular, we revisit the gambler’s ruin problem and generalize to random walks with absorption the arcsine law for the number of collisions on the half-line. DOI: 10.1103/PhysRevE.85.011132

PACS number(s): 05.40.Fb, 02.50.−r

I. INTRODUCTION

Many physical processes can be described in terms of a random walker evolving in the phase space [1–4], and one is often interested in assessing the portion of time tV that the system spends in a given region V of the explored space when observed up to time t [5–13]. This is key to understanding the dynamics of radiation transport, gas flows, research strategies, or chemical andbiological species migration in living bodies, just to name a few, the time spent in V being proportional to the interaction of the particle with the target medium [14–20]. For Brownian motion, a celebrated approach to characterizing the probability density of the residence time tV has been provided by Kac (based on Feynman path integrals) in a series of seminal papers, and later extended to Markov continuoustime processes [21–24]. For a review, see, for instance, [25]. The Feynman-Kac formalism basically allows us to write down the evolution equation for the moment generating function of tV for arbitrary domains, initial conditions, and displacement kernels. This approach has recently attracted a renewed interest [26–33], and it has also been extended to non-Markovian processes [10,34–36]. As a particular case, imposing leakage boundary conditions leads to the formulation of first-passage problems [5,6,37]. However, for those physical systems that are intrinsically discrete, the natural variable is the number of collisions nV in V when the process is observed up to the nth step, rather than time tV [33,38–41]. When nV is large, we can approximate the number of collisions in V by nV ∝ tV (the so-called diffusion limit), but this simple proportionality breaks down when V is small with respect to the typical step size of the walker and/or the effects of absorption are not negligible, so the diffusion limit is not attained [32,42]. In this paper, we derive a discrete Feynman-Kac equation for the evolution of the probability generating function of nV for a broad class of stochastic processes in absorbing and nonabsorbing media, and we illustrate this approach by explicitly working out calculations for some significant examples, such as the gambler’s ruin problem or the arcsine law. For the arcsine law, in particular, the Feynman-Kac formulas allow us to generalize the well-known Sparre Andersen results to random walks with absorption. Our analysis of the counting statistics is then completed by examining the moments of nV , which can also be obtained by building upon the Feynman-Kac

*

formalism, and their asymptotic behavior when n is large. In particular, we show that the asymptotic moments can be expressed as a function of the particle equilibrium distribution, which generalizes analogous results previously derived in terms of survival probabilities [32]. This paper is organized as follows. In Sec. II, we introduce a discrete Feynman-Kac formula for a class of random walkers in absorbing and nonabsorbing media. Then, in Sec. III we discuss some applications where the generating function can be explicitly inverted to give the probability of the number of collisions. In Sec. IV, we extend our analysis to the moments of nV , and in Sec. V we examine some examples of moment formulas. A short digression on the diffusion limit is given in Sec. VI. Perspectives are finally discussed in Sec. VII. II. FEYNMAN-KAC EQUATIONS

Consider the random walk of a particle starting from an isotropic point source S(r|r0 ) = δ(r − r0 ) located at r0 . At each collision, the particle can be either scattered (i.e., change direction) with probability ps or absorbed with probability pa = 1 − ps (in which case the trajectory terminates). We introduce the quantity T (r → r), namely the probability density of performing a displacement from r to r, between any two collisions [43,44]. For the sake of simplicity, we assume that scattering is isotropic and that displacements are equally distributed. Suppose that a particle emitted from r0 is observed up to entering the nth collision. Our aim is to characterize the distribution Pn (nV |r0 ), where nV is the number of collisions in a domain V . We can formally define nV (n) =

n 

(1)

where V (rk ) is the marker function of the region V , which takes the value 1 when the point rk ∈ V , and vanishes elsewhere. We adopt here the convention that the source is not counted, i.e., the sum begins at k = 1. Clearly, nV is a stochastic variable depending on the realizations of the underlying process and on the initial condition r0 . The behavior of its distribution, Pn (nV |r0 ), is most easily described in terms of the associated probability generating function, Fn (u|r0 ) = unV n (r0 ) =

[email protected]

1539-3755/2012/85(1)/011132(9)

V (rk ),

k=1

+∞ 

Pn (nV |r0 )unV ,

(2)

nV =0

011132-1

©2012 American Physical Society

A. ZOIA, E. DUMONTEIL, AND A. MAZZOLO

PHYSICAL REVIEW E 85, 011132 (2012)

which can be interpreted as the discrete Laplace transform (the transformed variable being u) of the collision number distribution. The derivation of an evolution equation for Fn (u|r0 ) is made simpler if we initially consider trajectories starting with a particle entering its first collision at r1 . Random flights are semi-Markovian (i.e., Markovian at collision points), which allows splitting the trajectory into a first jump, from r1 to r1 +  (the displacement  obeying the jump length density T ), and then a path from r1 +  to rn , conditioned to the fact that the particle is not absorbed at r1 . If the collision is an absorption, the trajectory ends and there will be no further events contributing to nV . Hence, F˜n+1 (u|r1 ) = ps uV (r1 )+V (r1 +)+···+V (rn+1 ) + pa uV (r1 ) ,

(3)

where the term uV (r1 ) is not stochastic and can be singled out. The tilde is used to recall that we are considering trajectories starting with a single particle entering the first collision at r1 . By observing that uV (r1 +)+···+V (rn+1 )  = F˜n (u|r1 + ), we have then the following equation for the generating function: F˜n+1 (u|r1 ) = uV (r1 ) [ps F˜n (u|r1 + ) + pa ],

(4)

where expectation is taken with respect to the random displacement . We make use then of the discrete Dynkin’s formula, which relates any sufficiently well behaved function f of a stochastic process with the adjoint kernel T ∗ (r → r) associated to T (r → r) [45], namely  f (r1 + ) = T ∗ (r → r1 )f (r )dr . (5) Intuitively, the adjoint kernel T ∗ displaces the walker backward in time. We therefore obtain the discrete Feynman-Kac equation    F˜n+1 (u|r1 ) = uV (r1 ) ps T ∗ (r → r1 )F˜n (u|r )dr + pa , (6) with the initial condition F˜1 (u|r1 ) = uV (r1 ) . Finally, by observing that the first collision coordinates r1 obey the probability density T (r0 → r1 ), it follows that  Fn (u|r0 ) = F˜n (u|r1 )T (r0 → r1 )dr1 . (7) Knowledge of Fn (u|r0 ) allows us to explicitly determine Pn (nV |r0 ). Indeed, by construction the probability generating function Fn (u|r0 ) is a polynomial in the variable u, the coefficient of each power uk being Pn (nV = k|r0 ). In particular, the probability that the particles never touch (or come back to, if the source r0 ∈ V ) the domain V is obtained by evaluating Fn (u|r0 ) at u = 0, i.e., Pn (0|r0 ) = Fn (0|r0 ). III. COLLISION NUMBER DISTRIBUTION: EXAMPLES OF CALCULATIONS

Direct calculations based on the discrete Feynman-Kac formulas, Eqs. (6) and (7), are in some cases amenable to exact results concerning Pn (nV |r0 ), at least for simple geometries

and displacement kernels. In this section, we shall discuss some relevant examples. A. The gambler’s ruin

Consider a gambler whose initial amount of money is x0  0. At each (discrete) time step, the gambler either wins or loses a fair bet, and his capital increases or decreases, respectively, by some fixed quantity s with equal probability. One might be interested to know the probability that the gambler is not ruined (i.e., that his capital has not reached zero, yet) at the nth bet, starting from the initial capital x0 . This well-known problem [46] can be easily recast in the Feynman-Kac formalism by setting a particle in motion on a straight line, starting from x0 , with scattering probability ps = 1 and a discrete displacement kernel T (x  → x) = δ(x − x  − s)/2 + δ(x − x  + s)/2. Setting s = 1 amounts to expressing the capital x0 in multiple units of the bet, and entails no loss of generality. The counting condition is imposed by assuming a Kronecker delta V (x) = δx,0 in Eq. (6): since the walker cannot cross x = 0 without touching it, solving the resulting equation for the quantity Fn (0|x0 ) gives, therefore, the required probability that the gambler is not ruined at the nth bet. We integrate now Eq. (6) and use Eq. (7): we start from the initial condition F˜1 (u|x1 ) = uV (x1 ) = uδx1 ,0 . Then, by observing that by symmetry T and T ∗ have the same functional form, and performing the integrals Eq. (7) over the δ functions, we obtain F1 (u|x0 ) = (uδx0 +1,0 + uδx0 −1,0 )/2. By injecting thus F˜1 (u|x1 ) in Eq. (6) and integrating again over the δ functions, we get F˜2 (u|x1 ) = uδx1 ,0 (uδx1 −1,0 + uδx1 +1,0 )/2. Finally, by integrating (7) we get F2 (u|x0 ) = (uδx0 −1,0 uδx0 −2,0 + uδx0 +1,0 uδx0 ,0 + uδx0 −1,0 uδx0 ,0 + uδx0 +1,0 uδx0 +2,0 )/4. Proceeding by recursion and identifying the coefficient of the zeroth-order term in the polynomial yields then the first terms in the series,   1 2 3 6 10 20 35 70 , , , , , , , ,... , Pn (0|1) = 2 4 8 16 32 64 128 256   2 3 6 10 20 35 70 126 , , , , , , , ,... , Pn (0|2) = 2 4 8 16 32 64 128 256 (8)   2 4 7 14 25 50 91 182 Pn (0|3) = , , , , , , , ,... , 2 4 8 16 32 64 128 256   2 4 8 15 30 56 112 210 , , , , , , , ,... Pn (0|4) = 2 4 8 16 32 64 128 256 for x0 = 1,2,3, . . ., respectively.1 After some rather lengthy algebra, by induction one can finally recognize the formula  

(n+x0 −1)/2   n n Pn (0|x0 ) = − 2−n , (9) k k − x0 k=0

· denoting the integer part. The quantity Pn (0|x0 ) is displayed in Fig. 1 as a function of n for a few values of x0 . The larger the initial capital x0 , the longer Pn (0|x0 ) 1 before decreasing. At large n, √ taking the limit of Eq. (9) leads to the scaling Pn (0|x0 ) 2/π x0 n−1/2 , in agreement with the findings in

The quantity uδx,0 evaluated at u = 0 is equal to 1 when x = 0, and vanishes otherwise. 1

011132-2

COUNTING STATISTICS: A FEYNMAN-KAC PERSPECTIVE

PHYSICAL REVIEW E 85, 011132 (2012)

to obey the so-called L´evy’s arcsine law tV itself is known √ Pt (tV ) = 1/ tV (t − tV )π , whose U shape basically implies that the particle will most often spend its time being always either on the positive or negative side of the axis [46,47,52]. This counterintuitive result has been shown to asymptotically hold also for discrete-time random √walks without absorption, for which one has Pn (nV |0) 1/ nV (n − nV )π when n and nV are large (see, for instance, [47]). The Feynman-Kac approach allows us to explicitly derive Pn (nV |0). Again, assume a displacement kernel with discrete jumps T (x  → x) = δ(x − x  − s)/2 + δ(x − x  + s)/2, with s = 1. Then, by integrating Eq. (6) and subsequently using Eq. (7), we compute the coefficients of the polynomial, which can be organized in an infinite triangle, whose first terms read

0

10

Pn (0|x0 )

2x20 /π −0.3

10

2x20 /π

−0.6

2x20 /π

10

−0.9

10

0

10

1

2

10

n

10

3

10

FIG. 1. (Color online) The probability Pn (0|x0 ) that the gambler is not ruined at the nth bet, given an initial capital x0 . Bets are modeled by discrete random increments of fixed size s = ±1. Blue circles: x0 = 5; red triangles: x0 = 10; green dots: x0 = 15. Lines have been added to guide the√eye. Dashed lines correspond to the asymptotic result Pn (0|x0 ) 2/π x0 n−1/2 . The interval 2x02 /π is also shown for each x0 .

Ref. [47]. This means that asymptotically the gambler is almost sure not to be ruined, yet, up to n 2x02 /π bets. Note that Eq. (9) is the survival probability of the gambler: the first-passage probability Wn (0|x0 ), i.e., the probability that the gambler is ruined exactly at the nth bet, can be obtained from Wn (x0 ) = Pn−1 (0|x0 ) − Pn (0|x0 ). As a particular case, for 0 < x0  n and n + x0 even, we recover the result in Ref. [46], namely  n x0 . (10) Wn (x0 ) = n 0 2 n n+x 2 Finally, observe that when x0 = 0,   1 2 3 6 10 20 35 ,... Pn (0|0) = 1, , , , , , , 2 4 8 16 32 64 128 for n  1. This is easily recognized as being the series

n−1 Pn (0|0) = n−1 21−n ,

(11)

n

nV = 0

0 1 2 3 4 5 6 7

1 1 2 1 4 2 8 3 16 6 32 10 64 20 128

1

2

3

4

5

6

7

1 2 1 4 1 8 2 16 3 32 6 64 10 128

2 4 2 8 2 16 4 32 6 64 12 128

3 8 3 16 3 32 6 64 9 128

6 16 6 32 6 64 12 128

10 32 10 64 10 128

20 64 20 128

35 128

Observe that this result is independent of s. To identify the elements Pn (nV |0), we initially inspect the column nV = 1 of the triangle, and we recognize the underlying series as being 2 )2−n . Then we realize given by terms of the kind ( (nn−−2)/2 that columns with nV  2 are related to the first column by a shift in the index n. The column nV = 0 can be obtained from normalization. Proceeding by induction, the elements in the triangle can be finally recast in the compact formula



n − nV − 1 nV nV 2−n . Pn (nV |0) = n−nV −1 (13) 2

2

(12)

2

which is, however, unphysical, since the gambler should not be allowed to bet when lacking an initial amount of money. B. The arcsine law with discrete jumps

Consider a walker on a straight line, starting from x0 . We are interested in assessing the distribution Pn (nV |x0 ) of the number of collisions nV that the walker performs at the right of the starting point when observed up to the nth collision. This amounts to choosing V (x) = H (x − x0 ), H being the Heaviside step function, in Eq. (6). To fix the ideas, without loss of generality we set x0 = 0, and we initially assume that ps = 1, i.e., the walker cannot be absorbed along the trajectory. This is a well known and long-studied problem for both Markovian and nonMarkovian processes [10,13,46,48–51]: for Brownian motion, the average residence time in V is simply tV t = t/2, whereas

Note that our result is slightly different from Ref. [46], where collisions are counted in pairs. When both √n and nV are large, we obtain the limit curve Pn (nV |0) 1/ nV (n − nV )π . When the scattering probability can vary in 0  ps  1, the triangle can be generated as above, and the first few terms read n

nV = 0

0 1 2 3 4

1 1 2 2−ps 4 4−2ps 8 8−4ps −ps3 16

1

2

1 2 2−ps 4 4−2ps −ps2 8 8−4ps −2ps2 16

2ps 4 2ps (2−ps ) 8 2ps (4−2ps −ps2 ) 16

3

3ps2 8

3ps2 (2−ps ) 16

4

6ps3 16

Now, the identification of the polynomial coefficients Pn (nV |0) becomes more involved, because each coefficient is itself a polynomial with respect to ps . The strategy in the

011132-3

A. ZOIA, E. DUMONTEIL, AND A. MAZZOLO

PHYSICAL REVIEW E 85, 011132 (2012)

identification is the same as above. By induction, the column for nV = 1 can be identified as being   4+2y  2 + 2y 1 − ps + 1 − ps2 ps + 1+y 4 2  3  2 2 F1 1, 2 + y,3 + y,ps (14) × 2(2 + y)

for n  2, with y = (n − 3)/2 , and P1 (1|0) = 1/2. Once Pn (1|0) is known, by inspection one realizes that the other columns Pn (nV |0) are related by  Pn (nV |0) =

ps 2

nV −1

nV nV Pn−nV +1 (1|0)

−2

10

(15)

−3

10

2

for nV  2. The probability Pn (0|0) is finally obtained from normalization, and reads   2+2z  2z ps − 1 + 1 − ps2 ps Pn (0|0) = + z 2ps 2 1  2 2 F1 2 + z,1,2 + z,ps , (16) × ps (1 + z) with z = n/2 . These results generalize Eq. (13), and are illustrated in Fig. 2, where we compare Pn (nV |0) as a function of nV for n = 10 and different values of ps . When ps = 1, the distribution approaches a U shape, as expected. As soon as ps < 1, the shape changes considerably, and in particular Pn (nV |0) becomes strongly peaked at nV = 0 as the effects of absorption overcome scattering. The presence of a second peak at nV = n is visible when ps 1 and progressively disappears as ps decreases: when ps is small, Pn (nV |0) has an exponential 0

10

−1

10

Pn (nV |x0 )

−1

10

Pn (nV |x0 )

Pn (1|0) =

0

10

−2

10

−3

10

0

10

20

nV

30

40

50

FIG. 3. (Color online) The arcsine law Pn (nV |x0 ) with discrete jump lengths, as a function of nV . The starting point is x0 = 0, and n= √ 50. Blue dots: ps = 1; dashed line: the asymptotic distribution 1/ nV (n − nV )π . Red triangles: ps = 0.95; dashed asymptotic Eq. (17).

tail. When n is large, Pn (nV |0) approaches the asymptotic curve

  nV −1

nV 1 − ps + 1 − ps2 ps (17) P∞ (nV |0) = nV 2 4 2  for nV  1, and P∞ (0|0) = (ps − 1 + 1 − ps2 )/2ps . Remark that when ps = 1, this means that the U shape for large n collapses on the two extremes at nV = 0 and nV = n. Equation (17) is an excellent approximation of Pn (nV |0) when the scattering probability is not too close to ps 1: as expected, the discrepancy between the exact and asymptotic probability is most evident when nV n, as shown in Fig. 2. Figure 3 displays Pn (nV |0) as a function of nV for n = 50 in order to emphasize the effects of ps : when ps = 1, the probability Pn (nV |0) √ is almost superposed to the asymptotic curve Pn (nV |0) 1/ nV (n − nV )π , whereas a deviation in the scattering probability as small as ps = 0.95 is sufficient to radically change the shape of the collision number distribution. Finally, observe that when nV is also large, which implies pa  1, Eq. (17) behaves as  1 − ps −(1−ps )nV P∞ (nV |0) e . (18) π nV

−4

10

C. The arcsine law with continuous jumps −5

10

0

2

4

nV

6

8

10

FIG. 2. (Color online) The arcsine law Pn (nV |x0 ) with discrete jump lengths, as a function of nV . The starting point is x0 = 0, and n = 10. Blue dots: ps = 1; red stars: ps = 3/4; green circles: ps = 1/2; black triangles: ps = 1/4. Lines have been added to guide the eye. Dashed curves are the asymptotic Eq. (17) for the corresponding value of ps .

When the displacement kernel T (x  → x) is continuous and symmetric, ps = 1, and x0 = 0, the distribution of the number of collisions nV falling in x  x0 is universal, in that it does not depend on the specific functional form of T (x  → x) (see [47] and references therein). This strong and surprising result stems from a Sparre Andersen theorem [53], whose proof is highly nontrivial (and does not apply to discrete jumps) [47]. This leaves the choice on the form of kernel T (x  → x), as far as it satisfies the hypotheses of the theorem. For the sake

011132-4

COUNTING STATISTICS: A FEYNMAN-KAC PERSPECTIVE

of simplicity, we have assumed an exponential distribution of jump lengths, i.e., T (x  → x) = s exp(−s|x − x  |)/2, with s = 1. Starting from Eqs. (6) and (7), we can generate the infinite triangle. n

nV = 0

0 1 2 3 4 5 6 7

1 1 2 3 8 3 16 35 128 63 256 231 1024 429 2048

1

2

3

4

5

6

7

1 2 2 8 5 16 20 128 35 256 126 1024 231 2048

3 8 5 16 18 128 30 256 105 1024 189 2048

3 16 20 128 30 256 100 1024 175 2048

35 128 35 256 105 1024 175 2048

63 256 126 1024 189 2048

231 1024 231 2048

429 2048

It is easy to verify that the triangle indeed does not depend on s, and that other functional forms of T (x  → x) would lead to the same coefficients for the polynomials Pn (nV |0). This holds true also for L´evy flights, where T (x  → x) is a L´evy stable law and jump lengths are unbounded [47]. We start from the column nV = 1, observe the relation with the subsequent columns nV  2, and finally derive the case nV = 0 from normalization. Proceeding therefore by induction, we recognize that the elements of the triangle obey   2nV 2n − 2nV 2−2n . (19) Pn (nV |0) = n − nV nV We recover here the celebrated results of the collision number distribution for discrete-time walks with symmetric continuous jumps in the absence of absorption [46,47]. When both n and nV are large, it√ is possible to show that Eq. (19) converges to the U shape 1/ nV (n − nV )π . When the scattering probability is allowed to vary in 0  ps  1, it turns out that the polynomial coefficients Pn (nV |0) are the same for several different continuous symmetric kernels T (x  → x) (L´evy flights included), and we are therefore led to conjecture that the universality result for the case ps = 1 carries over to random walks with absorption. This allows us to generalize the Sparre Andersen theorem for the collision number distribution on the half-line to a broader class of Markovian discrete-time processes. The first few terms in the triangle [which, for practical purposes, we have generated by resorting to T (x  → x) = s exp(−s|x − x  |)/2, with s = 1] read

n

nV = 0

0 1 2

1

3 4 5

1 2 4−ps 8 8−2ps −ps2 16 64−16ps −8pS2 −5ps3 128 128−32ps −16ps2 −10ps3 −7ps4 256

PHYSICAL REVIEW E 85, 011132 (2012)

As above, identification of the terms Pn (nV |0) becomes more involved, because each coefficient is itself a polynomial with respect to ps . By induction, the column for nV = 1 can be identified as being √  n  2n − 2 ps 1 − ps + Pn (1|0) = n−1 2 4  1  2 F1 − 2 + n,1,1 + n,ps (20) × n for n  1. Once Pn (1|0) is known, the subsequent columns Pn (nV |0) are observed to obey  nV −1  2nV − 1 ps Pn (nV |0) = Pn−nV +1 (1|0) (21) nV 4 for nV  2. The probability Pn (0|0) is finally obtained from normalization, and reads √  n  2n p s − 1 + 1 − ps ps + Pn (0|0) = n ps 4 1  2 F1 2 + n,1,2 + n,ps . (22) × 2(1 + n) These results are illustrated in Fig. 4, where we compare Pn (nV |0) as a function of nV for n = 10 and different values of ps . The findings for continuous jumps closely resemble those for discrete displacements. When ps = 1, the distribution approaches a U shape, as expected. As soon as ps < 1, the shape again changes abruptly, and in particular Pn (nV |0) becomes strongly peaked at nV = 0 when absorption dominates scattering. When ps is small, Pn (nV |0) decreases exponentially at large nV . When n is large, Pn (nV |0) approaches the asymptotic curve √  nV −1  2nV − 1 1 − ps ps P∞ (nV |0) = (23) nV 4 2 √ for nV  1, and P∞ (0|0) = (ps − 1 + 1 − ps )/ps . Again, Eq. (23) is an excellent approximation of Pn (nV |0) when the scattering probability is not too close to ps 1, as shown in Fig. 4. Figure 5 displays Pn (nV |0) as a function of nV for n = 50 in order to emphasize the effects of ps : when ps = 1, the probability Pn (nV |0) √ is almost superposed to the asymptotic curve Pn (nV |0) 1/ nV (n − nV )π , whereas a deviation in the scattering probability as small as ps = 0.95 is sufficient to radically change the shape of the collision number distribution. Finally, observe that when nV is also large, which implies pa  1, Eq. (23) yields the same scaling as Eq. (18). All analytical calculations discussed here have been

1

2

3

4

5

1 2 4−2ps 8 8−4ps −ps2 16 64−32ps −8ps2 −4ps3 128 128−64ps −16ps2 −8ps3 −5ps4 256

3ps 8 3ps (2−ps ) 16 6ps (8−4ps −ps2 ) 128 6ps (16−8ps −2ps2 −ps3 ) 256

5ps2 16 20ps2 (2−ps ) 128 10ps2 (8−4ps −ps2 ) 256

35ps3 128 35ps3 (2−ps ) 256

63ps4 256

011132-5

A. ZOIA, E. DUMONTEIL, AND A. MAZZOLO

PHYSICAL REVIEW E 85, 011132 (2012)

0

moment generating function for trajectories entering their first collision at r1 , which implies  (m)  ∂m ˜ (24) n˜ V n (r1 ) = (−1)m m G n (u|r1 )|u=1 , ∂u

−2

x (k) = x(x + 1) · · · (x + k − 1) being the rising factorial [54]. The tilde is used to recall that the moments refer to trajectories starting from the first collision at r1 . Combining Eqs. (6) and (24) yields the recursion property       (m)  n˜ V n+1 (r1 ) − ps T ∗ (r → r1 ) n˜ (m) V n (r )dr   = mV (r1 ) n˜ (m−1) (r ) (25) V n+1 1

10

Pn (nV |x0 )

10

−4

10

−6

10

0

2

4

nV

6

8

10

FIG. 4. (Color online) The arcsine law Pn (nV |x0 ) with exponential jump lengths, as a function of nV . The starting point is x0 = 0, and n = 10. Blue dots: ps = 1; red stars: ps = 3/4; green circles: ps = 1/2; black triangles: ps = 1/4. Lines have been added to guide the eye. Dashed curves are the asymptotic Eq. (23) for the corresponding value of ps .

verified by comparison with Monte Carlo simulations with 106 particles.

IV. MOMENTS FORMULAS

A complementary tool for characterizing the distribution Pn (nV |r0 ) is provided by the analysis of its moments. Toward ˜ n (u|r1 ) = this end, it is convenient to introduce the function G ˜ n (u|r1 ) is the (rising) factorial F˜n (1/u|r1 ). By construction, G

for m  1, with the conditions n˜ (0) V n (r1 ) = 1 and (m) n˜ V 1 (r1 ) = m!V (r1 ). Finally, the factorial moments n(m) V n (r0 ) for particles emitted at r0 are obtained from   (m)   (m)  n˜ V n (r1 )T (r0 → r1 )dr1 . (26) nV n (r0 ) = When trajectories are observed up to n → +∞, we can set (m) n(m) V  = limn→+∞ nV n , and from Eq. (25) we find       (m)  n˜ V (r1 ) − ps T ∗ (r → r1 ) n˜ (m) V (r )dr   = mV (r1 ) n˜ (m−1) (27) (r1 ) V for m  1, provided that n˜ (m) V (r1 ) is finite. It turns out that the asymptotic moments n˜ (m) V (r1 ) are related to the equilibrium distribution of the particles [32,39,40]. To see this, we introduce the incident propagator n (r|r0 ), i.e., the probability density of finding a particle emitted at r0 entering the nth collision (n  1) at r. We have  n+1 (r|r0 ) = ps T (r → r)n (r |r0 )dr , (28) with 1 (r|r0 ) = T (r0 → r). We introduce then the incident collision density

0

10

(r|r0 ) = lim

N→∞

Pn (nV |x0 )

−2

10

−3

0

n (r|r0 ),

(29)

n=1

which can be interpreted as the particle equilibrium distribution [32,33,43,44]. Now, by making use of the formal Neumann series [43], from Eq. (28) it follows that the collision density satisfies the stationary integral transport equation  (r|r0 ) = ps T (r → r)(r |r0 )dr + T (r0 → r). (30)

−1

10

10

N 

10

20

nV

30

40

50

FIG. 5. (Color online) The arcsine law Pn (nV |x0 ) with exponential jump lengths, as a function of nV . The starting point is x0 = 0, and n= √ 50. Blue dots: ps = 1; dashed line: the asymptotic distribution 1/ nV (n − nV )π . Red triangles: ps = 0.95; dashed line: asymptotic Eq. (23).

The transport equation (30) can be understood as follows: at equilibrium, the stationary particle density entering a collision at r for a source emitting at r0 is given by the sum of all contributions entering a collision at r , being scattered and then transported to r, plus the contribution of the particles emitted from the source and never collided up to entering r. Now, by resorting to the relation between T and its adjoint T ∗ in Eq. (A4), and observing that at equilibrium (for isotropic source and scattering)   T (r0 → r )(r|r )dr = (r |r0 )T (r → r)dr , (31)

011132-6

COUNTING STATISTICS: A FEYNMAN-KAC PERSPECTIVE

Eq. (27) can be inverted (see Appendix), and gives       (m)  (r )dr (r |r1 ) n˜ (m−1) n˜ V (r1 ) = mps V

n (x0 )

2

0 10

n

n (x0 )

n (x0 )

2 0 10

n

Ck (r0 ) = k!

··· V

  k

dri (ri |ri−1 )

(35)

V i=1

are k-fold convolution (Kac) integrals over the equilibrium distribution (r|r0 ), and the coefficients Lm,k = ( mk )(m − 1)!/ (k − 1)! are the Lah numbers [54]. Observe that a result analogous to Eq. (34) has been derived in Ref. [32] building upon survival probabilities. V. FIRST AND SECOND MOMENT OF COLLISION NUMBER: EXAMPLES OF CALCULATIONS

To illustrate the approach proposed in the preceding section, we compute here the average collision number n(1) V n (x0 ) and (2) the second factorial moment nV n (x0 ) in a bounded domain V in one dimension. We choose an exponential displacement kernel T (x  → x) = s exp(−s|x − x  |)/2, with s = 1, which is often adopted as a simplified random-walk model (the socalled exponential flights) to describe gas dynamics, radiation propagation, or biological species migration [55–60]. In this respect, the average n(1) V n (x0 ) is a measure of the passage of the particles through the region V (the deposited energy, for instance), whereas the second moment n(2) V n (x0 ) is proportional to the incertitude on the average [32,33]. The calculations stemming from Eqs. (25) and (26) are rather cumbersome, so it is preferable to visually represent our results instead of writing down the explicit formulas. In Fig. 6, we display the behavior of the moments n(1) V n (x0 ) and  (x ) for the volume V being the interval [−R,R], with n(2) V n 0 R = 1. When ps = 1, the moments diverge as n increases, since exponential flights in one dimension are recurrent random walks, and revisit their starting point infinitely many times. In particular, we observe that for large n, the average grows as n1/2 , and the second moment as n1 , which is coherent with the results in Ref. [35] for Brownian motion. When ps < 1, they converge instead to an asymptotic value, which can be computed based on Eq. (34) by observing that the collision density for this example is

5

0 10

x0

ps = 1

4

0

5

where 

n

20 10

(2)

(34)

k=1

0

5

x0

nV

Lm,k psk−1 Ck (r0 ),

0 10

5

0 10

(1)

=

m 

5

ps = 1

nV

 n(m) V (r0 )

0

5

V

Hence, by induction we finally get the desired relation between the factorial moments and the equilibrium distribution, namely

10

(2)

1

nV

n (x0 ) (1)

nV

(32)

for m  1, where  is the solution of Eq. (30). Thus, by using Eqs. (26) and (31), from Eq. (32) it follows that   (m)      nV (r0 ) = m (r |r0 ) n˜ (m−1) (r )dr . (33) V



p s = 3/4

p s = 3/4

V

  (r1 ) + mV (r1 ) n˜ (m−1) V

PHYSICAL REVIEW E 85, 011132 (2012)

5

0 10

x0

0 10

0

5

n

5

0 10

x0

FIG. 6. (Color online) One-dimensional exponential flights in V = [−1,1]. Upper half: First (left) and second factorial moment (right) of nV as a function of n and x0 , with ps = 3/4. The asymptotic values for x0 = 0 are shown as dashed lines for reference. Lower half: First (left) and second factorial moment (right) of nV in the same domain, when ps = 1.

as discussed in Ref. [32]. Moreover, Fig. 6 shows that the moments decrease as the distance of the source x0 from the region V increases, as expected. As we have chosen here a (m) symmetric interval, we have n(m) V n (x0 ) = nV n (−x0 ), so we can plot the moments only for positive values of x0 . All results presented in this section have been verified by comparison with Monte Carlo simulations with 106 particles. Other kinds of boundary conditions (leakage, for instance, which implies that particles are lost upon crossing the frontier of V [32,33]) have also been successfully tested, but will not be presented here. VI. DIFFUSION LIMIT

To conclude our analysis, in this section we comment on the scaling limit of the discrete Feynman-Kac equation, which is achieved when nV is large, and at the same time the typical jump length  is vanishingly small. We set tV = nV dt and t = ndt, where dt is some small time scale, related to  by the usual diffusion scaling  2 = 2Ddt, the constant D playing the role of a diffusion coefficient. When T is not symmetric, so that displacements have mean μ, we further require μ = vdt, where the constant v is a velocity. By properly taking the limit of large nV and vanishing dt, tV converges to the residence time in V . The quantity nV can only be large if the absorption probability pa is small, and it is natural to set pa = λa dt, the quantity λa being an absorption rate per unit of dt. Observe that when both  and μ are small for any displacement kernel, we have the Taylor expansion  1 T (r → r)f (r )dr f (r) − μ∂r f (r) +  2 ∂r2 f (r), (37) 2



e− 1−ps |x−x0 | (x|x0 ) = √ , 2 1 − ps

(36)

where the first-order derivative vanishes if the kernel is symmetric. A similar expansion holds for the kernel T ∗ ,

011132-7

A. ZOIA, E. DUMONTEIL, AND A. MAZZOLO

namely



PHYSICAL REVIEW E 85, 011132 (2012) ACKNOWLEDGMENTS

The authors wish to thank Dr. F. Malvagi for useful discussions.

T ∗ (r → r0 )f (r )dr

1 f (r0 ) + μ∂r0 f (r0 ) +  2 ∂r20 f (r0 ). (38) 2 It is expedient to introduce the quantity Qt (u|r0 ) = Ft (e−u |r0 ), which is the moment generating function of tV = nV dt, i.e.,  m ∂m tV t (r0 ) = (−1)m m Qt (u|r0 )|u=0 , (39) ∂u when trajectories are observed up to t = ndt. Under the previous hypotheses, combining Eqs. (6) and (7) yields Qt+dt (u|r0 ) − Qt (u|r0 ) L∗r0 Qt (u|r0 )dt − uV (r0 )Qt (u|r0 )dt + λa dt,

(40)

where we have neglected all terms vanishing faster than dt, and L∗r0 = D∂r20 + v∂r0 − λa . Taking the limit dt → 0, we recognize then the Feynman-Kac equation for a Brownian motion with diffusion coefficient D, drift v, and absorption rate λa , namely ∂Qt (u|r0 ) = L∗r0 Qt (u|r0 ) − uV (r0 )Qt (u|r0 ) + λa . (41) ∂t In other words, in the diffusion limit the statistical properties of the hit number in V behave as those of the residence time of a Brownian motion, as is quite naturally expected on physical grounds [27,28,33]. Finally, from Eq. (39) stems the recursion property for the moments   ∂ tVm t (r0 )     = L∗r0 tVm t (r0 ) + mV (r0 ) tVm−1 t (r0 ), (42) ∂t in agreement with the results in Refs. [29,30] for Brownian motion. VII. CONCLUSIONS

In this paper, we have examined the behavior of the distribution Pn (nV |r0 ) of the number of collisions nV in a region V for a broad class of stochastic processes in absorbing and nonabsorbing media. Key to our analysis has been a discrete version of the Feynman-Kac formalism. We have shown that this approach is amenable to explicit formulas for Pn (nV |r0 ), at least for simple geometries and displacement kernels. The moments of the distribution have also been detailed, and their asymptotic behavior for large n has been related to the walker equilibrium distribution. Finally, the diffusion limit and the convergence to the Feynman-Kac formulas for Brownian motion have been discussed. We conclude by observing that a generalization of the present work to more realistic transport kernels, including anisotropic source and scattering, would be possible, for instance by resorting to the formalism proposed in Ref. [61]. Moreover, while in this paper we have focused on counting statistics, and therefore chosen V (r) to be the marker function of a given domain in phase space, the Feynman-Kac formalism can be adapted with minor changes to describing the statistics of other kinds of functionals, such as, for instance, hitting probabilities [5,35,62].

APPENDIX: THE STATIONARY MOMENT EQUATION

We want to solve an integral equation  f (r1 ) − ps T ∗ (r → r1 )f (r )dr = g(r1 )

(A1)

for the function f (r1 ), where g(r1 ) is known. We propose a solution in the form  (A2) f (r1 ) = ps g(r )(r |r1 )dr + g(r1 ) and ask which is the equation satisfied by the integral kernel (r |r1 ). By injecting Eq. (A2) into Eq. (A1), one obtains  g(r )(r |r1 )dr   = ps (r |r )T ∗ (r → r1 )g(r )dr dr  + T ∗ (r → r1 )g(r )dr . (A3) Recall that the adjoint and direct displacement kernels are related to each other by the scalar products   g(r) T (r → r)f (r )dr dr   (A4) = f (r) T ∗ (r → r)g(r )dr dr for any test functions f and g [43]. From the definition of the scalar product in Eq. (A4), it follows that the second term on the right-hand side of Eq. (A3) is given by   T ∗ (r → r1 )g(r )dr = T (r1 → r )g(r )dr . (A5) From Eqs. (A4) and (31), the first term on the right-hand side of Eq. (A3) becomes   (r |r )T ∗ (r → r1 )g(r )dr dr   = (r |r1 )T (r → r )g(r )dr dr . (A6) Therefore, (r |r1 ) obeys  g(r )(r |r1 )dr   = ps (r |r1 )T (r → r )g(r )dr dr  + T (r1 → r )g(r )dr ,

(A7)

which for the arbitrariness of g(r ) finally implies Eq. (30), i.e., the required kernel  satisfies the integral transport equation.

011132-8

COUNTING STATISTICS: A FEYNMAN-KAC PERSPECTIVE [1] B. D. Hughes, Random Walks and Random Environments, Vol. I (Clarendon, Oxford, 1995). [2] G. H. Weiss, Aspects and Applications of the Random Walk (North Holland, Amsterdam, 1994). [3] J. Ph. Bouchaud and A. Georges, Phys. Rep. 195, 127 (1990). [4] D. ben Avraham and S. Havlin, Diffusion and Reactions in Fractals and Disordered Systems (Cambridge University Press, Cambridge, UK, 2000). [5] S. Redner, A Guide to First-passage Processes (Cambridge University Press, Cambridge, UK, 2001). [6] S. Condamin et al., Nature (London) 450, 40 (2007). [7] R. Metzler and J. Klafter, Phys. Rep. 339, 1 (2000). [8] S. Condamin, O. B´enichou, and M. Moreau, Phys. Rev. Lett. 95, 260601 (2005). [9] O. B´enichou et al., J. Stat. Phys. 142, 657 (2011). [10] E. Barkai, J. Stat. Phys. 123, 883 (2006). [11] D. S. Grebenkov, Phys. Rev. E 76, 041139 (2007). [12] D. S. Grebenkov, J. Stat. Phys. 141, 532 (2010). [13] S. N. Majumdar and A. Comtet, Phys. Rev. Lett. 89, 060601 (2002). [14] C. Cercignani, The Boltzmann Equation and Its Applications (Springer, New York, 1988). [15] M. Weinberg and E. P. Wigner, The Physical Theory of Neutron Chain Reactors (University of Chicago Press, Chicago, 1958). [16] C. Jacoboni and P. Lugli, The Monte Carlo Method for Semiconductor Device Simulation (Springer, New York, 1989). [17] F. Bartumeus et al., J. Theor. Biol. 252, 43 (2008). [18] G. Le Ca¨er, J. Stat. Phys. 140, 728 (2010). [19] A. Zoia, A. Rosso, and S. N. Majumdar, Phys. Rev. Lett. 102, 120602 (2009). [20] M. F. Shlesinger, Nature (London) 443, 281 (2006). [21] M. Kac, Trans. Am. Math. Soc. 65, 1 (1949). [22] M. Kac, in Proceedings of the Second Berkeley Symposium on Mathematical, Statistics and Probability (University of California Press, Berkeley, 1951), pp. 189–215. [23] D. A. Darling and M. Kac, Trans. Am. Math. Soc. 84, 444 (1957). [24] M. Kac, Probability and Related Topics in Physical Sciences, Lectures in Applied Mathematics (Wiley, New York, 1957). [25] S. N. Majumdar, Curr. Sci. 89, 2076 (2005). [26] N. Agmon, J. Chem. Phys. 81, 3644 (1984). [27] A. M. Berezhkovskii, V. Zaloj, and N. Agmon, Phys. Rev. E 57, 3937 (1998). [28] O. B´enichou et al., Europhys. Lett. 70, 42 (2005). [29] N. Agmon, Chem. Phys. Lett. 497, 184 (2010). [30] N. Agmon, J. Phys. Chem. A 115, 5838 (2011). [31] A. Zoia, E. Dumonteil, and A. Mazzolo, Phys. Rev. E 83, 041137 (2011). [32] A. Zoia, E. Dumonteil, and A. Mazzolo, Phys. Rev. Lett. 106, 220602 (2011). [33] A. Zoia, E. Dumonteil, and A. Mazzolo, Phys. Rev. E 84, 021139 (2011).

PHYSICAL REVIEW E 85, 011132 (2012) [34] L. Turgeman, S. Carmi, and E. Barkai, Phys. Rev. Lett. 103, 190201 (2009). [35] S. Carmi, L. Turgeman, and E. Barkai, J. Stat. Phys. 141, 1071 (2010). [36] S. Carmi and E. Barkai, Phys. Rev. E 84, 061104 (2011). [37] A. Bar-Haim and J. Klafter, J. Chem. Phys. 109, 5187 (1998). [38] P. Del Moral, Feynman-Kac Formulae. Genealogical and Interacting Particle Systems with Applications (Springer, New York, 2004). [39] P. J. Fitzsimmons and J. Pitman, Stoch. Proc. Appl. 79, 117 (1999). [40] M. Iosifescu, Finite Markov Processes and their Applications, Wiley Series in Probability and Mathematical Statistics (Wiley, New York, 1980). [41] E. Cs`aki, J. Stat. Plann. Inference 34, 63 (1993). [42] S. Blanco and R. Fournier, Phys. Rev. Lett. 97, 230604 (2006). [43] J. Spanier and E. M. Gelbard, Monte Carlo Principles and Neutron Transport Problems (Addison-Wesley, Reading, MA, 1969). [44] I. Lux and L. Koblinger, Monte Carlo Particle Transport Methods: Neutron and Photon Calculations (CRC, Boca Raton, FL, 1991). [45] S. P. Meyn and R. L. Tweedie, Adv. Appl. Probab. 24, 542 (1992). [46] W. Feller, An Introduction to Probability Theory and Its Applications, 3rd ed. (Wiley, New York, 1970). [47] S. N. Majumdar, Physica A 389, 4299 (2010). [48] J. Lamperti, Trans. Am. Math. Soc. 88, 380 (1958). [49] S. Sabhapandit, S. N. Majumdar, and A. Comtet, Phys. Rev. E 73, 051102 (2006). [50] C. Godr`eche and J. M. Luck, J. Stat. Phys. 104, 489 (2001). [51] A. Baldassarri, J. P. Bouchaud, I. Dornic, and C. Godreche, Phys. Rev. E 59, R20 (1999). [52] S. Watanabe, Proc. Symp. Pure Math. 57, 157 (1995). [53] E. Sparre Andersen, Math. Scand. 2, 195 (1954). [54] J. Riordan, Introduction to Combinatorial Analysis (Princeton University Press, Princeton, NJ, 1980). [55] A. Mazzolo, Europhys. Lett. 68, 350 (2004). [56] G. Milton Wing, An Introduction to Transport Theory (Wiley, New York, 1962). [57] H. T. Hillen and G. Othmer, Siam J. Appl. Math 61, 751 (2000). [58] H. G. Othmer, S. R. Dunbar, and W. Alt, J. Math. Biol. 26, 263 (1988). [59] G. H. Weiss, Physica A 311, 381 (2002). [60] P. L. Krapivsky, S. Redner, and E. Ben-Naim, A Kinetic View of Statistical Physics (Cambridge University Press, Cambridge, UK, 2010). [61] A. Zoia, E. Dumonteil, and A. Mazzolo, Phys. Rev. E 84, 061130 (2011). [62] S. N. Majumdar, A. Rosso, and A. Zoia, Phys. Rev. Lett. 104, 020602 (2010).

011132-9

Home

Search

Collections

Journals

About

Contact us

My IOPscience

Branching exponential flights: travelled lengths and collision statistics

This article has been downloaded from IOPscience. Please scroll down to see the full text article. 2012 J. Phys. A: Math. Theor. 45 425002 (http://iopscience.iop.org/1751-8121/45/42/425002) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 132.166.149.109 The article was downloaded on 06/11/2012 at 09:12

Please note that terms and conditions apply.

IOP PUBLISHING

JOURNAL OF PHYSICS A: MATHEMATICAL AND THEORETICAL

doi:10.1088/1751-8113/45/42/425002

J. Phys. A: Math. Theor. 45 (2012) 425002 (19pp)

Branching exponential flights: travelled lengths and collision statistics Andrea Zoia, Eric Dumonteil, Alain Mazzolo and Sameh Mohamed CEA/Saclay, DEN/DM2S/SERMA/LTSD, Bˆat. 470, 91191 Gif-sur-Yvette Cedex, France E-mail: [email protected]

Received 8 July 2012, in final form 11 September 2012 Published 8 October 2012 Online at stacks.iop.org/JPhysA/45/425002 Abstract The evolution of several physical and biological systems, ranging from neutron transport in multiplying media to epidemics or population dynamics, can be described in terms of branching exponential flights, a stochastic process which couples a Galton–Watson birth–death mechanism with random spatial displacements. Within this context, one is often called to assess the length V that the process travels in a given region V of the phase space, or the number of visits nV to this same region. In this paper, we address this issue by resorting to the Feynman–Kac formalism, which allows characterizing the full distribution of V and nV and in particular deriving explicit moment formulas. Some other significant physical observables associated to V and nV , such as the survival probability, are discussed as well, and results are illustrated by revisiting the classical example of the rod model in nuclear reactor physics. PACS numbers: 05.40.Fb, 05.40.−a, 02.50.−r (Some figures may appear in colour only in the online journal)

1. Introduction Consider a single walker initially emitted from a point source at time τ0 = 0 at position r0 , with velocity v0 . Once emitted, the walker undergoes a sequence of displacements (at constant speed), separated by collisions with the surrounding medium. When the scattering centre encountered by the travelling particle are spatially uniform, the inter-collision lengths are exponentially distributed [1, 2], so that the displacements from r to r in direction ω = v/|v| between any two collisions obey the probability density T (r → r|ω) = σ (r , v) e−

 ω·(r−r ) 0

σ (r +sω,v) ds

,

(1)

with v = |v| [3, 4]. The quantity σ (r, v) represents the interaction rate per unit length and takes the name of total cross section: σ (r, v) typically depends on the particle position and speed, and is proportional to the probability of particle–medium interaction along a straight 1751-8113/12/425002+19$33.00 © 2012 IOP Publishing Ltd

Printed in the UK & the USA

1

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

V

q

Figure 1. An illustration of branching exponential flights starting from a point source q and traversing a volume V in phase space.

line, carrying units of the inverse of a length [3]. At each collision, the incident particle disappears, and k particles (the descendants) are emitted with probability pk (r, v), whose velocities are randomly redistributed in angle and intensity according to a given probability density Ck (v → v|r), which in principle can vary as a function of the number k of descendants [3]. Each descendant will then behave as the mother particle, and undergo a new sequence of displacements and collisions, giving thus rise to a branched structure, as illustrated in figure 1. As a particular case, when pk = δk,1 we recover the well known Pearson random walk [1, 2]. Branching random flights as described above lie at the heart of physical and biological modelling [5, 6], and are key to the description of neutron transport in multiplying media and nucleon cascades [7], evolution of biological populations [8], diffusion of reproducing bacteria [9], and mutation-propagation of genes [10], just to name a few. For a detailed survey, ranging from the pioneering work by Galton and Watson on the extinction probability of birth–death processes to the recent developments, see, e.g., [5, 7]. A central question for random walks is to determine the occupation statistics of the stochastic paths in a given portion V of the phase space [11–22]. For exponential flights, the two natural observables of the system are the number nV of occurred visits to the volume V and the total length V travelled in V [3, 4, 17, 18, 20]. In reactor physics, for instance, knowledge of V allows assessing the neutron flux due to the chain reaction and hence the deposited power or the number of radiation-induced structural defects [7, 23]. In a model of epidemics outbreak, nV corresponds to the number of infections in a region V as a function of the position of the initial infected person (as long as the number of infected people is small, so that nonlinear effects due to the depletion of the susceptibles can be neglected, and that spatial displacements can be described by a simple random walk [24]). The quantity nV occurs also in population genetics, where one might be interested in quantifying the number nV of mutations of a given kind V , starting from a single character, as a function of the number of generations (this is closely related to the Ewens’ formula for the mutation partition, when mutations are allowed to be recurrent [25]). The goal of this paper is to characterize the statistical properties of the number of visits nV and of the travelled lengths V for branching exponential flights. This paper is structured as follows. In section 2 we first focus on the average quantities V  and nV . Then, in section 3 and 4 we assess the full distribution of V and nV , respectively, by resorting to the Feynman–Kac formalism. In particular, we show that recursive formulas for the higher moments of the travelled lengths and for the number of visits can be easily 2

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

derived based on this approach. Some other significant physical observables associated to V and nV are discussed in section 5. Then, in section 6 we illustrate the proposed formalism on one-dimensional exponential flights, the so-called rod model, and support our findings with Monte Carlo simulations. Perspectives are finally presented in section 7. Technical details are left to appendix A, appendix B and appendix C. 2. The average observables In the following, we introduce a few simplifying hypotheses, whose main advantage is to keep notation to a minimum, yet retaining the key physical mechanisms. Thus, we assume that displacements are performed at a constant speed v = v0 , i.e. that only the walker directions ω do change after collisions. We furthermore assume that the medium is spatially homogeneous, so that pk (r, v) and σ (r, v) can be taken to be constant. Finally, we assume that the probability density of the directions for the outgoing particles is isotropic, independent of the number of emitted descendants, namely, 1 , (2) Ck (v → v|r) = C(ω → ω|r) = d where the normalization factor d = 2π d/2 /(d/2) is the surface of the unit sphere in dimension d. Branching exponential flights, as defined above, are a Markovian stochastic process that can be observed both as a function of time τ and discrete generations n (this latter case corresponds to recording the particle position and direction at collision events only). Markovianity is granted by the fact that displacements between collisions are exponentially distributed [3, 4, 7], and implies that knowledge of the phase space variables r, ω at time τ or generation n is sufficient to determine the future evolution of the walker1. To begin with, we address first the average physical observables of branching exponential flights, which in most cases provide a reasonable first-order estimate of the system evolution. 2.1. The average total travelled length V  Let N(r, ω, τ |r0 , ω0 ) be the average number of particles that at time τ are found in the phase space element dr dω around r, ω, starting with a single particle emitted at r0 in direction ω0 , i.e. N(r, ω, 0|r0 , ω0 ) = q = δ(r − r0 )δ(ω − ω0 ). Consider a displacement in a small time dτ along a line oriented as ω: the time variation of N(r, ω, τ |r0 , ω0 ) reads  d N = −vσ N + ν1 dω vσC(ω → ω|r)N(r, ω , τ |r0 , ω0 ), (3) dτ  the quantity ν1 = k kpk being the average number of secondary particles emitted per collision event. This equation can be understood as a mass balance: in dτ , N decreases because of particles that at a time rate vσ interact, and thus change direction, and increases because of particles that, travelling in another direction ω , have a collision, are multiplied by a factor ν1 , and change their direction to ω [23]. By using dτd = ∂τ∂ + vω · ∇r , we finally have  ∂ N + vω · ∇r N = −vσ N + ν1 dω vσC(ω → ω|r)N(r, ω , τ |r0 , ω0 ), (4) ∂τ which is a Boltzmann-like conservation equation for the average particle density in phase space. Boundary conditions on N(r, ω, τ |r0 , ω0 ) depend on the specific problem under analysis. 1 Conversely, knowledge of the position r alone does not ensure Markovianity, as would instead be the case for branching Brownian motion [26–29].

3

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

Actually, instead of N it is often common to introduce the quantity φ = Nv, which takes the name of particle flux [23]. The stationary behaviour of the particle  ∞ density is provided by integrating over time, and for the stationary flux φ(r, ω|r0 , ω0 ) = 0 dτ φ(r, ω, τ |r0 , ω0 ) we get in particular  (5) ω · ∇r φ + σ φ = ν1 dω σC(ω → ω|r)φ(r, ω |r0 , ω0 ) + q. Equation (5) can be recast in the more compact formula (6) Lφ(r, ω|r0 , ω0 ) = −q,    where L = −ω · ∇r − σ + ν1 dω σC(ω → ω|r) takes the name of (forward) transport operator [3, 23]. The quantity φ(r, ω|r0 , ω0 ) can be interpreted as the stationary density of the total length travelled by the particles in the phase space element dr dω around r, ω: hence, the average travelled length in a given volume V of phase space will be given by   (7) V (r0 , ω0 ) = dr dωV (r, ω)φ(r, ω|r0 , ω0 ), where V (r, ω) denotes the marker function of the phase space volume V , i.e. V (r, ω) = 1 when r, ω belong to V , and V (r, ω) = 0 elsewhere. 2.2. The average total number of visits nV  If generations are considered instead of time, a different mass balance equation for exponential flights can be established. Let ψn (r, ω|r0 , ω0 ) be the average number of particles that enter a collision at r, having direction ω, at the nth generation. Then, the following recursive formula can be established   ψn+1 = ν1

dr

dω T (r → r|ω)C(ω → ω|r )ψn (r , ω |r0 , ω0 ),

(8)

with the initial condition ψ1 (r, ω|r0 , ω0 ) = T (r0 → r|ω0 ) [23]. The term ψ1 represents the average particle number entering a collision at the first generation (the so-called uncollided density). The stationary behaviour of ψn (r, ω|r0 , ω0 ) is obtained by summing over ∞ all generations: as customary, we define the collision density as being ψ (r, ω|r0 , ω0 ) = n=1 ψn (r, ω|r0 , ω0 ) [23], and we thus get the integral equation   (9) ψ = ν1 dr dω T (r → r|ω)C(ω → ω|r )ψ (r , ω |r0 , ω0 ) + ψ1 . The quantity ψ (r, ω|r0 , ω0 ) physically represents the stationary density of the number of particles entering a collision at r, ω: then, the average number of visits to a given volume V of phase space will be given by   (10) nV (r0 , ω0 ) = dr dωV (r, ω)ψ (r, ω|r0 , ω0 ). From nV (r0 , ω0 ) and V (r0 , ω0 ) being two average observables of the same stochastic process, and thus closely related to each other, one can easily imagine that the quantities ψ and φ must be intimately connected as well. Actually, it can be shown that ψ (r, ω|r0 , ω0 ) = σ φ(r, ω|r0 , ω0 ) [3, 23]: this provides the relation between the stationary densities N, ψ and φ, and implies in particular that equation (9) can be equivalently recast  into equation (5), by setting ψ = σ φ. As a consequence, we have also nV (r0 , ω0 ) = dr dωV (r, ω)σ φ(r, ω|r0 , ω0 ). The approach proposed in this section so to assess the behaviour of the average quantities V  and nV  cannot be straightforwardly extended to higher moments (which are often necessary to quantify the statistical fluctuations around the average), nor to other observables. In the following, we show that this difficulty can be overcome by resorting to the Feynman–Kac formalism, which allows characterizing the full distribution of the variables V and nV . 4

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

3. Total travelled length in V We formally define the total length V (t ) travelled by a branching exponential flight in a given volume V of the phase space, when observed up to a time t, as  t V (r , ω )v dt  , (11) V (t ) = 0

where the integral is intended over all the branching paths of a single realization up to time t. The quantity V (t ) is clearly a stochastic variable, which depends on the realizations of the underlying process, as well as on the initial conditions. Instead of studying the probability density function Pt (V |r0 , ω0 ), it is more convenient to introduce the associated moment generating function Qt (s|r0 , ω0 ) =  e−sV (t ) (r0 , ω0 ),

(12)

where s is the transformed variable with respect to V . Basically, Qt (s|r0 , ω0 ) can be interpreted as the Laplace transform of Pt (V |r0 , ω0 ). We derive then a backward equation for Qt (s|r0 , ω0 ) by closely following the approach originally proposed by Kac for Brownian motion [30], based on Feynman path integrals2. As detailed in appendix A, the resulting backward Feynman–Kac equation relates the generating function Qt of the travelled length V to the generating function G[z] = k pk zk of the offspring number k, and reads 1 ∂ Qt = ω0 · ∇r0 Qt − σ Qt − sV (r0 , ω0 )Qt + σ G[C∗ {Qt }], v ∂t where C∗ {Qt } is a shorthand for the direction-averaged Qt  C∗ {Qt } = C∗ (ω0 → ω0 |r0 )Qt (s|r0 , ω0 ) dω0 ,

(13)

(14)

C∗ being the adjoint probability density with respect to C. Equation (13) is completed by the initial condition Q0 (s|r0 , ω0 ) = 1 and by the appropriate boundary conditions, which depend on the problem at hand. 3.1. Moment equations Equation (13) is a partial differential equation with a nonlinear integral term, for which explicit solutions are hardly available. Moreover, one would still need to invert the solution Qt so to obtain the probability density of V in the direct space. A somewhat simpler approach consists in deriving the corresponding moment equations3: by the definition of Qt , the moments of the travelled length can be obtained from  m ∂m (15) V t (r0 , ω0 ) = (−1)m m Qt (s|r0 , ω0 )|s=0 . ∂s By taking the mth derivative of equation (13) and resorting to the Fa`a di Bruno’s formula for multiple derivatives of composite functions [39], we get the following recursive formula for the moments of the trace length m         1 ∂  m V t = L∗ Vm t + mV (r0 , ω0 ) Vm−1 t + σ (16) ν j Bm, j C∗ Vi t , v ∂t j=2 2 The Feynman–Kac formalism more generally applies to continuous-time Markov processes (see, e.g., [31–34]), and has recently been extended to non-Markovian walks [35–38]. 3 For Brownian motion without branching, a similar approach was proposed by Kac [33] and later extended, e.g., in [14–16]. Similarly, the moments of exponential flights without branching are discussed, e.g., in [17–20, 22].

5

J. Phys. A: Math. Theor. 45 (2012) 425002

for m  1, where L∗ = ω0 · ∇r0 − σ + σ ν1

A Zoia et al



C∗ (ω0 → ω0 |r0 ) dω0

(17)

is the (backward) transport operator adjoint to L [23]. Here Bm, j [zi ] = Bm, j [z1 , z2 , . . . , zm− j+1 ] are the Bell’s polynomials [39], and ν j = k(k − 1) · · · (k − j + 1) are the falling factorial moments of the descendant number, with ν0 = 1. Bell polynomials4 commonly appear in connection with the combinatorics of branched structures [39]: this might give a hint about their role in equation (16), which relates the moments Vm t of the travelled length to the moments ν j of the descendant number. The recurrence is initiated with the conditions V0 t = 1 (from normalization), and Vm 0 = 0. Observe that V1 t depends only on ν1 , V2 t on ν1 and ν2 , and so on. 3.2. Stationary behaviour Most often, the observation time t is much longer than the characteristic time scale of the system dynamics, which means that trajectories are followed up to t → ∞. In this case, the time derivative in equation (16) vanishes, provided that the moment Vm t does not diverge when t → ∞. We therefore get a recursive formula for the stationary moments Vm  = limt→∞ Vm t , namely,   L∗ Vm (r0 , ω0 ) = −Um−1 (r0 , ω0 ), (18) where

m       Um−1 (r0 , ω0 ) = mV (r0 , ω0 ) Vm−1 + σ ν j Bm, j C∗ Vi

(19)

j=2

can be interpreted as a (known) source term that depends at most on the moments of order m − 1. Now, from L∗ being the adjoint operator with respect to L, if one can solve equation (6) for a point source q = δ(r−r0 )δ(ω −ω0 ) and obtain the corresponding stationary flux φ, then equation (18) can be explicitly inverted, and gives   Vm (r0 , ω0 ) = dr dωφ(r, ω|r0 , ω0 )Um−1 (r, ω), (20) which means that the stationary moments of the travelled length can be obtained by convoluting the stationary flux with the source term Um−1 . As a particular case, when pk2 = 0 we reobtain the simpler recursive formula derived in [19, 22] for non-branching exponential flights. Finally,  for the average length travelled in V , i.e. m = 1, we recover the formula V1  = dr dωV (r, ω)φ(r, ω|r0 , ω0 ), since U0 (r, ω) = V (r, ω). 4. Total number of visits to V We address then the statistical properties of the total number of visits nV (n) performed by a branching exponential flight in a given volume V of the phase space, when observed up to the nth generation5. We formally define  V (ri , ωi ), (21) nV (n) = i

The first few polynomials read: B0,0 = 1; B1,1 [z1 ] = z1 ; B2,1 [z1 , z2 ] = z2 , B2,2 [z1 , z2 ] = z21 ; B3,1 [z1 , z2 , z3 ] = z3 , B3,2 [z1 , z2 , z3 ] = 3z1 z2 , B3,3 [z1 , z2 , z3 ] = z31 ; . . .. 5 The analysis of discrete (isotropic) branching walks can be extended to arbitrary flight length densities T , as shown in [40, 41], by resorting to an integral formulation. Here we stuck however to exponential flights, which imposes T as given in equation (1). 4

6

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

where the sum is intended over all the points visited by the branching path up to entering the nth generation. We adopt the convention that the source is not taken into account. The quantity nV (n) is again a stochastic variable depending on the realizations of the underlying process and on the initial conditions. Similarly as done for V , instead of studying the probability Pn (nV |r0 , ω0 ), it is more convenient to introduce the associated moment generating function Qn (u|r0 , ω0 ) =  e−unV (n) (r0 , ω0 ),

(22)

where u is the transformed variable with respect to nV . As shown in appendix B, the backward discrete Feynman–Kac equation for Qn (u|r0 , ω0 ) relates the generating function Qn of the number of visits nV to the generating function G of the offspring number k, and reads − ω0 · ∇r0 Qn+1 (u|r0 , ω0 ) + σ Qn+1 (u|r0 , ω0 ) = σ e−uV (r0 ,ω0 ) G[C∗ {Qn }], with the initial condition



Q1 (u|r0 , ω0 ) =

e−uV (r1 ,ω0 ) T ∗ (r1 → r0 |ω0 ) dr1 ,

(23)

(24)

 and the appropriate boundary conditions. We have used the shorthand C∗ {Qn } = C∗ (ω0 → ω0 |r1 )Qn (u|r1 , ω0 ) dω0 . As a particular case, when particles cannot move (i.e. σ → ∞), spatial dependences can be neglected and from equation (23) one recovers the counting statistics of a simple Galton–Watson process observed up to the nth generation [5]. 4.1. Moment equations Equation (23) is a nonlinear integro-differential and finite differences equation. Similarly as in the case of V , the analysis of the distribution of nV can be simplified by deriving the corresponding moment equations: by the definition of Qn , the moments of the number of visits can be obtained from  m ∂m (25) nV n (r0 , ω0 ) = (−1)m m Qn (u|r0 , ω0 )|u=0 . ∂u Then, by taking the mth derivative of equation (23), we get the following recursive formula for the moments of number of visits m         − ω0 · ∇r0 nVm n+1 + σ nVm n+1 = σ ν j Bm, j C∗ nVi n +σ

m  k=1

j=1

m−k     m V (r0 , ω0 ) ν j Bm−k, j C∗ nVi n , k j=0

(26)

for m  1. Equation (26) relates the moments nVm n of the number of visits to the moments ν j of the descendant number. Observe that nV1 n depends only on ν1 , nV2 n on ν1 and ν2 , and 0 so on. The  recurrence∗is initiated with the conditions nV n = 1 (from normalization), and m nV 1 = V (r1 , ω0 )T (r1 → r0 |ω0 ) dr1 . 4.2. Stationary behaviour Most often, one considers trajectories that are followed up to n → ∞, provided that the moment nVm n does not diverge. We therefore get a recursive formula for the stationary moments nVm  = limn→∞ nVm n , namely, L∗ nVm (r0 , ω0 ) = −Hm−1 (r0 , ω0 ),

(27) 7

J. Phys. A: Math. Theor. 45 (2012) 425002

where Hm−1 = σ

m m−k m          m V (r0 , ω0 ) +σ ν j Bm−k, j C∗ nVi ν j Bm, j C∗ nVi . k k=1 j=0 j=2

A Zoia et al

(28)

is a source term, and we have singled out the term of order m in the The   Bell polynomials. quantity Hm−1 is closely related to Um−1 , and the contribution σ mj=2 ν j Bm, j C∗ {nVi } is common to both. When m = 1, we have H0 = σU0 . As done for the moments of travelled lengths, equation (27) can be explicitly inverted in terms of the corresponding stationary flux φ, and gives   nVm (r0 , ω0 ) = dr dωφ(r, ω|r0 , ω0 )Hm−1 (r, ω), (29) which means that the stationary moments of the number of visits can be obtained by convoluting the stationary flux with the source term Hm−1 . In particular, for the average number of visits to V (m = 1) we recover the formula   nV1  = dr dωV (r0 , ω0 )σ φ(r, ω|r0 , ω0 ), since H0 (r, ω) = σV (r0 , ω0 ). The close relation between nV1  and V1  carries over to moments of any order m  1 (via equation (27)), although the simple proportionality that holds true for the average does not apply to higher moments. 5. Other physical observables The Feynman–Kac approach proposed in the previous sections allows fully characterizing the distribution of the travelled length and of the number of visits. In the following, we show that a number of other interesting features of the process can be assessed by relying upon the same formalism, with minimal modifications, and we discuss some significant examples. 5.1. Probability of never visiting a region V For instance, one might be interested in determining the probability Rn (r0 , ω0 ) that a branching exponential flight coming from a point source at r0 , ω0 never collides in a given domain V , up to generation n. This is intimately related to the well-known gambler’s ruin problem [1, 2]. It follows that Rn (r0 , ω0 ) = Pn (n V = 0|r0 , ω0 ). Now, if we introduce the probability generating i function unV (n) (r0 , ω0 ) = i u Pn (nV = i|r0 , ω0 ), then by construction Rn (r0 , ω0 ) = nV (n) (r0 , ω0 )|u=0 . By comparing unV (n) (r0 , ω0 ) to Qn (u|r0 , ω0 ) = e−unV (n) (r0 , ω0 ), it u is apparent that Rn (r0 , ω0 ) satisfies − ω0 · ∇r0 Rn+1 (r0 , ω0 ) + σ Rn+1 (r0 , ω0 ) = σ V¯ (r0 , ω0 )G[C∗ {Rn }],

(30)

¯ 0 ) = 1 − V (r0 , ω0 ). As for the initial conditions, we have R1 (r0 , ω0 ) = where V (r0 , ω ¯ V (r1 , ω0 )T ∗ (r1 → r0 |ω0 ) dr1 . When n → ∞, we get the stationary probability equation − ω0 · ∇r0 R(r0 , ω0 ) + σ R(r0 , ω0 ) = σ V¯ (r0 , ω0 )G[C∗ {R}],

(31)

where we have set R(r0 , ω0 ) = limn→∞ Rn (r0 , ω0 ). 5.2. Survival probability The quantities V and nV , as defined above, are cumulative, i.e. their knowledge integrates the whole history of the walk, from the source to the measure time, or generation. Sometimes, it is necessary to provide information about local (instantaneous) countings, namely the number mV (t ) and mV (n) of particles in the volume V when an observation is performed at time t, or at 8

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

generation n, respectively (see, e.g., [7]). As shown in appendix A, the probability generating function Wt (s|r0 , ω0 ) = smV (t )  satisfies 1 ∂ Wt (s|r0 , ω0 ) = ω0 · ∇r0 Wt (s|r0 , ω0 ) − σWt (s|r0 , ω0 ) + σ G[C∗ {Wt }], (32) v ∂t with initial condition W0 (s|r0 , ω0 ) = sV (r0 ,ω0 ) . Equation (32) is known in reactor physics as the P´al–Bell equation [42, 43, 7]. Furthermore, as detailed in appendix B, the probability generating function Wn (u|r0 , ω0 ) = umV (n)  satisfies (33) − ω0 · ∇r0 Wn+1 (u|r0 , ω0 ) + σWn+1 (u|r0 , ω0 ) = σ G[C∗ {Wn }],  V (r ,ω ) ∗ 1 0 T (r1 → r0 |ω0 ) dr1 . with initial condition W1 (u|r0 , ω0 ) = u Assume now that V is bounded, i.e. particles are lost upon leaving the volume: then, we might want to assess the survival probability at time t or generation n, due to the interplay between the branching mechanism and the spatial leakages. As particles cannot re-enter V after crossing the boundaries, if mV (t ) = 0, then also mV (t  ) = 0 for t   t (i.e. the process goes to extinction), and the same holds for mV (n). Hence, by definition, the probability of having zero particles in the volume V at a time t is given by Wt (s = 0|r0 , ω0 ), which equivalently yields the probability that extinction is reached for times smaller than t, since V is bounded [7]. We define then the survival probability as St (r0 , ω0 ) = 1 − Wt (s = 0|r0 , ω0 ), which by direct substitution in equation (32) satisfies 1 ∂ St (r0 , ω0 ) = ω0 · ∇r0 St (r0 , ω0 ) − σ St (r0 , ω0 ) − σ F[C∗ {St }], (34) v ∂t ∞ where we set F[z] = k=1 αk zk , with αk = (−1)k νk /k! [43]. At the boundaries, St must vanish when ω0 is directed towards the exterior of V . The probability of having zero particles in the volume V at generation n is similarly given by Wn (u = 0|r0 , ω0 ), which therefore yields the probability that extinction is reached for generations smaller than n. We define then the associated survival probability as Sn (r0 , ω0 ) = 1 − Wn (s = 0|r0 , ω0 ), which by direct substitution in equation (33) satisfies ω0 · ∇r0 Sn+1 (r0 , ω0 ) − σ Sn+1 (r0 , ω0 ) = σ F[C∗ {Sn }],

(35)

where Sn must again vanish at the boundaries when ω0 is directed towards the exterior of V . Finally, by either taking the limit S = limt→∞ St (r0 , ω0 ) or S = limn→∞ Sn (r0 , ω0 ), respectively, the probability of ultimate survival S(r0 , ω0 ) satisfies ω0 · ∇r0 S − σ S = σ F[C∗ {S}].

(36)

Observe that in principle S = 0 is always a solution to equation (36), which would imply almost sure extinction, i.e. a vanishing probability that infinitely long branching chains exist in V . From the Galton–Watson theory, we know that when ν1  1 the branching process goes to extinction even in the absence of spatial leakages, hence S = 0. However, when ν1 > 1 the branching process would grow indefinitely, and it may happen that the particle loss due to finite geometry is not sufficient to compensate the population growth. In this case, the solution S = 0 would become unstable, and St (or Sn ) would converge towards a nontrivial survival probability S = S∞ > 0. The stability analysis of the solution S = 0 can be carried out by introducing a small perturbation, for instance in the form Sˆ X (t )Y (r0 , ω0 ), the amplitude  > 0 being a small positive constant, with X (t ) > 0 and Y (r0 , ω0 ) > 0. Now, if we inject Sˆ into equation (34), and take the limit  → 0, we obtain an equation for the perturbation amplitude 1 1 ∂X (t ) ω0 · ∇r0 Y (r0 , ω0 ) − σY (r0 , ω0 ) + σ ν1C∗ {Y } = , (37) v X (t ) ∂t Y 9

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

where at the numerator of the right-hand side we recognize the adjoint operator L∗ . From the separation of the variables, equation (37) shows that the evolution of the perturbation amplitude with respect to time is determined by the ratio λ = L∗Y/Y , hence by the eigenvalue equation L∗Y = λY . The spectrum of the eigenvalues of L∗ depends on the geometry of V and on the boundary conditions. If all eigenvalues λ are negative, the amplitude of the small perturbation will shrink in time, so that eventually S = 0; if on the contrary at least one eigenvalue is positive, then the small perturbation will grow in time, which means that S = 0 is unstable, and eventually S = S∞ . For a given branching process, the crossover between these two regimes depends on the size and shape of the volume V , and generally speaking one would expect that S = S∞ for a volume size larger than some critical value Vc [23, 7], which is attained when λ = 0. When ν1  1, Vc = ∞. The stability analysis of Sˆ XnY (r0 , ω0 ) in equation (35) leads to Xn+1 σ ν1C∗ {Y } . =− Xn ω0 · ∇r0 Y (r0 , ω0 ) − σY (r0 , ω0 )

(38)

The evolution of the perturbation amplitude is determined then by the eigenvalue equation ω0 · ∇r0 Y (r0 , ω0 ) − σY (r0 , ω0 ) + σβν1 C∗ {Y } = 0: when β < 1 the perturbation shrinks and when β > 1 the perturbation grows, the crossover occurring for a critical volume Vc such that β = 1. 6. The rod model In order to illustrate the previous results, we revisit here a relevant example inspired by reactor physics. Neutrons in a multiplying medium undergo branching exponential flights, where conceptually radiative capture represents absorption (p0 ), scattering corresponds to p1 and fission to pk2 : in realistic situations, the cross sections and all the other physical parameters depend on energy and position, and angular distributions are often mildly or even strongly anisotropic [23]. To simplify the matter, we assume that cross sections are constant, scattering is isotropic and particles travel at an average constant speed v = 1. Moreover, we address a one-dimensional configuration, namely the interval [0, L], where only two angular directions (forward or backward) are allowed: all physical quantities will be then denoted with a superscript ± according to whether ω is taken along the x-axis (+), or in the opposite direction (−). Despite these many simplifications, the so-called rod model yet captures the key features of neutron transport, and has been widely adopted in neutronics [5, 44]. To begin with, we impose leakage boundary conditions at x = 0 and x = L and study the survival probabilities. By setting Y + (x0 ) = Y (x0 , ω0 = +) and Y − (x0 ) = Y (x0 , ω0 = −), the eigenvalue equation L∗Y = λY gives rise to a system of two coupled linear differential equations σ ν1 + ∂ + (Y (x0 ) + Y − (x0 )) = λY + (x0 ), Y (x0 ) − σY + (x0 ) + ∂x0 2 ∂ − σ ν1 + (Y (x0 ) + Y − (x0 )) = λY − (x0 ), − Y (x0 ) − σY − (x0 ) + ∂x0 2

(39)

where we have used C∗ {Y } = (Y + (x0 ) + Y − (x0 ))/2, thanks to isotropy. The general integrals Y + (x0 ) = Y1 (x0 ; κ1 , κ2 ) and Y − (x0 ) = Y2 (x0 ; κ1 , κ2 ) of equation (39) can be easily obtained as a combination of exponential functions, up to two integration constants, say κ1 and κ2 , that are to be imposed by boundary conditions. Due to the leakages, at the boundaries we have Y + (L) = 0 and Y − (0) = 0. The solutions Y + = Y − = 0 clearly satisfy the system 10

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

in equation (39), together with boundary conditions. Searching for nontrivial solutions leads then to solving

Y1 (L; κ1 , κ2 ) = 0 =0 (40) det Y2 (0; κ1 , κ2 ) = 0 with respect to the basis of κ1 and κ2 . This in turn gives an implicit equation for the eigenvalues λ as a function of the system parameters L, σ , and ν1 , namely,   √  λσ − ν21 sinh(Lσ λσ (λσ − ν1 )) cosh(Lσ λσ (λσ − ν1 )) + = 0, (41) √ λσ (λσ − ν1 ) where we have set the dimensionless variable λσ = λ/σ + 1. When λ = λ(L, σ, ν1 ) < 0, then S = 0 and the neutron chain reaction will die out because of leakages and possibly absorptions; when on the contrary λ > 0, then S = S∞ (x0 ), and the chain reaction will diverge, as particles born from fission are not sufficiently compensated by leakages and absorptions; the crossover between these two regimes is reached for λ = 0, which therefore defines the portion of the parameter space (L, σ , ν1 ) for which the branching process will attain ultimate extinction. When ν1  1, all λ stay negative. When ν1 > 1, imposing λ = 0 in equation (41) yields the explicit relation   −1 √ 1 2 tan ν1 −1 , (42) Lc = Lλ=0 = √ σ ν1 − 1 where Lc is the maximum system size such that for L > Lc there exists a finite survival probability that the neutron population will grow indefinitely. By resorting to the same arguments, the analysis of the eigenvalue equation ω0 ·∇r0 Y (r0 , ω0 )−σY (r0 , ω0 )+ σβν1 C∗ {Y } = 0 leads to   

(β − ν1 ) sinh Lσ 1 − ν1  2 β ν1 + cosh Lσ 1 − = 0, (43) √ β β(β − ν1 ) which closely resembles equation (41). Imposing β = 1 not surprisingly yields again equation (42). In a previous paper we derived equation (42) based on the analysis of the moment nV1  as a function of the system parameters [41]. The survival probabilities in equations (34) and (35) can be integrated numerically: in figures 2 and 3 we compare the resulting curves (as a function of time or generations, respectively) with Monte Carlo simulations for different configurations. In particular, once the probabilities pk have been chosen, by varying the rod size L it is possible to impose L < Lc or L > Lc . In the former case, the survival probabilities converge to zero independent of the starting direction, as expected, whereas in the latter the survival probabilities saturate to an asymptotic value S∞ that depends on the starting point as well as on the initial direction of the walker. Observe that St+ up to a time of the order of τ + |L − x0 |/v does not feel the effects of the boundaries, yet, and the same holds for St− up to a time of the order of τ − x0 /v. Therefore, we expect St+ St− up to min(τ + , τ − ). For the configuration where L > Lc , the asymptotic survival probability as a function of starting point and initial direction is displayed in figure 4. We examine then the statistics of the travelled length V and number of visits nV via the moment equations (16) and (26), respectively. In particular, we consider here the average and second moment, which in most cases are sufficient to characterize the typical behaviour of the stochastic variables and their dispersion. The equations are given in appendix C, for a single neutron initially emitted at x0 . We consider the same geometrical configuration as above, and take V (x, ω) = V (x) for x ∈ [0, L], i.e. we measure lengths and collisions independent of the local direction of the particles. As for boundary conditions, leakages impose Vm t+ (L) = 0 11

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

1

S t (x 0 )

0.8

0.6

0.4

0.2

0 0

5

10

15

t

20

Figure 2. Survival probabilities St± (x0 ) for p0 = 0.2, p1 = 0.3, and p2 = 0.5 (ν1 = 1.3, and Lc = 3.9987). Blue crosses and red circles: St+ (x0 ) and St− (x0 ), respectively, with x0 = 1.75 and L = 2 (λ < 0). Green squares and black triangles: St+ (x0 ) and St− (x0 ), respectively, with x0 = 3.75 and L = 6 (λ > 0). Solid lines are numerical integrals of equation (34), symbols Monte Carlo simulations with 106 histories. Dashed curves: asymptotic survival probabilities S∞ from equation (36).

1

S n (x 0 )

0.8

0.6

0.4

0.2

0 0

5

10

n

15

20

25

Figure 3. Survival probabilities Sn± (x0 ) for p0 = 0.2, p1 = 0.3, and p2 = 0.5 (ν1 = 1.3, and Lc = 3.9987). Blue crosses and red circles: Sn+ (x0 ) and Sn− (x0 ), respectively, with x0 = 1.75 and L = 2 (λ < 0). Green squares and black triangles: Sn+ (x0 ) and Sn− (x0 ), respectively, with x0 = 3.75 and L = 6 (λ > 0). Solid lines are numerical integrals of equation (35), symbols Monte Carlo simulations with 106 histories. Dashed curves: asymptotic survival probabilities S∞ from equation (36). m − and Vm t− (0) = 0; similarly, nVm + n (L) = 0 and nV n (0) = 0. For L > Lc the moments would diverge as time and generations increase. Therefore, we choose a configuration such that L < Lc and in figures 5 and 6 we plot the evolution of the moments of the travelled length and number of visits, respectively. Observe that the moments depend on the starting point as

12

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

0.4 0.35 0.3

S (x 0 )

0.25 0.2 0.15 0.1 0.05 0 0

1

2

3

4

x0

5

6

Figure 4. Asymptotic survival probabilities S± (x0 ) for p0 = 0.2, p1 = 0.3, and p2 = 0.5 (ν1 = 1.3, and Lc = 3.9987). Blue circles and red squares: S+ (x0 ) and S− (x0 ), respectively,with L = 6 (λ > 0). Solid lines are numerical integrals from equation (36), symbols Monte Carlo simulations with 106 histories. 3.5

3

2

m V

(x 0 )

2.5

1.5

1

0.5

0 0

1

2

t

3

4

5

Figure 5. Average and second moment of V , for x0 = 1.75 and L = 2, with p0 = 0.2, p1 = 0.3, p2 = 0.5 (ν1 = 1.3 and L < Lc ). Blue crosses: V1 t+ (x0 ); red circles: V1 t− (x0 ). Green squares: V2 t+ (x0 ); black triangles: V2 t− (x0 ). Solid lines are numerical integrals from equation (16), symbols Monte Carlo simulations with 106 histories. Dashed lines are the asymptotic limits in equation (20).

well as on the initial direction of the neutron, and level off to an asymptotic value, which is the signature of the process going to ultimate extinction. Furthermore, again for Vm t we have Vm t+ Vm t− up to min(τ + , τ − ). 13

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

4

3.5

3

nm V (x 0 )

2.5

2

1.5

1

0.5

0 1

2

3

4

n

5

6

7

8

Figure 6. Average and second moment of nV , for x0 = 1.75 and L = 2, with p0 = 0.2, p1 = 0.3, 1 − p2 = 0.5 (ν1 = 1.3 and L < Lc ). Blue crosses: nV1 + n (x0 ); red circles: nV n (x0 ). Green squares: 2 − (x ). Solid lines are numerical integrals from equation (26), nV2 + (x ); black triangles: n 0 0 n V n symbols Monte Carlo simulations with 106 histories. Dashed lines are the asymptotic limits in equation (29).

Observe finally that the rod model has been widely adopted to describe, among others, gas dynamics or biological species migration [45–49], so that the results described in this Section concerning an application in reactor physics could perhaps be of interest in other areas of science as well.

7. Summary and conclusions In this paper we have examined the statistics of travelled lengths and number of collisions for branching exponential flights. Moment formulas have been derived by resorting to the backward Feynman–Kac formalism, based on a minimal number of simplifying hypotheses. Moreover, we have shown that this same formalism can be extended with slight modifications to the analysis of other physical observables, such as the survival probability. The proposed formulas have been compared to Monte Carlo simulations for an example of one-dimensional transport inspired by reactor physics, and an excellent agreement was found. A generalization to more complex transport problems would also be possible, by relaxing for instance the requirement on the isotropy of the scattering kernel [50] and by introducing energy and space dependent cross sections. We conclude by observing that, although throughout this paper we have used our knowledge of ψ and φ so to assess the statistical properties of the observables V and nV , the inverse is also possible: for instance, knowledge of the average total travelled length in V allows inferring the average particle flux in that volume, and similarly knowledge of the total number of visits to V allow inferring the average collision density in that volume. 14

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

Appendix A. The Feynman–Kac equations for Qt and Wt Consider a single walker initially at r0 , ω0 at observation time t = 0. We want to write an equation for the moment generating function Qt = e−sV (t ) . Assume an observation time t + dt: this can be split into a first interval, from 0 to dt, and then a second interval from dt to t + dt. The only requirement is that the process is Markovian: after dt the particle continues its path without memory of the past. We start by observing that in a vanishing small time interval dt, from the definition of the underlying process, only two mutually exclusive events are possible: either the particle does not interact with the medium, in which case the walker keeps going in the same direction by a space interval dr0 = vω0 dt, or the particle interacts, in which case the walker disappears and gives rise to k descendants at the same position, with random directions ωk obeying the same probability density C. From the definition of the cross section σ , the former event happens at a rate 1 − σ v dt, whereas the latter at a rate σ v dt. If no particles are emitted (p0 ), the trajectory is terminated, and no further contribution is added to V . When k  1 (identical) particles are generated, the probability that the contribution to the total travelled length coming from each walker adds up precisely to V is given by the convolution of the probability that the first particle spends a length 1 , the second 2 , and the kth a length V − 1 − 2 − · · ·. In the transformed space, the convolution products amount to a simple product of generating functions. This simple argument leads to the equation Qt+dt (s|r0 , ω0 ) = (1 − σ v dt ) e−svV (r0 ,ω0 ) dt Qt (s|r0 + dr0 , ω0 )     + σ v dt p0 + p1 Qt s|r0 , ω1 + p2 Qt (s|r0 , ω21 )Qt (s|r0 , ω22 ) + · · ·], (A.1) where brackets denote expectation with respect to the random directions ω k . Now, if we suppose that the descendant directions are independent, the expectation of a product of random variables becomes the product of expectations, so that we are led to Qt+dt = (1 − σ v dt ) e−svV (r0 ,ω0 ) dt Qt (s|r0 + dr0 , ω0 ) + σ v dtG[Qt (s|r0 , ω0 )], (A.2) 2 where G[z] = p0 + p1 z + p2 z + · · · is the generating function associated to pk . Observe that the average over the random directions can be expressed in terms of the associated probability density as  Qt (s|r0 , ω0 ) =

C∗ (ω0 → ω0 |r0 )Qt (s|r0 , ω0 ) dω0 ,

(A.3)



where C is formally the adjoint density with respect to C. As a shorthand, we will denote C∗ {Qt } = C∗ (ω0 → ω0 |r0 )Qt (s|r0 , ω0 ) dω0 . Now, when dt is small, at the leading order we have Qt (s|r0 + dr0 , ω0 ) = Qt (s|r0 , ω0 ) + vω0 · ∇r0 Qt dt + · · ·, along the direction of ω0 . Furthermore, for vanishing dt we have exp(−svV (r0 , ω0 ) dt ) = 1 − svV (r0 , ω0 ) dt + · · ·. By recollecting all terms we then get Qt+dt = Qt + vω0 · ∇r0 Qt dt − vσ Qt dt − svV (r0 , ω0 ) dt + σ v dtG[C∗ {Qt }]. (A.4) By dividing by v dt and taking the limit dt → 0, we finally obtain the backward Feynman–Kac equation for the moment generating function Qt , namely 1 ∂ Qt = ω0 · ∇r0 Qt − σ Qt − sV (r0 , ω0 )Qt + σ G[C∗ {Qt }]. (A.5) v ∂t Now, from the same argument as above, it follows that the probability generating function Wt = smV (t )  satisfies Wt+dt (s|r0 , ω0 ) = (1 − σ v dt )Wt (s|r0 + dr0 , ω0 ) + σ v dt[p0 + p1 Wt (s|r0 , ω1 ) + p2 Wt (s|r0 , ω21 )Wt (s|r0 , ω22 ) + · · ·], (A.6) hence 1 ∂ (A.7) Wt = ω0 · ∇r0 Wt − σWt + σ G[C∗ {Wt }]. v ∂t 15

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

Appendix B. The discrete Feynman–Kac equation for Qn Let a single walker be initially at r0 , ω0 . When considering generations, it is more convenient to begin with a single particle entering its first collision with coordinates r1 , ω0 , and we denote by Q˜ n (u|r1 , ω0 ) = e−unV (n) (r1 , ω0 ) the corresponding moment generating function. The separation between r1 and r0 , at a first glance somewhat artificial, is actually due to the special role of the source: a particle emitted from the source is just transported to the first collision point, and cannot be absorbed nor multiplied at r0 [40, 41]. When entering the collision at r1 , k particles are created, with probability pk , and random directions. Exponential flights are Markovian at collision points, which allows splitting each of the k subsequent trajectories into a first jump, from r1 to rk in direction ωk , and a branching path from rk to the positions held at the (n + 1)th generation. The displacement rk − r1 obeys the jump length density T , and the direction ωk the density C. If k = 0, the trajectory ends at r1 and there will be no further events contributing to nV . Hence, we have Q˜ n+1 (u|r1 , ω0 ) = p0 e−uV (r1 ,ω0 ) + p1 e−uV (r1 ,ω0 ) Q˜ n (u|r1 , ω1 ) +p2 e−uV (r1 ,ω0 ) Q˜ n (u|r21 , ω21 )Q˜ n (u|r22 , ω22 ) + · · · ,

(B.1)

where expectation is taken with respect to the random displacements and directions, and the term e−uV (r1 ,ω0 ) can be singled out because it is not stochastic. The terms at the righthand side in equation (B.1) can be understood as follows: the probability that k identical and indistinguishable particles (born at r1 ) give rise to nV collisions in V is given by the convolution product that the first makes n1 collisions, the second n2 , ..., and the kth nV − n1 − n2 − · · ·. In the transformed space, this convolution becomes a simple product of generating functions. If we assume that the descendant particles are independent, the expectation of the products in equation (B.1) becomes the product of the expectations. Now, observe that the average over displacements and directions can be expressed in terms of the associated densities, namely   (B.2) Q˜ n (u|rk , ωk ) = C∗ (ω0 → ω0 |r1 ) T ∗ (r1 → r1 |ω0 )Q˜ n (u|r1 , ω0 ) dr1 dω0 , where T ∗ is the adjoint density associated to T [40]. Intuitively, T ∗ displaces the walker backward in time. Observe in particular that the first collision coordinates r1 , ω0 obey the probability density T ∗ (r1 → r0 |ω0 ), namely,  Qn (u|r0 , ω0 ) = Q˜ n (u|r1 , ω0 )T ∗ (r1 → r0 |ω0 ) dr1 . (B.3)  Therefore, by using C∗ {Qn } = C∗ (ω0 → ω0 |r1 )Qn (u|r1 , ω0 ) dω0 as above, we obtain the discrete Feynman–Kac equation in integral form, namely Q˜ n+1 (u|r1 , ω0 ) = e−uV (r1 ,ω0 ) G[C∗ {Qn }].

(B.4)

Finally, by integrating over T ∗ both sides of equation (B.4) we get  Qn+1 (u|r0 , ω0 ) = dr1 T ∗ (r1 → r0 |ω0 ) e−uV (r1 ,ω0 ) G[C∗ {Qn }].

(B.5)

It can be shown [23] that integral equations in the form  f (r0 , ω0 ) = dr1 T ∗ (r1 → r0 |ω0 )g(r1 , ω0 )

(B.6)

can be equivalently recast into ω0 · ∇r0 f (r0 , ω0 ) − σ f (r0 , ω0 ) + σ g(r0 , ω0 ) = 0. 16

(B.7)

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

Therefore, equation (B.4) gives the discrete Feynman–Kac equation in integro-differential form − ω0 · ∇r0 Qn+1 (u|r0 , ω0 ) + σ Qn+1 (u|r0 , ω0 ) = σ e−uV (r0 ,ω0 ) G[C∗ {Qn }].

(B.8)

Now, from the same argument as above, it follows that the probability generating function Wn = umV (n)  satisfies W˜ n+1 (u|r1 , ω0 ) = p0 + p1 W˜ n (u|r1 , ω1 ) + p2 W˜ n (u|r21 , ω21 )W˜ n (u|r22 , ω22 ) + · · · , with W˜ n (r1 , ω0 ) = u

mV (n)

(B.9)

(r1 , ω0 ). Finally,

− ω0 · ∇r0 Wn+1 (u|r0 , ω0 ) + σWn+1 (u|r0 , ω0 ) = σ G[C∗ {Wn }].

(B.10)

Appendix C. The rod model equations From equation (16), the average travelled length obeys     1 ∂  1 V t = L∗ V1 t + V (x0 ) V0 t , v ∂t which gives then  + σ ν1  1 +  1 −  1 ∂  1 + ∂  1 + V t = V t + V t + V (x0 ), V t − σ V1 t + v ∂t ∂x0 2  − σ ν1  1 +  1 −  1 ∂  1 − ∂  1 −  V t + V t + V (x0 ).  =− − σ V1 t + v ∂t V t ∂x0 V t 2 For the second moment of the travelled length we have     2   1 ∂  2 V t = L∗ V2 t + 2V (x0 ) V1 t + σ ν2C∗ V1 t , v ∂t which gives  + σ ν1  2 +  2 −  1 ∂  2 + ∂  2 + V t = V t + V t V t − σ V2 t + v ∂t ∂x0 2   + σ ν2  1 +  1 − 2 V t + V t , + 2V (x0 ) V1 t + 4  − σ ν1  2 +  2 −  1 ∂  2 − ∂  2 − V t = − V t + V t V t − σ V2 t + v ∂t ∂x0 2  − σ ν2  1 +  1 − 2 V t + V t . + 2V (x0 ) V1 t + 4 From equation (26), the average number of visits obeys    

  − ω0 · ∇r0 nV1 n+1 + σ nV1 n+1 = σ ν1C∗ nV1 n + V (x0 ),

(C.1)

(C.2)

(C.3)

(C.4)

(C.5)

which gives  + ∂  1 + σ ν1  1 +  1 −  nV n + nV n + V (x0 ), nV n+1 + σ nV1 n+1 = ∂x0 2  − ∂  1 − σ ν1  1 +  1 −  nV n + nV n + V (x0 ). n + σ nV1 n+1 = ∂x0 V n+1 2 For the second moment of the number of visits we have    

  2   − ω0 · ∇r0 nV2 n+1 + σ nV2 n+1 = σ ν1C∗ nV2 n + σ ν2C∗ nV1 n  

+ 2σ ν1V (x0 )C∗ nV1 n + σV (x0 ), −

(C.6)

(C.7) 17

J. Phys. A: Math. Theor. 45 (2012) 425002

which leads to  + ∂  2 + σ ν1  2 +  2 −  nV n + nV n n − + σ nV2 n+1 = ∂x0 V n+1 2     σ ν2 1 + σ ν1  1 +  1 −  − 2 nV n + nV1 n + nV n + nV n + σV (x0 ) + 4 2  − ∂  2 + σ ν1  2 +  2 −  nV n + nV n nV n+1 + σ nV2 n+1 = ∂x0 2 σ ν2  1 +  1 − 2 σ ν1  1 +  1 −  nV n + nV n + nV n + nV n + σV (x0 ). + 4 2

A Zoia et al

(C.8)

References [1] Hughes B D 1995 Random Walks and Random Environments vol I (Oxford: Clarendon) [2] Weiss G H 1994 Aspects and Applications of the Random Walk (Amsterdam: North-Holland) [3] Spanier J and Gelbard E M 1969 Monte Carlo Principles and Neutron Transport Problems (Reading, MA: Addison-Wesley) [4] Lux I and Koblinger L 1991 Monte Carlo Particle Transport Methods: Neutron and Photon Calculations (Boca Raton, FL: CRC Press) [5] Harris T E 1963 The Theory of Branching Processes (Berlin: Springer) [6] Jagers P 1975 Branching Processes with Biological Applications (Wiley Series in Probability and Mathematical Statistics) (London: Wiley) [7] P´azsit I and P´al L 2008 Neutron Fluctuations: A Treatise on the Physics of Branching Processes (Oxford: Elsevier) [8] Lawson D J and Jensen H J 2007 Neutral evolution in a biological population as diffusion in phenotype space: reproduction with local mutation but without selection Phys. Rev. Lett. 98 098102 [9] Golding I, Kozlovsky Y, Cohen I and Ben-Jacob E 1998 Studies of bacterial branching growth using reactiondiffusion models for colonial development Physica A 260 510 [10] Sawyer S and Fleischman J 1979 Maximum geographic range of a mutant allele considered as a subtype of a Brownian branching random field Proc. Natl Acad. Sci. 76 872 [11] Condamin S, B´enichou O, Tejedor J, Voituriez R and Klafter J 2007 First passage times in complex scale invariant media Nature 450 40 [12] Condamin S, B´enichou O and Moreau M A 2005 First-passage times for random walks in bounded domains Phys. Rev. Lett. 95 260601 [13] Grebenkov D S 2007 Residence times and other functionals of reflected Brownian motion Phys. Rev. E 76 041139 [14] Berezhkovskii A M, Zaloj V and Agmon N 1998 Residence time distribution of a Brownian particle Phys. Rev. E 57 3937 [15] Agmon N 2010 The residence time equation Chem. Phys. Lett. 497 184 [16] Agmon N 2011 Single molecule diffusion and the solution of the spherically symmetric residence time equation J. Phys. Chem. A 115 5838 [17] Blanco S and Fournier R 2006 Short-path statistics and the diffusion approximation Phys. Rev. Lett. 97 230604 [18] Mazzolo A 2004 Properties of diffusive random walks in bounded domains Europhys. Lett. 68 350 [19] B´enichou O, Coppey M, Moreau M, Suet P H and Voituriez R 2005 Averaged residence times of stochastic motions in bounded domains Europhys. Lett. 70 42 [20] Zoia A, Dumonteil E and Mazzolo A 2011 Collision densities and mean residence times for d-dimensional exponential flights Phys. Rev. E 83 041137 [21] Zoia A, Dumonteil E and Mazzolo A 2011 Collision-number statistics for transport processes Phys. Rev. Lett. 106 220602 [22] Zoia A, Dumonteil E and Mazzolo A 2011 Residence time and collision statistics for exponential flights: the rod problem revisited Phys. Rev. E 84 021139 [23] Bell G I and Glasstone S 1970 Nuclear Reactor Theory (New York: Van Nostrand Reinhold Company) [24] J Bailey N T 1957 The Mathematical Theory of Infectious Diseases and its Applications (London: Griffin) [25] Bertoin J 2010 A limit theorem for trees of alleles in branching processes with rare neutral mutations Stoch. Process. Appl. 120 678 [26] Brunet E and Derrida B 2009 Statistics at the tip of a branching random walk and the delay of traveling waves Europhys. Lett. 87 60010 18

J. Phys. A: Math. Theor. 45 (2012) 425002

A Zoia et al

[27] Derrida B and Simon D 2007 The survival probability of a branching random walk in presence of an absorbing wall Europhys. Lett. 78 60006 [28] Derrida B and Spohn H 1988 Polymers on disordered trees, spin glasses and traveling waves J. Stat. Phys. 51 817 [29] van Saarloos W 2003 Front propagation into unstable states Phys. Rep. 386 29 [30] Kac M 1949 On distributions of certain Wiener functionals Trans. Am. Math. Soc. 65 1 [31] Kac M 1951 On some connections between probability theory and differential and integral equations Proc. 2nd Berkeley Symp. on Mathematical Statistics and Probability (Berkeley, CA: University of California Press) pp 189–215 [32] Darling D A and Kac M 1957 On occupation times for Markoff processes Trans. Am. Math. Soc. 84 444 [33] Kac M 1957 Probability and related topics in physical sciences (Lectures in Applied Mathematics) (New York: Wiley) [34] Majumdar S N 2005 Brownian functionals in physics and computer science Curr. Sci. 89 2076 [35] Barkai E 2006 The survival probability of a branching random walk in presence of an absorbing wall J. Stat. Phys. 123 883 [36] Turgeman L, Carmi S and Barkai E 2009 Fractional Feynman–Kac equation for non-Brownian functionals Phys. Rev. Lett. 103 190201 [37] Carmi S, Turgeman L and Barkai E 2010 On distributions of functionals of anomalous diffusion paths J. Stat. Phys. 141 1071 [38] Carmi S and Barkai E 2011 Fractional Feynman–Kac equation for weak ergodicity breaking Phys. Rev. E 84 061104 [39] Pitman J 2006 Combinatorial Stochastic Processes (Berlin: Springer) [40] Zoia A, Dumonteil E and Mazzolo A 2012 Counting statistics: a Feynman–Kac perspective Phys. Rev. E 85 011132 [41] Zoia A, Dumonteil E and Mazzolo A 2012 Discrete Feynman–Kac formulas for branching random walks Europhys. Lett. 98 40012 [42] P´al L 1958 On the theory of stochastic processes in nuclear reactors Il Nuovo Cimento 7 (suppl. 1) 25 [43] Bell G I 1965 On the stochastic theory of neutron transport Nucl. Sci. Eng. 21 390 [44] Milton Wing G 1962 An Introduction to Transport Theory (New York: Wiley) [45] Kac M 1974 A stochastic model related to the Telegrapher’s equation Rocky Mt. J. Math. 4 497 [46] Hillen H T and Othmer G 2000 The diffusion limit of transport equations derived from velocity jump processes SIAM J. Appl. Math. 61 751 [47] Othmer H G, Dunbar S R and Alt W 1988 Models of dispersal in biological systems J. Math. Biol. 26 263 [48] Weiss G H 2002 Some applications of persistent random walks and the telegrapher’s equation Physica A 311 381 [49] Krapivsky P L, Redner S and Ben-Naim E 2010 A Kinetic View of Statistical Physics (Cambridge: Cambridge University Press) [50] Zoia A, Dumonteil E and Mazzolo A 2011 Collision statistics for random flights with anisotropic scattering and absorption Phys. Rev. E 84 061130

19

Spatial extent of an outbreak in animal epidemics Eric Dumonteila, Satya N. Majumdarb, Alberto Rossob,1, and Andrea Zoiaa a DEN/DM2S/SERMA/LTSD, Commissariat à l’Energie Atomique/Saclay, 91191 Gif-sur-Yvette Cedex, France; and bUnité Mixte de Recherche 8626, Centre National de la Recherche Scientifique—Université Paris-Sud, LPTMS, 91405 Orsay Cedex, France

Characterizing the spatial extent of epidemics at the outbreak stage is key to controlling the evolution of the disease. At the outbreak, the number of infected individuals is typically small, and therefore, fluctuations around their average are important: then, it is commonly assumed that the susceptible–infected–recovered mechanism can be described by a stochastic birth–death process of Galton– Watson type. The displacements of the infected individuals can be modeled by resorting to Brownian motion, which is applicable when long-range movements and complex network interactions can be safely neglected, like in the case of animal epidemics. In this context, the spatial extent of an epidemic can be assessed by computing the convex hull enclosing the infected individuals at a given time. We derive the exact evolution equations for the mean perimeter and the mean area of the convex hull, and we compare them with Monte Carlo simulations. branching Brownian motion

| extreme value statistics

M

odels of epidemics traditionally consider three classes of populations—namely, the susceptible (S), the infected (I), and the recovered (R). This framework provides the basis of the so-called SIR model (1, 2), a fully connected mean-field model where the population sizes of the three species evolve with time t by the coupled nonlinear equations: dS/dt = −βIS, dI/dt = βIS − γI, and dR/dt = γI. Here, γ is the rate at which an infected individual recovers, and β denotes the rate at which it transmits the disease to a susceptible (3–5). In the simplest version of these models, the recovered cannot be infected again. These rate equations conserve the total population size I(t) + S(t) + R(t) = N; one assumes that, initially, there is only one infected individual, and the rest of the population is susceptible: I(0) = 1, S(0) = N − 1, and R(0) = 0. Of particular interest is the outbreak stage (i.e., the early times of the epidemic process), when the susceptible population is much larger than the number of infected or recovered. During this regime, for large N, the susceptible population hardly evolves and stays S(t) ∼ N; therefore, nonlinear effects can be safely neglected, and one can just monitor the evolution of the infected population alone: dI/dt ∼ (βN − γ)I(t). Thus, the ultimate fate of the epidemics depends on the key dimensionless parameter R0 = βN/ γ, which is called the reproduction rate. If R0 > 1, the epidemic explodes and invades a finite fraction of the population; if R0 < 1, the epidemic goes to extinction, and in the critical case R0 = 1 the infected population remains constant (6–8). This basic deterministic SIR has been generalized to a variety of both deterministic as well as stochastic models, and distinct advantages and shortcomings are discussed at length in refs. 9–11. Generally speaking, stochastic models are more suitable in the presence of a small number of infected individuals, when fluctuations around the average may be relevant (9, 10). During the outbreak of epidemics, the infected population is typically small: in this regime, the evolution can be modeled by resorting to a stochastic birth–death branching process of the Galton–Watson type for the number of infected (9–11), where each infected individual transmits the disease to another individual at rate βN and recovers at rate γ. The epidemic may become endemic for R0 > 1 and becomes extinct for R0 < 1, whereas for R0 = 1 fluctuations are typically long-lived and completely control the time evolution of the infected population (3–5). www.pnas.org/cgi/doi/10.1073/pnas.1213237110

How far in space can an epidemic spread? Branching processes alone are not sufficient to describe an outbreak, and spatial effects must necessarily be considered (1, 4, 12–14). Quantifying the geographical spread of an epidemic is closely related to the modeling of the population displacements. Brownian motion is often considered as a paradigm for describing the migration of individuals, despite some well-known shortcomings: for instance, finite speed effects and preferential displacements are neglected. Most importantly, a number of recent studies have clearly shown that individuals geographically far apart can actually be closely related to each other through the so-called small-world connections, such as air traffic, public transportation, and so on; then, the spread of epidemics among humans cannot be realistically modeled without considering these complex networks of interconnections (15–18). Nonetheless, Brownian motion provides a reasonable basis for studying disease propagation in animals and possibly plants (here, pathogen vectors are insects) (5). Although theoretical models based on branching Brownian motion have provided important insights on how the population size grows and fluctuates with time in a given domain (1, 4, 13, 14), another fundamental question is how the spatial extension of the infected population evolves with time. Assessing the geographical area traveled by a disease is key to the control of epidemics, which is especially true at the outbreak, when confinement and vaccination could be most effective. One major challenge in this very practical field of disease control is how to quantify the area that needs to be quarantined during the outbreak. For animal epidemics, this issue has been investigated experimentally (for instance, in the case of equine influenza) (19). The most popular and widely used method consists of recording the set of positions of infected animals and, at each time instant, constructing a convex hull (i.e., a minimum convex polygon surrounding the positions; a precise definition of the convex hull is given below) (Fig. 1). The convex hull at time t then provides a rough measure of the area over which the infections have spread up to time t. The convex hull method is also used to estimate the home range of animals (i.e., the territory explored by a herd of animals during their daily search for food) (20, 21). In this paper, we model the outbreak of an epidemic as a Galton– Watson branching process in presence of Brownian spatial diffusion. Despite infection dynamics being relatively simple, the corresponding convex hull is a rather complex function of the trajectories of the infected individuals up to time t, whose statistical properties seem to be a formidable problem. Our main goal is to characterize the time evolution of the convex hull associated with this process, particularly its mean perimeter and area. The rest of the paper is organized as follows. We first describe precisely the model and summarize our main results. Then, we provide a derivation of our analytical findings supported by extensive numerical simulations. We conclude with perspectives and discussions. Some details of the computations are relegated to SI Text.

Author contributions: E.D., S.N.M., A.R., and A.Z. performed research and wrote the paper. The authors declare no conflict of interest. This article is a PNAS Direct Submission. 1

To whom correspondence should be addressed. E-mail: [email protected].

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. 1073/pnas.1213237110/-/DCSupplemental.

PNAS | March 12, 2013 | vol. 110 | no. 11 | 4239–4244

PHYSICS

Edited by Susan N. Coppersmith, University of Wisconsin, Madison, WI, and approved January 29, 2013 (received for review July 31, 2012)

Day 1

Day 2

Day 3

O

O

O

y

y

y

x

x

x

Model and Main Results Consider a population of N individuals uniformly distributed in a 2D plane, with a single infected at the origin at the initial time. At the outbreak, it is sufficient to keep track of the positions of the infected, which will be marked as particles. The dynamics of the infected individuals are governed by the following stochastic rules. In a small time interval dt, each infected alternatively has three actions.

Consider now a particular history of the assembly of the trajectories of all of the infected individuals up to time t, starting from a single infected initially at the origin (Fig. 1). For every realization of the process, we construct the associated convex hull C. To visualize the convex hull, imagine stretching a rubber band so that it includes all of the points of the set at time t inside it and then releasing the rubber band. It shrinks and finally gets stuck when it touches some points of the set; therefore, it cannot shrink any more. This final shape is precisely the convex hull associated with this set. In this paper, we show that the mean perimeter 〈L(t)〉 and the mean area 〈A(t)〉 of the convex hull are ruled by two coupled nonlinear partial differential equations that can be solved numerically for all t (Fig. 2). The asymptotic behavior for large t can be determined analytically for the critical (R0 = 1), subcritical (R0 < 1), and supercritical (R0 > 1) regimes. In particular, in the critical regime, the mean perimeter saturates to a finite value as t → ∞, whereas the mean area diverges logarithmically for large t:

i) The infected individual recovers with probability γdt. This event corresponds to the death of a particle with rate γ. ii) The infected individual infects, by local contact, a new susceptible individual from the background with probability bdt. This event corresponds to the birth of a new particle that can subsequently diffuse. The originally infected particle still remains infected, which means that the trajectory of the originally infected particle branches into two new trajectories. The rate b replaces the rate βN in the SIR or the Galton– Watson process mentioned before. iii) The infected individual diffuses with diffusion constant D with probability 1 − (γ + b)dt. The coordinates {x(t), y(t)} of the particle get updated to the new values {x(t) + ηx(t)dt, y (t) + ηy(t)dt}, where ηx(t) and ηy(t) are independent Gaussian white noises with zero mean and correlators 〈ηx(t)ηx(t′)〉 = 2Dδ(t − t′), 〈ηy(t)ηy(t′)〉 = 2Dδ(t − t′), and 〈ηx(t)ηy(t′)〉 = 0.

sffiffiffiffiffiffi   6D + O t−1=2 hLðt → ∞Þi = 2π γ

hAðt → ∞Þi =

=1 R0

1.01

2000

Prob(A,t)

< A(t) >

R0=

R0=1.15

3500

2500

R0=0.99

1500

10

-2

10

-3

R0=0.85

10

1

10

2

10 t

3

10

4

[2]

10-4 10

-5

10

-6

10

0

24πD ln t + Oð1Þ: 5γ

10-7

1000 500

[1]

and

The only dimensionless parameter in the model is the ratio R0 = b/γ (i.e., the basic reproduction number).

3000

Fig. 1. The snapshots of the trajectories of an assembly of infected individuals at the epidemics outbreak at three different times (schematic) starting from a single infected at the origin O at time t = 0. Individuals that are still infected at a given time t are displayed as red dots, whereas individuals already recovered are shown as black dots. The convex hull enclosing the trajectories (shown as a dashed line) is a measure of the geographical area covered by the spreading epidemic. As the epidemic grows in space, the associated convex hull also grows in time.

10

5

-8

10-9 100

101

102

103 A

104

105

106

Fig. 2. (Left) The average area 〈A(t)〉 of the convex hull as a function of the observation time. For the parameter values, we have chosen D = 1/2 and b = R0γ = 0.01. We considered five different values of R0. We have obtained these results by two different methods. (i) One method is by the numerical integration of Eqs. 9 and 14 and using Eq. 15. These results are displayed as solid lines. (ii) The other method is by Monte Carlo simulations of the 2D branching Brownian motion with death with the same parameters averaged over 105 samples. Monte Carlo simulations are displayed as symbols. The dashed lines represent the asymptotic limits as given in Eq. 2 for the critical case R0 = 1. Additional details of the numerical simulations are provided in SI Text. (Right) Distribution of the area of the convex hull for the critical case R0 = 1, with γ = 0.01 and D = 1/2, as obtained by Monte Carlo simulations with 2 × 106 realizations. The dashed line corresponds to the power-law (24πD/5γ)A−2 as predicted by expression 20.

4240 | www.pnas.org/cgi/doi/10.1073/pnas.1213237110

Dumonteil et al.

This prediction seems rather paradoxical at a first glance. How can the perimeter of a polygon be finite while its area is divergent? The resolution to this paradox owes its origin precisely to statistical fluctuations. The results in Eqs. 1 and 2 are true only on average. Of course, for each sample, the convex hull has a finite perimeter and a finite area. However, as we later show, the probability distributions of these random variables have power-law tails at long time limits. For instance, although Prob(L, t → ∞) ∼ L−3 for large L (thus leading to a finite first moment), the area distribution behaves as Prob(A, t → ∞) ∼ A−2 for large A. Hence, the mean area is divergent as t → ∞ (Fig. 2). When R0 ≠ 1, the evolution of the epidemic is controlled by a characteristic time t*, which scales like t* ∼ jR0 − 1j−1. For times t < t* the epidemic behaves as in the critical regime. In the subcritical regime, for t > t*, the quantities 〈L(t)〉 and 〈A(t)〉 rapidly saturate, and the epidemic goes eventually to extinction. In contrast, in the supercritical regime (which is the most relevant for virulent epidemics that spread fast), a new timedependent behavior emerges when t > t*, because there exists a finite probability (namely 1 − 1/R0) that the epidemic never goes to extinction (Fig. 3). More precisely, we later show that   D  E 1 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi L t  t* = 4π 1 − DγðR0 − 1Þt [3] R0 and   D  E 1 A t  t* = 4π 1 − Dγ ðR0 − 1Þt2 : R0

[4]

The ballistic growth of the convex hull stems from an underlying traveling front solution of the nonlinear equation governing the convex hull behavior. Indeed, the prefactor of perimeter growth pthe ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi is proportional to the front velocity v* = 2 DγðR0 − 1Þ. As time increases, the susceptible population decreases because of the growth of the infected individuals: this depletion effect leads to a breakdown of the outbreak regime and a slowing down of the epidemic propagation. Statistics of the Convex Hull Characterizing the fluctuating geometry of C is a formidable task even in absence of branching (b = 0) and death (γ = 0), i.e., 5

hLðtÞi = 2πhxm ðtÞi

[5]



 hAðtÞi = π x2m ðtÞ − y2 ðtm Þ ;

[6]

and

where xm is the maximum displacement of our 2D stochastic process in the x direction up to time t, tm is the time at which the maximum displacement along x direction occurs, and y(tm) is the ordinate of the process at tm (i.e., when the displacement along the x direction is maximal). A schematic representation is provided in Fig. 4, where the global maximum xm is achieved by one single infected individual, whose path is marked in red. A crucial observation is that the y component of the trajectory connecting O to this red path is a regular 1D Brownian motion. Hence, given tm and t, clearly 〈y2(tm)〉 = 2D〈tm〉. Therefore, 

hAðtÞi = π x2m ðtÞ − 2Dhtm ðtÞi :

[7]

Eqs. 5 and 7 thus show that the mean perimeter and area of the epidemics outbreak are related to the extreme statistics of a 1D branching Brownian motion with death. Indeed, if we can compute the joint distribution Pt(xm, tm), we can, in turn, compute the three moments 〈xm〉, hx2m i, and 〈tm〉 that are needed in Eqs. 5 and 7. We show below that this calculation can be performed exactly.

1 PHYSICS

10

purely for diffusion process in 2Ds. Major recent breakthroughs have, nonetheless, been obtained for diffusion processes (22, 23) by a clever adaptation of the Cauchy integral geometric formulae (24, 25) for the perimeter and area of any closed convex curve in 2Ds. In fact, the problem of computing the mean perimeter and area of the convex hull of any generic 2D stochastic process can be mapped, using Cauchy formulae, to the problem of computing the moments of the maximum and the time at which the maximum occurs for the associated 1D component stochastic process (22, 23). This technique was used for computing the mean perimeter and area of the convex hull of a 2D regular Brownian motion (22, 23) and a 2D random acceleration process (26). Our main idea here is to extend this method to compute the convex hull statistics for the 2D branching Brownian motion. Using this general mapping and isotropy in space (SI Text), the average perimeter and area of the convex hull are given by

.25 =1 R0

.5

=1

=2 .5 R0

R0

0

t

103

40

102

40

t=

101

t=∞

30

0

00

100 100

x

10

20

1

0

t=

10

R0=1

t=1

10

2

0 t=40

103

t=200

< A(t) >

104

0 0

20

40

60

80

100

120

140

x

Fig. 3. (Left) The time behavior of the average area in the supercritical regime for different values of R0 > 1. Dashed lines represent the asymptotic scaling as in Eq. 4. The red curve corresponds to the critical regime. (Right) The behavior of Wt(x) = 1 − Qt(x) for R0 = 1.5 at different times as in Eq. 21. When t → ∞, Wt ðxÞ → 1 − R−1 the traveling 0 , and for large but finite times,p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi front behavior is clearly visible. Inset displays the exponential convergence of Wt(x) to the asymptotic limit. The dashed line represents ξ = D=γðR0 − 1Þ.

Dumonteil et al.

PNAS | March 12, 2013 | vol. 110 | no. 11 | 4241

Fig. 4. (Left) A branching random walk composed of five individuals. At time t = 0, a single infected individual is at the origin O and starts diffusing (blue line). At later times, this individual branches and gives rise to other infected individuals. Among these individuals, the red path reaches the maximum xm along the x component up to the final time t. infected individuals at a given time t are displayed as red dots, whereas recovered individuals are displayed as black dots. (Center) The displacement along the x direction as a function of time. The red path reaches the global maximum xm at time tm. (Right) The displacement along the y direction as a function of time. When the red path reaches the global maximum xm at time tm, its y coordinate attains the value y(tm). A crucial observation is that the y component of the trajectory connecting O to the red path is a regular Brownian motion. This observation does not apply to the x component, which is constrained to reach the global maximum of the branching process.

Convex Hull Perimeter and Maximum xm. ForRthe average perimeter, ∞ we just need the first moment hxm ðtÞi = 0 xm qt ðxm Þdxm , where qt(xm) denotes the probability density of the maximum of the 1D component process. It is convenient to consider the cumulative distribution Qt(xm) (i.e., the probability that the maximum x displacement stays below a given value R ∞ xm up to time t). Then, qt(xm) = dQt(xm)/dxm, and hxm ðtÞi = 0 ½1 − Qt ðxm Þdxm . Because the process starts at the origin, its maximum x displacement, for any time t, is necessarily nonnegative (i.e., xm ≥ 0). We next write down a backward Fokker–Planck equation describing the evolution of Qt(xm) by considering the three mutually exclusive stochastic moves in a small time interval dt: starting at the origin at t = 0, the walker during the subsequent interval [0, dt] dies with probability γdt, infects another individual (i.e., branches) with probability bdt = R0γdt, or diffuses by a random displacement Δx = ηx(0)dt with probability 1 − γ(1 + R0)dt. In the last case, its new starting position is Δx for the subsequent evolution. Hence, for all xm ≥ 0, one can write

Qt+dt ðxm Þ = γdt + R0 γdtQ2t ðxm Þ + ½1 − γðR0 + 1ÞdthQt ðxm − ΔxÞi ; [8] where the expectation 〈〉 is taken with respect to the random displacements Δx. The first term means that if the process dies right at the start, its maximum up to t is clearly zero and hence, necessarily less than xm. The second term denotes the fact that, in case of branching, the maximum of each branch stays below xm: because the branches are independent, one gets a square. The third term corresponds to diffusion. By using 〈Δx〉 = 0 and 〈Δx2〉 = 2Ddt and expanding Eq. 8 to the first order in dt and second order in Δx, we obtain ∂ ∂2 Q = D 2 Q − γðR0 + 1ÞQ + γR0 Q2 + γ ∂xm ∂t

[9]

for xm ≥ 0, satisfying the boundary conditions Qt(0) = 0 and Qt(∞) = 1 and the initial condition Q0(xm) = Θ(xm), where Θ is the Heaviside step function. Hence, from Eq. 5, Z∞ hLðtÞi = 2π ½1 − Qt ðxm Þdxm :

[10]

0

Eq. 9 can be solved numerically for all t and all R0 values, which allows for subsequent computation of 〈L(t)〉 in Eq. 10 (details are provided in SI Text and Figs. S1, S2, and S3). 4242 | www.pnas.org/cgi/doi/10.1073/pnas.1213237110

Convex Hull Area. To compute the average area in Eq. 7, we need

to evaluate hx2m ðtÞi as well as 〈tm〉. After the cumulative distri2 bution Qt(xm) is known, the second moment R ∞hxm ðtÞi can be directly computed by integration, namely hx2m ðtÞi = 0 dxm 2xm ð1 − Qt ðxm ÞÞ. To determine 〈tm〉, we need to also compute the probability density pt(tm) of the random variable tm. Unfortunately, writing down a closed equation for pt(tm) is hardly feasible. Instead, we first define Pt(xm, tm) as the joint probability density that the maximum of the x component achieves the value xm at time tm, when the full process is observed up to time t. Then, we derive a backward evolution equation for Pt(xm, tmR) and integrate out xm ∞ to derive the marginal density pt ðtm Þ = 0 Pt ðxm ; tm Þdxm . Following the same arguments as for Qt(xm) yields a backward equation for Pt(xm, tm), Pt+dt ðxm ; tm Þ = ½1 − γðR0 + 1ÞdthPt ðxm − Δx; tm − dtÞi : + 2γR0 dtQt ðxm ÞPt ðxm ; tm − dtÞ

[11]

The first term at the right-hand side represents the contribution from diffusion. The second term represents the contribution from branching: we require that one of them attains the maximum xm at the time tm − dt, whereas the other stays below xm [Qt(xm) being the probability that this condition is satisfied]. The factor 2 comes from the interchangeability of the particles. Developing Eq. 11 to leading order gives

∂ ∂ ∂2 + Pt = D 2 − γðR0 + 1Þ + 2γR0 Qt Pt : ∂t ∂tm ∂xm

[12]

This equation describes the time evolution of Pt(xm, tm) in the region xm ≥ 0 and 0 ≤ tm ≤ t. It starts from the initial condition P0(xm, tm) = δ(xm)δ(tm) (because the process begins with a single infected with x component located at x = 0, it implies that, at t = 0, the maximum xm = 0 and also tm = 0). For any t > 0 and xm > 0, we have the condition Pt(xm, 0) = 0. We need to also specify the boundary conditions at xm = 0 and xm → ∞, which read (i) Pt(∞, tm) = 0 (because for finite t, the maximum is necessarily finite) and (ii) Pt ð0; tm Þ = δðtm Þ qt ðxm Þjxm =0 . The latter condition comes from the fact that, if xm = 0, the x component of the entire process, starting at 0 initially, stays below zero in the time interval [0, t], which happens with probability qt ðxm Þjxm =0 : consequently, tm must necessarily be zero. Furthermore, by integrating Pt(xm, tm) with respect to tm, we recover the marginal density qt(xm). Dumonteil et al.

Zt Tt ðxm Þ =

tm Pt ðxm ; tm Þdtm ;

[13]

Thus, interestingly, the leading order result is universal [i.e., independent of the details of the cutoff function f(z); the c dependence is only in the subleading term]. To complete the characterization of 〈A(t)〉 in Eq. 7, we still need to determine 〈tm〉: in SI Text, we explicitly determine the stationary solution P∞(xm, tm) for R0 = 1. By following the same arguments as for hx2m ðtÞi, we show that

0

R from which the average follows as htm i = dxm Tt ðxm Þ. Multiplying Eq. 12 by tm and integrating by parts, we get ∂ ∂2 [14] Tt − qt ðxm Þ = D 2 + 2γR0 Qt − γðR0 + 1Þ Tt ∂xm ∂t with the initial condition T0(xm) = 0 and the boundary conditions Tt(0) = 0 and Tt(∞) = 0. Eq. 14 can be integrated numerically together with Eq. 9 (details are provided in SI Text), and the behavior of Z∞ hAðtÞi = π

dxm ½2xm ð1 − Qt ðxm ÞÞ − Tt ðxm Þ

[15]

0

as a function of time is illustrated in Fig. 2. Critical Regime. We now focus on the critical regime R0 = 1. We begin with the average perimeter: for R0 = 1, Eq. 9 admits a stationary solution as t → ∞, which can be obtained by setting ∂Q/∂t = 0 and solving the resulting differential equation. In fact, this stationary solution was already known in the context of the genetic propagation of a mutant allele (27). Taking the derivative of this solution with respect to xm, we get the stationary probability density of the maximum xm: rffiffiffiffiffiffi γ 2 6D q∞ ðxm Þ = ∂xm Q∞ ðxm Þ =  [16] rffiffiffiffiffiffi 3 : γ 1+ xm 6D pffiffiffiffiffiffiffiffiffiffiffi R∞ The average is hxm i = 0 xm q∞ ðxm Þdxm = 6D=γ , which yields Eq. 1 for the average perimeter of the convex hull at late times. To compute the average area in Eq. 7, we need to also evaluate the second moment hx2m ðtÞi, which diverges as t → ∞ because of the power-law tail of the stationary probability density q∞ ðxm Þ ∝ x−3 m for large xm. Hence, we need to consider large but finite t. In this case, the time-dependent probability density qt(xm) displays a scaling form, which can be conveniently written as   xm qt ðxm Þ ’ q∞ ðxm Þf pffiffiffiffiffi ; [17] Dt

where f(z) is a rapidly decaying function with f(z  1) ’ 1 and f(z  1) ’ 0. Using the scaling form of expression 17 and Eq. 9, one can derive a differential equation for f(z). However, it turns out that we do not really need the solution of f(z). From expression 17, we see that the asymptotic power-law decay pffiffiffiffiffi of qt(xm) for large xm has a cutoff around x*m ∼ Dt and that f(z) is the cutoff function. The R ∞second moment at finite but large times t is given by hx2m ðtÞi = 0 x2m   qt ðxm Þdxm . Substituting pffi the scaling form and cutting off the integral over xm at x*m = c t [where the constant c depends on the precise form of f(z)], we get to leading order for large t:  2 xm ðtÞ ’

*

Zxm x2m q∞ ðxm Þdxm ’ 0

Dumonteil et al.

6D ln t: γ

[18]

htm i ’

3 ln t 5γ

[19]

for large t, which leads again to a logarithmic divergence in time. Finally, substituting expressions 18 and 19 in Eq. 7 gives the result announced in Eq. 2. A deeper understanding of the statistical properties of the process would demand knowing the full distribution Prob(L, t) and Prob(A, t) of the perimeter and area. These quantities seem rather hard to compute, but one can obtain the asymptotic tails of the distributions by resorting to scaling arguments. Following the lines of the Cauchy formula (SI Text), it is reasonable to assume that, for each sample, the perimeter scales as L(t) ∼ xm(t). We have seen that the distribution of xm(t) has a power-law tail for large t: q∞ ðxm Þ ∼ x−3 m for large xm. Then, assuming the scaling L(t) ∼ xm(t) and using Prob(L, t → ∞)dL ∼ q∞(xm)dxm, it follows that, at late times, the perimeter distribution also has a power-law tail: Prob(L, t → ∞) ∼ L−3 for large L. Similarly, using the Cauchy formula for the area, we can reasonably assume that, for each sample, AðtÞ ∼ x2m ðtÞ in the scaling regime. Once again, using Prob(A, t → ∞)dA = q∞(xm)dxm, we find that the area distribution also converges, for large t, to a stationary distribution with a power-law tail: Prob(A, t → ∞) ∼ A−2 for large A. Moreover, the logarithmic divergence of the mean area calls for a precise ansatz on the tail of the area distribution, namely   24πD −2 A A h ; [20] ProbðA; tÞ



! A1 5γ Dt where the scaling function h(z) satisfies the conditions h(z  1) = 1 and h(z  1) ’ 0. It is not difficult to verify that expression is the only scaling compatible with Eq. 2. These two results are consistent with the fact that, for each sample, typically A(t) ∼ L2(t) at late times in the scaling regime. Our scaling predictions are in agreement with our Monte Carlo simulations (Fig. 2). The power-law behavior of Prob(A, t) implies that the average area is not representative of the typical behavior of the epidemic area, which is actually dominated by fluctuations and rare events, with likelihood given by expression 20. Supercritical Regime. When R0 > 1, it is convenient to rewrite Eq.

9 in terms of W(xm, t) = 1 − Q(xm, t):

∂ ∂2 W = D 2 W + γðR0 − 1ÞW − γR0 W 2 ∂xm ∂t

[21]

starting from the initial condition W(xm, 0) = 0 for all xm > 0 (Fig. 3). R∞ From Eq. 10, hLðtÞi = 2π 0 W ðxm ; tÞdxm is just the area under the curve W(xm, t) vs. xm up to a factor 2π. As t → ∞, the system approaches a stationary state for all R0 ≥ 1, which can be obtained by setting ∂tW = 0 in Eq. 21. For R0 > 1, the stationary solution W(xm, ∞) approaches the constant 1 − 1/R0 exponentially fast as xm → ∞, namely W ðxmp ; ∞Þ − 1 + R−1 0 ffi→ exp½−xm =ξ, with a characffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi teristic length scale ξ = D=γðR0 − 1Þ. However, for finite but large t, W(xm, t) as a function of xm has a two-step form: it first decreases from 1 to its asymptotic stationary value 1 − 1/R0 over the length scale ξ, and then decreases rather sharply from 1 − 1/R0 to 0. The frontier between the stationary asymptotic value 1 − 1/R0 (stable) and 0 (unstable) moves forward with time at constant velocity, thus PNAS | March 12, 2013 | vol. 110 | no. 11 | 4243

PHYSICS

The numerical integration of the full Eq. 12 would be rather cumbersome. Fortunately, we do not need this calculation. Because we are only interested in 〈tm〉, it is convenient to introduce

creating a traveling front at the right end, which separates the stationary value 1 − 1/R0 to the left of the front and 0 to the right. This front advances with a constant velocity v* that can be estimated using the standard velocity selection principle (28–30). Near the front, where the nonlinear term is negligible, the equation admits a traveling front solution: W(xm, t) ∼ exp[−λ(xm − vt)], with a one parameter family of possible velocities v(λ) = Dλ + γ(R0 −1)/λ, parametrized by λ. This v(λ) has a minimum at pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi dispersion relation p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi λ = λ* = γðR0 − 1Þ=D, where v* = vðλ* Þ = 2 DγðR0 − 1Þ. According to the standard velocity selection principle (28–30), for a sufficiently sharp initial condition, the system will choose this minimum velocity v*. The width of the front remains ∼O(1) at large t. Thus, because of this sharpness of the front, to leading order for large t, one can approximate W(xm, t) ’ (1 − 1/R0)Θ (v*t − xm) near the front. Hence, to leading order for large t, one gets 〈xm(t)〉 ’ (1 − 1/R0)v*t and hx2m i ’ ð1 − 1=R0 Þðv* tÞ2 . The former gives, from Eq. 5, the result announced in Eq. 3. For the mean area in Eq. 7, the term hx2m i ∼ t2 for large t dominates over 〈tm〉 ∼ t (which can be neglected), and we get the result announced in Eq. 3.

invading a thermodynamical fraction of the total population. In addition, we have assumed that the susceptible individuals are homogeneously distributed in space, which is not the case in reality. Nonetheless, it must be noticed that, in practical applications, whenever strong heterogeneities appear, such as mountains, deserts, or oceans, one can split the analysis of the evolving phenomena by conveniently resorting to several distinct convex hulls— one for each separate region. For analogous reasons, the convex hull approach would not be suitable to characterize birth–death processes with long-range displacements, such as for instance branching Lévy flights. The model discussed in this paper based on branching Brownian motion is amenable to exact results. More generally, realistic models could be taken into account by resorting to cumbersome Monte Carlo simulations. The approach proposed in this paper paves the way for assessing the spatial dynamics of the epidemic by more conveniently solving two coupled nonlinear equations under the assumption that the underlying process be rotationally invariant. We conclude with an additional remark. In our computations of the mean perimeter and area, we have averaged over all realizations of the epidemics up to time t, including realizations that are already extinct at time t. It would also be interesting to consider averages only over the ensemble of epidemics that are still active at time t. In this case, we expect different scaling laws for the mean perimeter and the mean area of the convex hull. In particular, in the critical case, we believe that the behavior would be much closer to that of a regular Brownian motion.

Conclusions In this paper, we have developed a general procedure for assessing the time evolution of the convex hull associated with the outbreak of an epidemic. We find it extremely appealing that one can successfully use mathematical formulae (Cauchy) from 2D integral geometry to describe the spatial extent of an epidemic outbreak in relatively realistic situations. Admittedly, there are many assumptions in this epidemic model that are not quite realistic. For instance, we have ignored the fluctuations of the susceptible populations during the early stages of the epidemic: this hypothesis clearly breaks down at later times, when depletion effects begin to appear because of the epidemic

ACKNOWLEDGMENTS. S.N.M. acknowledges support from Agence Nationale de la Recherche (ANR) Grant 2011-BS04-013-01 WALKMAT. S.N.M. and A.R. acknowledge support from the Indo-French Centre for the Promotion of Advanced Research under Project 4604-3.

1. Bailey NTJ (1975) The Mathematical Theory of Infectious Diseases and Its Applications (Griffin, London). 2. McKendrick AG (1925) Applications of mathematics to medical problems. Proc Edinburgh Math Soc 44(1):98–130. 3. Murray JD (1989) Mathematical Biology (Springer, Berlin). 4. Bartlett MS (1960) Stochastic Population Models in Ecology and Epidemiology (Wiley, New York). 5. Andersson H, Britton T (2000) Stochastic Epidemic Models and their Statistical Analysis Lecture Notes in Statistics 151 (Springer, New York). 6. Antal T, Krapivsky PL (2012) Outbreak size distributions in epidemics with multiple stages. arXiv:1204.4214. 7. Anderson R, May R (1991) Infectious Diseases: Dynamics and Control (Oxford Univ Press, Oxford). 8. Antia R, Regoes RR, Koella JC, Bergstrom CT (2003) The role of evolution in the emergence of infectious diseases. Nature 426(6967):658–661. 9. Whittle P (1955) The outcome of a stochastic epidemic—a note on Bailey’s paper. Biometrika 42(1/2):116–122. 10. Kendall DG (1956) Deterministic and stochastic epidemics in closed populations. Proc 3rd Berkeley Symp Math Statist Prob 4:149–165. 11. Bartlett MS (1956) An Introduction to Stochastic Processes (Cambridge Univ Press, Cambridge, United Kingdom). 12. Elliott P, Wakefield JC, Best NG, Briggs DJ (2000) Spatial Epidemiology: Methods and Applications (Oxford Univ Press, Oxford). 13. Radcliffe J (1976) The convergence of a position-dependent branching process used as an approximation to a model describing the spread of an epidemic. J Appl Probab 13(2):338–344. 14. Wang JS (1980) The convergence of a branching brownian motion used as a model describing the spread of an epidemic. J Appl Probab 17(2):301–312. 15. Riley S (2007) Large-scale spatial-transmission models of infectious disease. Science 316(5829):1298–1301. 16. Fraser C, et al. (2009) Pandemic potential of a strain of influenza A (H1N1): Early findings. Science 324(5934):1557–1561.

17. Colizza V, Barrat A, Barthélemy M, Vespignani A (2006) The role of the airline transportation network in the prediction and predictability of global epidemics. Proc Natl Acad Sci USA 103(7):2015–2020. 18. Brockmann D, Hufnagel L, Geisel T (2006) The scaling laws of human travel. Nature 439(7075):462–465. 19. Cowled B, Ward MP, Hamilton S, Garner G (2009) The equine influenza epidemic in Australia: Spatial and temporal descriptive analyses of a large propagating epidemic. Prev Vet Med 92(1–2):60–70. 20. Worton BJ (1995) A convex hull-based estimator of home-range size. Biometrics 51(4): 1206–1215. 21. Giuggioli L, Abramson G, Kenkre VM, Parmenter RR, Yates TL (2006) Theory of home range estimation from displacement measurements of animal populations. J Theor Biol 240(1):126–135. 22. Randon-Furling J, Majumdar SN, Comtet A (2009) Convex hull of N planar Brownian motions: Exact results and an application to ecology. Phys Rev Lett 103(14):140602. 23. Majumdar SN, Comtet A, Randon-Furling J (2010) Random convex hulls and extreme value statistics. J Stat Phys 138(6):955–1009. 24. Cauchy A (1850) Mémoire sur la rectification des courbes et la quadrature des surfaces courbes. Mém Acad Sci Paris 22:3–15. 25. Santaló LA (1976) Integral Geometry and Geometric Probability (Addison–Wesley, Reading, MA). 26. Reymbaut A, Majumdar SN, Rosso A (2011) The convex hull for a random acceleration process in two dimensions. J Phys A Math Theor 44:415001. 27. Sawyer S, Fleischman J (1979) Maximum geographic range of a mutant allele considered as a subtype of a Brownian branching random field. Proc Natl Acad Sci USA 76 (2):872–875. 28. van Saarloos W (2003) Front propagation into unstable states. Phys Rep 386(2–6): 29–222. 29. Brunet E, Derrida B (2009) Statistics at the tip of a branching random walk and the delay of traveling waves. Europhys Lett 87:60010. 30. Majumdar SN, Krapivsky PL (2003) Extreme value statistics and traveling fronts: Various applications. Physica A 318(1–2):161–170.

4244 | www.pnas.org/cgi/doi/10.1073/pnas.1213237110

Dumonteil et al.

Supporting Information Dumonteil et al. 10.1073/pnas.1213237110 Qn+1 ðiÞ = Qn ðiÞ + γdt½1 − Qn ðiÞ2 + D

SI Text 1. Cauchy Formula. The problem of determining the perimeter and

the area of the convex hull of any 2D stochastic process [x(τ), y(τ)] with 0 ≤ τ ≤ t can be mapped to the problem of computing the statistics of the maximum and the time of occurrence of the maximum of the 1D component process x(τ) (1, 2, 3). This goal is achieved by resorting to a formula by Cauchy, which applies to any closed convex curve C. A sketch of the method is illustrated in Fig. S1. Choose the coordinates system such that the origin is inside the curve C, and take a given direction θ. For fixed θ, consider a stick perpendicular to this direction, and imagine bringing the stick from infinity; stop on first touching the curve C. At this point, the distance M(θ) of the stick from the origin is called the support function in the direction θ. Intuitively, the support function measures how close one can get to the curve C in the direction θ, coming from infinity. After the support function M(θ) is known, then Cauchy equations (4, 5) give the perimeter L and the area A enclosed by C, namely Z2π L=

MðθÞdθ 0

and 1 A= 2

Z2πh

 2 i M 2 ðθÞ − M′ðθÞ dθ;

[S1]

0

where M′(θ) = dM/dθ. For example, for a circle of radius R = r, M(θ) = r, and one recovers the standard equations: L = 2πr and A = πr2. When C is the convex hull associated with the process at time t, we first need to compute its associated support function M(θ). A crucial point is to realize that actually M(θ) = max0 ≤ τ ≤ t[x(τ)cos(θ) + y(τ)sin(θ)] (1, 2). Furthermore, if the process is rotationally invariant, any average is independent of the angle θ. Hence, for the average perimeter, we can simply set θ = 0 and write 〈L(t)〉 = 2π〈M(0)〉, where brackets denote the ensemble average over realizations. Similarly, for the average area, 〈A(t)〉 = π[〈M2(0)〉 − 〈M′(0)2〉]. Clearly, M(0) = max0 ≤ τ ≤ t[x(τ)] is then the maximum of the 1D component process x(τ) for τ ∈ [0, t]. Assuming that x(τ) takes its maximum value x(tm) at time τ = tm (Fig. 4), then, M(0) = x(tm) = xm(t), and M′(0) = y(tm). [Actually, tm implicitly depends on θ; hence, formally, M′ðθÞ = − xðtm ÞsinðθÞ + yðtm ÞcosðθÞ + dtdθm dzdtθ ðtÞjt=tm . Nonetheless, because zθ(t) is maximum at t = tm, by definition, dzθ ðtÞ=dtjt=tm = 0.] Now, by taking the average over Cauchy formulas and using isotropy, we simply have Eqs. 5 and 6 from the text for the mean perimeter and the mean area of the convex hull C at time t. Note that this argument is very general and applicable to any rotationally invariant 2D stochastic process. Because the branching Brownian motion with death is rotationally invariant, we can use these formulae. 2. Numerical Methods. Numerical integration. Eqs. 9 and 14 in the text have been integrated numerically by finite differences in the following way. Time has been discretized by setting t = ndt, and space has been discretized by setting x = idx, where dt and dx are small constants. For the sake of simplicity, here, we consider the case R0 = 1. We, thus, have Dumonteil et al. www.pnas.org/cgi/content/short/1213237110

dt ðdxÞ2

½Qn ði + 1Þ − 2Qn ðiÞ

+ Qn ði − 1Þ

[S2]

and Tn+1 ðiÞ = Tn ðiÞ + 2γdt  Tn ðiÞ½Qn ðiÞ − 1 + D + Tn ði − 1Þ +

dt ðdxÞ2

½Tn ði + 1Þ − 2Tn ðiÞ

dt ½Tn ðiÞ − Tn ði − 1Þ: dx

[S2]

As for the initial conditions, Q0(0) = 0, Q0(i > 0) = 1, and T0(i) = 0 ∀i. The boundary conditions at the origin are Qn(0) = 0 and Tn(0) = 0. To implement the boundary condition at infinity, we impose Qn(imax) = 1 and Tn(imax) = 0 ∀n, where the large value imax is chosen so that Tn(imax) − Tn(imax − 1) < 10−7. We have verified that numerical results do not change when passing to the tighter condition Tn(imax) − Tn(imax − 1) < 10−9. After Qn(i) and Tn(i) are known, we use Eqs. 10 and 15 from the text to determine the average perimeter and area, respectively. Monte Carlo simulations. The results of numerical integrations have been confirmed by running extensive Monte Carlo simulations. Branching Brownian motion with death has been simulated by discretizing time with a small dt: in each interval dt, with probability bdt, the walker branches and the current walker coordinates are copied to create a new initial point, which is then stored for being simulated in the next dt; with probability γdt, the walker dies and is removed, and with probability 1 − (b + γ)dt, the walker diffuses: the x and y displacements pffiffiffiffiffiffiffiffiffiffiare sampled from Gaussian densities of zero mean and SD 2Ddt, and the particle position is updated. The positions of all of the random walkers are recorded as a function of time, and the corresponding convex hull is then computed by resorting to the algorithm proposed in ref. 6. Perimeter statistics. To complete the analysis of the convex hull statistics, in Figs. S2 and S3, we show the results for the perimeter. 3. Analysis of tm. In the critical case R0 = 1, the stationary joint probability density P∞(xm, tm) satisfies (on setting ∂Pt/∂t = 0 in Eq. 12 in the text) 3 2

7 6 ∂2 ∂ 2γ 7P∞ ðxm ; tm Þ: D 2 − P∞ ðxm ; tm Þ = 6  r ffiffiffiffiffiffi 5 4 2 ∂tm ∂xm γ xm 1+ 6D

[S4]

For any xm > 0, we have the condition P∞(xm, 0) = 0. The boundary conditions for p Eq. S4 are ffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi P∞(xm → ∞, tm) = 0 and P∞ ð0; tm Þ = q∞ ð0Þδðtm Þ = 2 γ=ð6DÞδðtm Þ. We first take the Laplace transform of Eq. S4, namely ~ ∞ ðxm ; sÞ = P

Z∞

e−stm P∞ ðxm ; tm Þdtm :

[S5]

0

Hence, for all xm > 0, 2

3

6 7 12 61 + 7 D ∂2 ~ sffiffiffiffiffiffi ~ 6 !2 7P ðx P ; sÞ = ∞ m 6 7 ∞ ðxm ; sÞ; 2 s ∂xm s 6D 4 5 + xm D γ

[S6]

1 of 3

where we have used the condition P∞(xm, 0) = 0 for any xm > 0. This second-order differential equation satisfies two pffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi boundary ~ ∞ ð∞; sÞ = 0 and P ~ ∞ ð0; sÞ = 2 γ=ð6DÞ. The latter conditions: P condition pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi is obtained by Laplace transforming P∞ ð0; tm Þ = 2 γ=ð6DÞδðtm Þ. By setting sffiffiffiffiffiffi !rffiffiffiffi 6D s + xm z= ; γ D

[S7]

only in the asymptotic tail of p∞(tm), it suffices to investigate the small s behavior of ~p∞ ðsÞ. Integrating Eq. S10 over xm and taking the s → 0 limit, we obtain, after some straightforward algebra, ~ p∞ ðsÞ = 1 +

we rewrite the equation as

  d 1d 49 FðzÞ − 1 + 2 FðzÞ = 0: FðzÞ + z dz 4z dz2 2

[S9]

The general solution of this differential equation can be expressed as a linear combination of two independent solutions: F(z) = AI7/2(z) + BK7/2(z), where Iν(z) and Kν(z) are modified Bessel functions. Because Iν(z) ∼ ez for large z, it is clear that, to satisfy ~ ∞ ð∞; sÞ = 0 [which means F(z → ∞) = the boundary condition P 0], we need to choose A = 0. Hence, we are left with F(z) = BK7/2(z), where the constant B is from the second pdetermined ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ~ ∞ ð0; sÞ = 2 γ=ð6DÞ. By reverting to the boundary condition P variable xm, we finally get rffiffiffiffiffiffirffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi K7=2 γ γ ~ P∞ ðxm ; sÞ = 2 1+ xm 6D 6D

rffiffiffiffiffiffi rffiffiffiffi  6D s + xm γ D : [S10] rffiffiffiffiffi 6s K7=2 γ

Now, we are interested in determining the Laplace transform of R∞ marginal density ~p∞ ðsÞ = 0 e−stm p∞ ðtm Þdtm , where p∞ ðtm Þ = Rthe ∞ R ∞ transform of this last relation 0 P∞ ðxm ; tm Þdxm . Taking Laplace ~ ∞ ðxm ; sÞdxm . After we know with respect to tm gives ~p∞ ðsÞ = 0 P ~p∞ ðsÞ, we can invert it to obtain p∞(tm). Because we are interested 1. Randon-Furling J, Majumdar SN, Comtet A (2009) Convex hull of N planar Brownian motions: Exact results and an application to ecology. Phys Rev Lett 103(14): 140602. 2. Majumdar SN, Comtet A, Randon-Furling J (2010) Random convex hulls and extreme value statistics. J Stat Phys 138(6):955–1009. 3. Cauchy A (1850) Mémoire sur la rectification des courbes et la quadrature des surfaces courbes. Mém Acad Sci Paris 22:3–15.

Dumonteil et al. www.pnas.org/cgi/content/short/1213237110

[S11]

We further note that Z∞

∂2 ~ ~ ∞ = 0: ~ ∞ − 12 P P∞ − P [S8] ∂z2 z2 ~ ∞ ðzÞ = pffiffizFðzÞ, the function F(z) On making the transformation P then satisfies the Bessel differential equation:

3 s ln ðsÞ + ⋯: 5γ

e−stm t2m p∞ ðtm Þdtm =

0

d2 3 ~ p ðsÞ ’ ; ds2 ∞ 5γs

[S12]

which can then be inverted to give the following asymptotic behavior for large tm: p∞ ðtm Þ ’

3 : 5γt2m

[S13]

Analogously as for hx2m i, the moment 〈tm〉 → ∞ because of the power-law tail p∞ ðtm Þ∝t−2 m . Hence, we need to compute 〈tm〉 for large but finite t: in this case, the time-dependent solution displays a scaling behavior: t

m pt ðtm Þ ’ p∞ ðtm Þg ; [S14] t where the scaling function g(z) satisfies the conditions g(z  1) ’ 1 and g(z  1) = 0. Much like in expression 17 in the text for the marginal density qt(xm), we have a power-law tail of pt(tm) for large tm that has a cutoff at a scale t*m ∼ t, and g(z) is the cutoff function. As in the case of xm, we do not need the precise form of g(z) R ∞ to compute the leading term of the first moment htm i = * 0 pt ðtm Þtm dtm for large t. Cutting off the integral at tm = c1 t [where c1 depends on the precise form of g(z)] and performing the integration gives Zt tm p∞ ðtm Þdtm ’

htm i ’

3 ln t; 5γ

[S15]

0

which is precisely the result announced in expression 19 in the text. 4. Santaló LA (1976) Integral Geometry and Geometric Probability (Addison–Wesley, Reading, MA). 5. Reymbaut A, Majumdar SN, Rosso A (2011) The convex hull for a random acceleration process in two dimensions. J Phys A Math Theor 44:415001. 6. Gehrman DC (2005) Module: Finding the convex hull of a set of 2D points. Python Cookbook, eds Martelli A, Ravenscroft A, Ascher D (O’Reilly, Paris).

2 of 3

Cauchy construction of the 2D convex hull, with support function M(θ) representing the distance along the direction θ.

R0= 1.1 5

Fig. S1.

100

.01

1 R= 0

10-2

R0=1

R0=0.99

60

Prob(L,t)

< L(t) >

80 R0=0.85

40 20 0

101

102

103 t

104

105

10

-3

10

-4

10

-5

10-6

101

102 L

103

R0 =2

.5

Fig. S2. (Left) The average perimeter〈L(t)〉of the convex hull as a function of the observation time. For the parameter values, we have chosen D = 1/2 and b = R0γ = 0.01. We considered five different values of R0. We have obtained these results by two different methods. (i) One method is by the numerical integration of Eq. S9 and using Eq. 10 in the text (with the choices dt = 0.003125 and dx = 0.1768). These results are displayed as solid lines. (ii) Another method is by Monte Carlo simulations of the 2D branching Brownian motion with death with the same parameters and the choice of the Monte Carlo time step dt = 0.25 with the results averaged over 105 samples. Monte Carlo simulations are displayed as symbols. The dashed lines represent the asymptotic limits as given in Eq. 1 in the text for the critical case R0 = 1. (Right) Distribution of the perimeter of the convex hull for the critical case R0 = 1, with γ = 0.01 and D = 1/2, as obtained by Monte Carlo simulations with time step dt = 1 and t = 4 × 105. The number of realizations is 2 × 106. The dashed line in Right corresponds to the power-law L−3 (up to an arbitrary prefactor).

1.5 0= R

R

0

.25

=1

R0=1

Fig. S3. The time behavior of the average perimeter in the supercritical regime for different values of R0 > 1. Dashed lines represent the asymptotic scaling as in Eq. 3 in the text. The red curve corresponds to the critical regime.

Dumonteil et al. www.pnas.org/cgi/content/short/1213237110

3 of 3

ANNEXE B. PUBLICATIONS PRINCIPALES

B.2

160

liées au Chapitre 4

E. Dumonteil

Université Paris-Sud

ON A NEW VARIANCE REDUCTION TECHNIQUE: NEURAL NETWORK BIASING—A STUDY OF TWO TEST CASES WITH THE MONTE CARLO CODE TRIPOLI4

RADIATION PROTECTION KEYWORDS: Monte Carlo, variance reduction, neural network

ERIC DUMONTEIL* Commissariat à l’Energie Atomique Service d’Etude des Réacteurs et de Mathématiques Appliquées (DEN0DANS0DM2S0SERMA0LTSD), F91191 Gif-sur-Yvette CEDEX, France Received April 16, 2008 Accepted for Publication December 3, 2008

probability density function ~pdf ! that a given part of the phase-space contributes to the detector: Various variance-reduction techniques are used in Monte Carlo particle transport. Most of them rely either on a hypothesis made by the user (parameters of the exponential biasing, mesh and weight bounds for weight windows, etc.) or on a previous calculation of the system with, for example, a deterministic solver. This paper deals with a new acceleration technique, namely, autoadaptative neural network biasing. Indeed, instead of using any a priori knowledge of the system, it is possible, at a given point in a simulation, to use the Monte Carlo histories previously simulated to train a neural network, which, in return, should be able to provide an estimation of the adjoint flux, used then for biasing the simulation. We will describe this method, detail its implementation in the Monte Carlo code Tripoli4, and discuss its results on two test cases.

pdf : ~ r,? V, < E ! 僆 ᑬ 7 r I 僆 @0, 1# . This function is also known as the adjoint flux C ⫹ or, in the frame of the zero variance game, as the ideal importance function I.

II. TRIPOLI4 CURRENT BIASING SCHEME To precisely detail the implementation of the neural network in Tripoli4 ~Ref. 1!, it is worthwhile to first introduce its standard variance-reduction techniques. Indeed, beyond the well-known “weight watching” functionalities ~like splitting and Russian roulette!, Tripoli4 owns its full biasing scheme, based on conformal symmetry preserving. Let us introduce the conformal operator: C~ r,? v! ? r C * ~ r,? v! ? ⫽ I ~ r,? v!C~ ? r,? v! ? .

~1!

If we consider the Boltzmann transport equation, ? ⫹ S T ~ r,? v!C~ ? r,? v! ? V{ < ¹< C~ r,? v!

I. INTRODUCTION Variance-reduction techniques in the field of Monte Carlo shielding calculations often rely on a deterministic solving of the Boltzmann transport equation, aiming at providing an estimation of the adjoint flux. This adjoint flux is then used to bias the simulation. Another attempt to guess the form of the adjoint flux can be done with feeding a neural network with all the collisions @points P~ x,? V, < E ! or P~ x,? v? !# . If a given point belongs to a history that reaches the target detector, then the neural network target value is 1; otherwise, it is 0. Thus, the neural network will approximate the

VOL. 168



S T ~ r,? v!C~ ? r,? v? ' !C~ v? ' r v,? r! ? d 3 v? ' ⫹ Q~ r,? v! ? , ~2!

where Q~ r,? v! ? ⫽ source ? ⫽ macroscopic total cross section S T ~ r,? v! ? ⫽ collision operator, C~ v? ' r v,? r! then one can rewrite it using Eq. ~1!, and after development one obtains

*E-mail: [email protected] NUCLEAR TECHNOLOGY



DEC. 2009

793

Dumonteil

VARIANCE-REDUCTION TECHNIQUE USING NEURAL NETWORK BIASING TEST CASES WITH TRIPOLI4



V{ < ¹< C * ~ r,? v! ? ⫹ S T ~ r,? v! ? ⫺ V{ < ⫽



¹< I ~ r,? v! ?

I ~ r,? v! ?



C * ~ r,? v! ?

I ~ r,? v! ?

* S T ~ r,? v!C ? ~ r,? v? ' !

I ~ r,? v? ' !

? d 3 v? ' ⫹ Q~ r,? v!I ? ~ r,? v! ? . ⫻ C~ v? ' r v,? r!

~3!

calculation of the K coefficients. It is worth noting that all these options suppose to solve Eqs. ~4! and ~5! on a spatial, angular, and energy grid @i.e., I ~ r,? v! ? is calculated on a grid#. This paper proposes to compute I ~ r,? v! ? with a neural network, for each point of the phase-space @i.e., I ~ r,? v! ? is a field#. Section III describes how this neural network is implemented.

If we define K~ r,? v! ? ⫽

冨 I ~ r,? v!? 冨 ¹< I ~ r,? v! ?

~4!

and V< 0 ~ r,? v! ? ⫽

¹< I ~ r,? v! ?

K~ r,? v!I ? ~ r,? v! ?

,

~5!

then we can introduce an effective or biased cross section that acts on the left-side term of Eq. ~3!: S T* ⫽ S T ~ r,? v! ? ⫺ V{ 1, the population will grow unbounded (super-critical condition); finally, if m1 = 1, the population will stay constant on average (exactly critical condition). In the first and second case, the decrease or increase of population number, respectively, will happen exponentially fast. Here, having in mind the application to criticality calculations, we will focus on the critical case, where population is globally conserved. Moreover, in neutronics N is typically very large, namely, N  1. Branching Brownian motion has been long studied in connection with the mathematical modeling of biological populations, such as bacteria, plankton or amoebae (sometimes under the name of Brownian bugs) (Athreya and Ney, 1972; Young et al., 2001; Houchmandzadeh, 2008). In such systems, several experiences have shown the emergence of an astonishing tendency for neutral clustering: even in the absence of particle–particle interactions, individuals that are initially uniformly distributed spontaneously aggregate into spatial clusters (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). This behavior is illustrated in Fig. 4 for a three-dimensional system. At time t = 0, a collection of N branching Brownian particles is uniformly distributed over space. The system is then left free to evolve, starting from the initial positions. At later times, the particle positions begin to fluctuate because of diffusion and branching events. These fluctuations grow in time and finally induce spatial clustering: while the average number of particles is constant on

E. Dumonteil et al. / Annals of Nuclear Energy 63 (2014) 612–618

615

Fig. 4. The evolution of a collection of N = 500 Brownian branching particles whose initial positions are initially homogenously distributed in a three-dimensional domain. As time grows, the diffusive mixing is not sufficient to overcome the spatial clustering effect induced by the death–birth mechanism, and particles at later times occupy only a small number of spatial sites (while the average population stays constant on average).

average, only a few spatial sites are actually occupied at the final observation time. A simple, yet effective, explanation for this kind of behavior has been actually proposed: clustering is the result of particles dying uniformly in space, while births only happen close to a parent particle. Clusters will appear whenever diffusive mixing effects are too weak to overcome this asymmetry: this intuitively happens when the initial particle density is low (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). These features, together with the formal analogy between branching Brownian motion and neutron transport in multiplying media, provide the main motivation for introducing this model. It is important to notice that Brownian motion, while surely being a fairly adequate model in the context of biological systems, is perhaps oversimplified so as to realistically represent neutron trajectories. However, Brownian motion is expected to provide a good description of neutron transport at least when the one-speed diffusion approximation applies. 3.2. The pair correlation function In the context of branching Brownian motion, it has been shown that spatial clustering can be quantified by computing the ensemble-averaged particle concentration ct(x) and the ensemble-averaged 2-point particle correlation function gt(x, y) (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). The particle concentration is defined as

ct ðxÞ ¼

hM t ðxÞi ; dx

ð1Þ

where Mt(x) is the random number of particles that are in a small element volume dx around x at time t, and brackets denote

ensemble averages. On the basis of the previous considerations, concentration alone is not sufficient to fully characterize the evolution of the particle population: in order to assess the effects of space and time fluctuations, it is customary to resort to the (normalized and centered) pair correlation function, which is defined as

g t ðx; yÞ ¼

hM t ðxÞM t ðyÞi  hM t ðxÞihMt ðyÞi  hM t ðxÞidðx  yÞ : hMt ðxÞihM t ðyÞi

ð2Þ

The numerator in Eq. (2) represents the space correlation, to which we subtract the square of the mean (for centering) and the auto-correlation (the delta-function term). The denominator is used for normalization. In (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009) it has been shown that Eq. (2) can be given a probabilistic interpretation, namely, it amounts to the probability density of finding a particle pair at positions x and y at time t, up to a normalization factor (for more details, we refer the reader to the above-mentioned references). An elegant and far-reaching formalism based on a master equation for the balance of the particle number in the phase space allows deriving the equations for the evolution of both ct(x) and gt(x, y) (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). For an unbounded domain in dimension d, the equations for critical conditions m1 = 1 have been shown to read

@ ct ðxÞ ¼ Dr2x ct ðxÞ; @t

ð3Þ

for the concentration, and

@ 2km2 dðx  yÞ; g ðx; yÞ ¼ Dr2x g t ðx; yÞ þ Dr2y g t ðx; yÞ þ ct ðxÞ @t t

ð4Þ

616

E. Dumonteil et al. / Annals of Nuclear Energy 63 (2014) 612–618

for the pair correlation function, respectively, r2z being the ddimensional Laplacian (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). Eqs. (3) and (4) can be in principle solved together with the initial and boundary conditions. For one- and two-dimensional critical systems, it has been shown that the amplitude of the pair correlation function gt(x, y) grows unbounded with time: this means that the clustering mechanism is so strong that the population, starting from a uniform distribution, asymptotically gathers around a countable number of sites having an infinitesimal support. Strictly speaking, this is the phenomenon that mathematicians call clustering (Cox and Griffeath, 1985). For systems of dimension greater than 2, the population will not display these features; yet, depending on the physical parameters, spatial fluctuations can still override the average concentration (Cox and Griffeath, 1985; Houchmandzadeh, 2002): when this happens, strong heterogeneities appear in the particle distribution. With an abuse of language, we will refer also to this case as clustering. For an exactly critical (m1 = 1) three-dimensional system starting from a uniform initial c0(x) = c0, the ensemble-average concentration will stay constant at each subsequent time, namely, ct(x) = c0, "t. As for the ensemble-average pair correlation function, one gets

! km2 1 jx  yj2 ; g t ðx; yÞ ¼ C ; 2 8Dt c0 p3=2 jx  yj

ð5Þ

P where m2 ¼ k kðk  1ÞpðkÞ is the second factorial moment of the number of particles per collision, and C(a, z) is the incomplete Gamma function (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). The space and time behavior of gt(x, y) is illustrated in Fig. 5 (left). The features of spatial clustering are embedded in the shape of the function gt(x, y), whose amplitude grows as a function of time while developing a strong peak at x ’ y. Physically, this means that particles have an increasing probability of occupying spatial sites situated at short distance from each other. As mentioned above, the amplitude of gt(x, y) for three-dimensional systems does not grow unbounded but saturates to a constant at later times. Due to the aforementioned asymmetry between birth and death mechanisms, in each stochastic realization particles cluster together because of fluctuations, and the initial homogeneous distribution progressively breaks down since the size of fluctuations grows in time. We expect spatial clusters to appear when local fluctuations have the same magnitude as local concentrations:

from the definition of the pair correlation function, this happens when gt(x, y) ’ 1 (Young et al., 2001; Houchmandzadeh, 2008; Houchmandzadeh, 2002; Houchmandzadeh, 2009). Finally, observe that the function gt(x, y) is inversely proportional to the initial particle concentration c0, which means that increasing the particle number c0 per unit volume acts against the appearance of spatial clustering: this is indeed coherent with the Monte Carlo simulation results presented in Section 2. 4. A diagnostic tool for spatial clustering in criticality simulations The previous analysis has shown that invaluable information on the emergence of spatial clustering phenomena can be extracted from the 2-points correlation function gt(x, y), which rules the space–time evolution of the particle fluctuations. Motivated by these considerations, we argue that a reliable diagnostic tool for detecting spatial clustering in Monte Carlo criticality simulations could be also based on a similar mathematical object. In principle, the phase space distribution for neutrons in realistic configurations is fairly more complex than that of Brownian particles in simple geometries: for instance, one should also consider energy as well as direction variables. To simplify the matter, we consider only the space variable r as a function of the cycle number n (which replaces time for the Brownian model). Even under these assumptions, an explicit analytic calculation of gn(x, y) would be hardly feasible. Rather, we propose to replace the exact 2-points correlation function with an empirical quantity that can be computed during the Monte Carlo simulation run. To be more precise, we introduce an empirical correlation function as

Gn ðrÞ ¼

  þ r; nÞ  /ðnÞÞi hð/ðx; nÞ  /ðnÞÞð/ðx x ; 2 r ðnÞ

ð6Þ

where /(x, n) stands for the neutron flux at spatial mesh position x  at cycle n; /ðnÞ ¼ h/ðx; nÞix , and brackets denote average over mesh positions. The quantity appearing at the denominator of Eq. (6) is the variance of the neutron flux with respect to the spatial coordinate. Observe that, as opposed to the definition of the 2-points correlation function given in Eq. (2), we have substituted the content of each spatial bin with the value of the neutron flux at the same bin, and the ensemble-average over the possible realizations with a spatial average over all bins. This latter hypothesis holds true provided that some sort of ergodicity is ensured, i.e., the two kinds of averages can be replaced with each other. Alternatively, the empirical pair correlation function could be obtained by computing the

Fig. 5. Left. The pair correlation function gt(x, y) dispalyed as a function of the spatial coordinate r = jx  yj for an unbounded three-dimensional system. The function displays a peak at r = 0, whose amplitude grows with time and asymptotically saturates to a constant value. Right. The empirical correlation function Gn(r) for the PWR pin-cell simulation with L = 400 cm, displayed as a function of the mesh distance i and the cycle number n. The behavior of the empirical correlation function closely resembles that of the pair correlation function of branching Brownian motion.

E. Dumonteil et al. / Annals of Nuclear Energy 63 (2014) 612–618

histogram of the particle distances at each cycle; here, however, we stick to the former definition. Also, the normalization factor appearing in Eq. (6) has been taken to be the variance of the flux, whereas in Eq. (2) the normalization factor is the product of the averages. Eq. (6) has been implemented in the development version of the Monte Carlo code TRIPOLI-4. In order to apprehend the use of the empirical correlation function as a diagnostic tool, we have rerun the criticality simulations considered in Section 2 and computed Eq. (6) as a function of the spatial coordinate r and the cycle number n. Results are shown in Fig. 5 (right). The empirical correlation function displays a peak at short distances (the signature of clustering effects), whose amplitude grows with the number of cycles. We are not able to clearly state whether the amplitude of the correlation function asymptotically saturates or continually grows at later cycles. As observed above, the 2-points correlation function is inversely proportional to the initial particle density. Then, clustering effects are roughly speaking more often observed in systems with a small number of particles per unit volume. Analogously, we expect criticality simulations to be typically affected by spatial clustering effects when the number of simulated particles is relatively ‘small’ as compared to the typical system size. Since the typical separation between particles is approximately given by the mean free path qffiffiffiffiffiffiffiffiffi ‘¼ r 2a (where ra stands for the distance to absorption (Bell and Glasstone, 1970; Duderstadt and Hamilton, 1976)), the number N0 of simulated particles to suppress clustering effects for a system of typical size L should be roughly N0  (L/‘)3. For the examples considered in Section 2, the order of magnitude of the mean free path is ‘ ’ 6 cm, so that for L = 10 cm we have N0  4: for the chosen number of simulated particles (N = 104 per cycle), then N  N0 and clustering effects are therefore not expected to affect the simulation. For L = 100 cm we have N0  5  103, N ’ N0 and clustering effects should begin to manifest themselves. Finally, for L = 400 cm we have N0  3  105, N  N0 and clustering should seriously affect the particle transport. This hand-waving analysis is coherent with our Monte Carlo simulation results. Eq. (2) is a measure of local spatial fluctuations, i.e., fluctuations occurring at a given site r: in a Monte Carlo simulation, some sort of spatial binning is needed, so that what we actually measure is the fluctuations pertaining to a bin. The larger the bin size, the smoother will the fluctuations appear, due to spatial averaging effects. 5. Conclusions and perspectives In this paper, we have examined the issue of space and cycle fluctuations arising in Monte Carlo criticality simulations. In particular, the analysis of a simple PWR pin-cell configuration with reflective boundary conditions has revealed that such space-cycle fluctuations manifest themselves in the form of particle clusters: along cycles, the simulated neutrons gather together instead of converging to the expected flat equilibrium distribution. A physical explanation for this phenomenon has been provided by resorting to a simplified Brownian transport model coupled with a birth–death Galton–Watson process. This model, which mimics the neutron transport in multiplying media, allows extracting the key features responsible for the observed spatial clustering. In particular, it has been possible to conclude that clustering is intimately connected to the fission process (branching), since particle fluctuations are greatly emphasized by the asymmetry between particles being absorbed uniformly in the system and appearing only at fission sites. Moreover, these mechanisms have been shown to be inversely proportional to the number of particles per unit

617

volume, so that Monte Carlo simulations with a relatively small number of transported neutrons are expected to be more prone to spatial clustering. In this respect, ad hoc variance reduction techniques specifically designed for fission chains (Sjenitzer and Hoogenboom, 2011) could perhaps contribute to effectively reduce clustering effects. In order to quantify space-cycle fluctuations, we have proposed to adopt an empirical correlation function, whose behavior helps in detecting the appearance of strong heterogeneities in the particle distribution. The analysis presented in this work is very preliminary, in that many relevant aspects have been neglected for the sake of clarity. Indeed, we have entirely ignored the energetic and angular dependence of the particle trajectories, as well as the effects of delayed neutrons. While all these ingredients surely play a non negligible role in determining the behavior of the particle correlations, we nonetheless believe that retaining the space and cycle coordinates allows grasping the key features behind particle clustering without being hindered by a cumbersome notation. Future work will be aimed at extending the validity of the present results to a more realistic phase-space formalism. A final observation concerns the applicability of the proposed estimator to other systems than the reactor pin-cell. Indeed, for the configuration considered here the expected equilibrium distribution is flat, which highly simplifies the subsequent analysis. For configurations where the equilibrium distribution is not uniform in space, the analysis of the empirical correlation function should accommodate for the emergence of multiple scales due to spatial heterogeneities.

Acknowledgments The authors thank AREVA and Electricité de France (EDF) for partial financial support. References Athreya, K.B., Ney, P., 1972. Branching processes. Grundlehren Series. SpringerVerlag, New York. Bell, G.I., Glasstone, S., 1970. Nuclear Reactor Theory. Van Nostrand Reinhold Company, New York. Brown, F., 2005. LA-UR-05-4983, Los Alamos National Laboratory. Brown, F., 2009. In: Proceedings of M& C2009, Saratoga Springs, NY, USA. Cox, J.T., Griffeath, D., 1985. Annals Prob. 13, 1108. Dawson, D.A., 1972. Z. Wahrsch. Verw. Gebiete 40, 125. Duderstadt, J.J., Hamilton, L.J., 1976. Nuclear Reactor Analysis. Wiley, New York. Dumonteil, E., Malvagi, F., 2012. In: Proceedings of ICAP2012, Chicago. Dumonteil, E., Courau, T., 2010. Nucl. Technol. 172, 120. Dumonteil, E., Le Peillet, A., Lee, Y.K., Petit, O., Jouanne, C., Mazzolo, A., 2006. In: Proceedings of PHYSOR2006, Vancouver. Watson, H.W., Galton, F., 1875. J. Anthropol. Inst. Great Britain 4, 138. Gelbard, E.M., Prael, R., 1990. Prog. Nucl. Energy 24, 237. Houchmandzadeh, B., 2002. Phys. Rev. E 66, 052902. Houchmandzadeh, B., 2008. Phys. Rev. Lett. 101, 078103. Houchmandzadeh, B., 2009. Phys. Rev. E 80, 051920. Jacquet, O., Chajari, R., Bay, X., Nourri, A., Carraro, L., 2000. In: Proceedings of ICRS2000, Lisbon. L’Abbate, A., Courau, T., Dumonteil, E., 2007. In: Proceedings of PHYTRA1, Marrakech. Lux, I., Koblinger, L., 1991. Monte Carlo Particle Transport Methods: Neutron and Photon Calculations. CRC Press, Boca Raton. Osborn, R.K., Yip, S., 1966. The Foundations of Neutron Transport Theory. Gordon and Breach, New York. Pázsit, I., Pál, L., 2008. Neutron Fluctuations: A Treatise on the Physics of Branching Processes. Elsevier, Oxford. Rief, H., Kschwendt, H., 1967. Nucl. Sci. Eng. 30, 395–418. Sjenitzer, B.L., Hoogenboom, J.E., 2011. Ann. Nucl. Eng. 38, 2195. TRIPOLI-4 Project Team, 2008. TRIPOLI-4 User Guide, Rapport CEA-R-6169. Brun, E., Dumonteil, E., Hugot, F.X., Huot, N., Jouanne, C., Lee, Y.K., Malvagi, F., Mazzolo, A., Petit, O., Trama, J.C., Zoia, A., 2011. Overview of TRIPOLI-4 version 7 continuous-energy Monte Carlo transport code. In: Proceedings of ICAPP2011, Nice. Ueki, T., Brown, F., 2003. In: Proceedings of M& C2003, Gatlinburg, Tennessee.

618

E. Dumonteil et al. / Annals of Nuclear Energy 63 (2014) 612–618

Ueki, T., Brown, F., Parsons, D.K., Kornreich, D.E., 2003. Nucl. Sci. Eng. 145, 279. Ueki, T., Brown, F., Parsons, D.K., Warsa, J.S., 2004. Nucl. Sci. Eng. 148, 374. Ueki, T., 2012. J. Nucl. Sci. Technol. 49, 1134–1143. Williams, M.M.R., 1974. Random Processes in Nuclear Reactors. Pergamon Press, Oxford.

Young, W.R., Roberts, A.J., Stuhne, G., 2001. Nature 412, 328. Zoia, A., Dumonteil, E., Mazzolo, A., Mohamed, S., 2012. J. Phys. A: Math. Theor. 45, 425002. Zoia, A., Dumonteil, E., Mazzolo, A., 2012. Europhys. Lett. 100, 40002.

NUCLEAR SCIENCE AND ENGINEERING: 167, 165–170 ~2011!

Biases and Statistical Errors in Monte Carlo Burnup Calculations: An Unbiased Stochastic Scheme to Solve Boltzmann/Bateman Coupled Equations E. Dumonteil* and C. M. Diop Commissariat à l’Energie Atomique, Centre de Saclay Service d’Etude des Réacteurs et de Mathématiques Appliquées (DEN0DANS0DM2S0SERMA0LTSD) F91191 Gif-sur-Yvette CEDEX, France Received December 14, 2009 Accepted July 1, 2010

Abstract – External linking scripts between Monte Carlo transport codes and burnup codes, and complete integration of burnup capability into Monte Carlo transport codes, have been or are currently being developed. Monte Carlo–linked burnup methodologies may serve as an excellent benchmark for new deterministic burnup codes used for advanced systems; however, there are some instances where deterministic methodologies break down (i.e., heavily angularly biased systems containing exotic materials without proper group structure) and Monte Carlo burnup may serve as an actual design tool. Therefore, researchers are also developing these capabilities in order to examine complex, three-dimensional exotic material systems that do not contain benchmark data. Providing a reference scheme implies being able to associate statistical errors to any neutronic value of interest like keff , reaction rates, fluxes, etc. Usually in Monte Carlo, standard deviations are associated with a particular value by performing different independent and identical simulations (also referred to as “cycles,” “batches,” or “replicas”), but this is only valid if the calculation itself is not biased. And, as will be shown in this paper, there is a bias in the methodology that consists of coupling transport and depletion codes because Bateman equations are not linear functions of the fluxes or of the reaction rates (those quantities being always measured with an uncertainty). Therefore, we have to quantify and correct this bias. This will be achieved by deriving an unbiased minimum variance estimator of a matrix exponential function of a normal mean. The result is then used to propose a reference scheme to solve Boltzmann/ Bateman coupled equations, thanks to Monte Carlo transport codes. Numerical tests will be performed with an ad hoc Monte Carlo code on a very simple depletion case and will be compared to the theoretical results obtained with the reference scheme. Finally, the statistical error propagation during the timedependent depletion process will be discussed. It will be proven that the nonlinearity of Bateman equations induces a bias in the coupling. The paper provides a formula to correct it, which is verified in a simple example.

I. INTRODUCTION

dependent calculations. Indeed, these codes @e.g., Tripoli4 ~Refs. 4 and 5!, MONTEBURN ~Ref. 6!, ALEPH ~Ref. 7!, MCNPX ~Ref. 8!, MCWO ~Ref. 9!, MCNP0 REBUS ~Ref. 10!, MCB ~Ref. 11!, RACER ~Ref. 12!, and MCOR ~Ref. 13!# calculate time-dependent material burnup properties. If these capabilities may involve either external coupling or integrated codes, they all rely on a coupling between the Monte Carlo code, which computes the flux ~or reaction rates! in a given volume, and a depletion code, which computes concentrations for different nuclides in this volume, for a given burning time

From a mathematical point of view, the aim of this paper is to provide a reference numerical scheme aiming at solving the coupled Boltzmann0Bateman equations. This scheme is based on Refs. 1, 2, and 3. The physics motivation behind it consists of improving the methodology used by a particular ~and quite recent! kind of Monte Carlo–linked burnup capability performing time*E-mail: [email protected] 165

166

DUMONTEIL and DIOP

and a given flux. The common scheme consists of computing a mean flux with the Monte Carlo transport code and using this mean flux to compute mean concentrations with the depletion code. This last concentration “vector” is used to reload the transport simulation; this whole process is performed iteratively. If this calculation scheme seems at first sight safe, it presents nevertheless a consequent drawback because, as will be seen in the following, there is a bias function of the number of particles in each transport. It is thus impossible not to parallelize it ~here, we mean a full parallelization of the transport0depletion sequences—not only the transport!. For the same reason it is not possible to estimate its statistical error by using a brute force method ~that is, in the sense of performing independent “replicas” of the burnup calculation; see Refs. 14 and 15 for more details on independent replicas!. Indeed, let us consider a sample $fi % of I flux estimation ~the same considerations held with reaction rates!. These estimations come from the different batches ~cycles! of a Monte Carlo simulation ~that solve the Boltzmann equation! and are the realizations of independent and identically distributed ~iid! random variables f of a normal law N~fR , sf ! ~where the R stands for “real”!. The estimator of the flux that is usually built from these estimations is the mean estimator: fN ⫽

( fi I

.

Now the concentration N of a given nuclide is calculated from the Bateman equation. This equation is a function g of the flux ~for one burning step!, and if we consider the real concentration NR , we have

processes is used to perform natural parallelization of such calculations: Each node of a cluster computes a value of N, and the distribution of N over all nodes is the true statistical distribution. This parallelization cannot be performed since our estimator of N is biased. In Sec. II, we will recall recent results in the field of nonlinear function estimation, using unbiased minimum variance estimator ~UMVE! of nonlinear functions. We will show how they can be applied to solve the coupled Boltzmann0Bateman equation. II. UNBIASED MINIMUM VARIANCE ESTIMATOR OF A NONLINEAR FUNCTION OF A NORMAL MEAN

Following the work initiated by Kolmogorov in 1950 ~see Ref. 16! concerning the theory of estimation, recent results allow estimation of a nonlinear function of a normal mean ~those functions are defined on iid random variables distributed according to a normal law!, with the help of unbiased minimum variance estimators. In Ref. 3 two estimators are considered: One of them, from Gray, Watkins, and Schucany 2 involves infinite series, and the other, from Stefanski, Novick, and Devanarayan, 1 uses complex numbers. Both estimators are minimum variance, which is important for iterative processes like Monte Carlo burnup calculations, as will be seen in Sec. IV. The Gray, Watkins, and Schucany estimator,2 from now on denoted gI 1 ~fR !, can be written as follows: ~⫺1! k G `

N ⫹ gI 1 ~fR ! ⫽ g~ f!

N NR ⫽ g~fR ! ⫽ lim ~g~ f!! Ir`

~because lim Ir` ~ fN ⫽ ( fi 0I ! ⫽ fR !. Nevertheless, since this function is nonlinear, in the general case the function of the mean flux g~ f! N is not an unbiased estimator of the real concentration NR : N NR ⫽ g~fR ! ⫽ E~g~ f!! since the expected value of g~ f!, N called E~g~ f!!, N is the mean of g over all possible realizations $ fN j %: E~g~ f!! N ⫽ lim

Jr`



( g~fj! J





冉 冊 s[ 2 4

k!G

I ⫺1 2

I ⫺1 2

g ~2k! ~ f! N

⫹k

k

,

~1!

where s[ is the estimated square root of the variance of the sample $fi %~i ⫽ 1. . . I !. The Stefanski, Novick, and Devanarayan estimator, 1 denoted gI 2 ~fR !, is

再冉

gI 2 ~fR ! ⫽ Re g fN ⫹ i

.

The fact that g~ f! N is biased has three important consequences. First, since one given simulation does not have an infinite number of batches, but only I of them, the result is false ~the error being a decreasing function of I !. Second, it means that if one performs J independent simulations, the resulting distribution of N does not represent the statistical distribution of N ~no “error bars”!. Finally, this statistical independence of true Monte Carlo

(

k⫽1

再 冎 再 冎



I ⫺1 I

Z1

s[

M

Z 12



Z 22

⫹ {{{ ⫹ Z n2

冊冎

,

~2! where ~Z 1 , . . . , Z n ! ⫽ vector made of N iid N ~0, 1! components i ⫽ M ⫺1. Those two formulas are UMVE of N: N ⫽ E ~ gI 1 ~fR !! ⫽ E ~ gI 2 ~fR !!; the first formula supposes the ability to calculate all the pair derivatives of g, and NUCLEAR SCIENCE AND ENGINEERING

VOL. 167

FEB. 2011

MC BIASES0ERRORS IN BURNUP CALCULATIONS

the second one admits that one can evaluate the real part of a complex value of g and uses a Monte Carlo procedure to build the imaginary part of the random variable. To introduce Sec. III, one can now consider that g is an exponential function of the form g~f! ⫽ e ⫺Ltf , where L is a constant and t is the time. All the pair derivatives of g, which we call 2k, can then be written ~Lt ! 2k g~f!. An equivalent formulation of Eq. ~1! is given in Ref. 11 as `

N ⫹ gI 1 ~fR ! ⫽ g~ f!

(

冉 冊

~⫺1! k g ~2k! ~ f! N s2 k!

k⫽1

k

2

,

~⫺1! `

gI 1 ~fR ! ⫽ g~ f! N

( k⫽0

冉 M冊 Lt

2

k!

,

where we recognize the Taylor series development of an exponential function: gI 1 ~fR ! ⫽ g~ f!e N

N

Lt N N~t < ! ⫽ N~0!e < .

~5!

Thus, if g is the function that gives the concentration N of a given nuclide i, one can write N

Lt N g~f! ⫽ $ N~t < !% i ⫽ $ N~0!e < %i .

~6!

If we choose Eq. ~2! to estimate g~f I R !, we can then write LN t gI 2 ~fR ! ⫽ Re$$ N~0!e < %i % ,

~7!

N ij ⫽ l ij ⫹ sij f * and where LN * is defined by $ L% f * ⫽ fN ⫹ i

2k

s

⫺~Lt~s0M 2!! 2

Equation ~3! can be integrated, and one finds

N*

where s is the known standard deviation of $ fN i % ~instead of the estimated standard deviation!. If we use that last formulation and the expression of the pair derivatives, Eq. ~1! can be rewritten as follows: k

167



I ⫺1 I

Z1

s[

⫹ Z 22 ⫹ {{{ ⫹ Z N2

2 1

MZ

which we can rewrite f * ⫽ fN ⫹ ij for the sake of simN ij ⫽ plicity. Thanks to this last notation, Eq. ~4! becomes $ L% l ij ⫹ sij fN ⫹ isij j, and Eq. ~7! can be rewritten as follows: N

T

Lt⫹i N sjt T < %i % , gI 2 ~fR ! ⫽ Re$$ N~0!e

.

N

T

Lt N T < e i sjt %% i , gI 2 ~fR ! ⫽ $Re$ N~0!e

This last result can be rewritten as follows: 2

and

g~ f! N ⫽ gI 1 ~fR !e ~Lt~s0M 2!! . The conclusion is that in this case one finds the mean of the lognormal law. Indeed, g has been chosen as an exponential function, valued on a normal law. The density of this function is therefore lognormal, and its mean has 2 to be corrected by a factor of e ~Lts! 02 . In Sec. III we generalize this result to a matrix exponential function, in order to calculate a correction factor in the general case of several nuclides.

N

Lt N < cos~ sjt T !% i . gI 2 ~fR ! ⫽ $ N~0!e

~8!

This last expression is therefore an unbiased minimum variance estimator of the depletion operator valued on a normal mean. If we write N< B ~t ! as the usual biased expression of the concentration vector, the nonbiased estimate of this vector can be written N< NB ~t !, and we finally obtain T ! , N< NB ~t ! ⫽ N< B ~t !cos~ sjt

III. UMVE OF A MATRIX EXPONENTIAL FUNCTION OF A NORMAL MEAN

T ij ⫽ sij and with $ s%

As indicated by, for example, Ref. 17, the Bateman equation that gives nuclides depletion can be written as follows:

j⫽

dN~t < ! dt

⫽ LN N~t < ! ,

~3!

where N~t < ! is a vector made of all nuclide concentraN ij tions and LN is a transition matrix with elements $ L% given by N ij ⫽ l ij ⫹ sij f , $ L% ~4! where l ij ⫽ decay constants from nuclide i to j sij f ⫽ alimentation from i to j due to neutron interactions. NUCLEAR SCIENCE AND ENGINEERING

VOL. 167

,

FEB. 2011



I ⫺1 I

s[

Z1

MZ

2 1

⫹ Z 22 ⫹ {{{ ⫹ Z N2

~9!

.

This last expression, as demonstrated in Ref. 1, can even be simplified if each mean flux is estimated with only two estimations of fi ~i.e., I ⫽ 2!: j⫽

f1 ⫺ f 2 2

.

~10!

Obviously, in pretty much the same way as the matrix exponential is calculated with its limited development, N

e LtN ⫽

`

(

k⫽0

LN k t k k!

,

168

DUMONTEIL and DIOP

and the matrix cosine is calculated via T !⫽ cos~ sjt

`

( ~⫺1! k k⫽0

sT ~jt ! 2k

~2k!!

2k

.

~11!

IV. SOLVING THE COUPLED BOLTZMANN0BATEMAN EQUATION

Since we have derived an UMVE for the matrix exponential function ~the Bateman equation! of a normal mean ~from the stochastic solving of the Boltzmann equation!, we are able to estimate the concentrations after one time step. Now, it is legitimate to wonder if the result still holds for an arbitrary number of time steps. If we consider the whole process—which consists of transporting neutrons with a Monte Carlo code, calculating a mean flux, then computing in Eq. ~9! the unbiased value of the concentration vector, restarting the Monte Carlo calculation with this new vector, and repeating this nT times—the question is, is the final estimation unbiased and is it minimum variance? The answer is yes, by trivial recurrence, and only because the estimator for one time step is minimum variance. Concerning the implementation of such an unbiased estimator of concentration, two solutions are possible. In the first solution, the deterministic depletion solver uses Runge Kutta type methods to solve Eq. ~3!, and then it is possible to linearize the nuclide decay chain ~this is called the linear decay chain method; see Ref. 18! and to use Eq. ~1! @if the chain is linear, the pair derivatives of Eq. ~3! can be formally derived#. In the second solution, the depletion solver calculates the matrix exponential of Eq. ~5!, and in this case one can just evaluate Eq. ~9!, calculating the matrix cosine of Eq. ~11! ~with a limited development! in exactly the same way as the matrix exponential. It is always possible to use Eq. ~9! in which formulation is “ready to use,” instead of Eq. ~1!, which requires all the pair derivatives of g.

Three estimators of the concentration have been implemented in a simple Monte Carlo code: the biased one, denoted gI 0 ~fR !, and the previously mentioned gI 1 ~fR ! and gI 2 ~fR !. To calculate gI 1 ~fR !, the pair derivatives of g were derived: g ~2k! ~f! ⫽ ~st ! 2k g~f!. The maximum value of k ~the index of the sum! was arbitrarily fixed to 10; gI 2 ~fR ! was evaluated thanks to Eq. ~9!. Their mean value at each batch number is presented in Fig. 1, where a straight line indicates the correct value ~g~fR !!. As can be seen in Fig. 1, both corrected estimators converge toward the correct value, whereas the noncorrected one exhibits a clear bias ~NNB 0NB ' 2.0202.055 ' 0.984!. This bias can be calculated formally. If we apply Eq. ~9! to our problem, using Eq. ~10! to evaluate j we find NNB ⫽ NB cos

冉 冊 Df 2

st

.

~12!

The first order of Eq. ~12! gives NNB NB

⬇ 1⫺

冉 冊

1 Df 2

2

2

st

,

where Df is a difference of 2 iid random variables, and thus, ^~f1 ⫺ f 2 ! 2 & ⫽ x 2 ~2!sf2N and ^~f1 ⫺ f 2 ! 2 & ⫽ 2sf2N , so we find

冓 冔 冓 冉 冊冔 NNB NB

⫽ 1⫺

1 Df

2

2

2

st

⫽1 ⫺

冉 冊 sfN st 2

2

. ~13!

V. RESULTS IN A SIMPLE CASE

In this section we consider a homogeneous material made of one nuclide with cross section S and initial concentration N~0!. This material is irradiated by a constant flux f~ x,? V, < E ! ⫽ fR . V.A. Numerical Results for One Step To exhibit the difference between the biased evaluation of N~t ! and its nonbiased estimation, a numerical test has been performed. The law of the random variable f is N~10, 2!. There are 5000 stochastic evaluations ~called batches! of the random variable f,N which is itself made of 30 evaluations of f. Finally, the following numerical values were chosen: st ⫽ 0.5 cm⫺1 {s and N~0! ⫽ 300.

Fig. 1. Mean of g~f I R ! for three different estimators versus batch number ~estimator 0 to 2 from top to bottom!. NUCLEAR SCIENCE AND ENGINEERING

VOL. 167

FEB. 2011

MC BIASES0ERRORS IN BURNUP CALCULATIONS

This formula is an approximation, since we used a truncation of the Taylor series development of the cosine. Moreover, we used the particular expression of Eq. ~12! with I ⫽ 2 to derive the general formula Eq. ~13! that only depends on sfN . This was possible here since we are in a simple case. If one then applies it to I ⫽ 30, one obtains NNB 0NB ⬇ 0.991. This is in good agreement with the value of 0.984 calculated from Fig. 1.

f~NNB ! f~NB !

冔 冓 冔 NB



NNB

step2

~sfN !stepi ⫽ O

,

step1

冉M 冊 1

n

,

᭙i ,

then the effect of the initial bias of the concentration on the flux is

冉冊

冓 冔

⫽O

冓 冔

⬍⬍ sfN .

dfNB

and therefore,

Still considering this simple example, we would like to quantify the propagation of this bias through an arbitrary number of time steps nT . Let us imagine that a first Monte Carlo transport produced a flux f at step 1. This flux has its own standard deviation, and this introduces a bias on the concentration N at step 1. This impact was quantified in Sec. V.A. Now how does this propagate at the second step? One can try to model the flux dependence on the concentration ~this approximates the answer of the Monte Carlo code!. If we take S ⫽ Ns and f ⫽ S0 0S ~this seems reasonable since in the frame of coupled transport0depletion one always makes the hypothesis that the flux has a flat distribution!, then we can write



Since we know that

fB

V.B. Error Propagation Due to the Bias for N Steps

169

step2

dfNB fB

1

n

,

~15!

step2

At step 2 the order of magnitude of the bias on the flux is negligible compared to the statistical noise on the flux due to the Monte Carlo method. On top of that, this reasoning can be reproduced for any time step I, and consequently, at the end of the nT time steps, the bias is also negligible compared to the statistical error. However, it should be noted that correcting this bias remains crucial. Indeed, the standard procedure to evaluate the statistical errors in Monte Carlo simulations consists of performing independent evaluations ~replicas! of a given observable and looking at the dispersion of the results. This procedure allows decreasing arbitrarily the statistical error. If there is a bias in each replica, the statistical error can become smaller than the bias, and the final result will be wrong.

and using Eq. ~13! this becomes

冓 冔 冓 fNB fB



step2

f~NNB ! f~NB !

⫽ 1⫺

⬇1 ⫹



1⫹

dfNB fB





step2

1⫺

⬇1 ⫹

冉 冉



VII. CONCLUSIONS AND FUTURE WORK

step2

1

冊 冊

~sfN !step1 st 2 ~sfN !step1 st 2

冉 冉

1

2

冊 冊

~sfN !step1 st 2

~sfN !step1 st 2

2

,

2

2

,

and

冓 冔 冉 dfNB fB



step2

~sfN !step1 st 2

NUCLEAR SCIENCE AND ENGINEERING



2

. VOL. 167

~14! FEB. 2011

Previous attempts to calculate the error propagation during Monte Carlo burnup calculations were made during the last years ~see Ref. 19, for example!. The most convincing one consists of performing “brute force” Monte Carlo simulation. Indeed, the usual numerical scheme to solve Boltzmann0Bateman coupled equations can be replicated many times ~with different initial random seed!, and the final statistical errors for any quantity can be estimated by looking at the standard deviation of the replicas for this quantity. The major problem doing so is the bias of this method, since the mean of a nonlinear function of a random normal variable does not coincide with the value of the function for the mean of the random variable. One way to treat this problem consists of finding an unbiased minimum variance estimator of this nonlinear function ~here, the concentration vector over all nuclides!. In the particular case of one nuclide depletion, the nonlinear function is a simple exponential, and using Eq. ~1! allows finding the well-known result for the mean of a lognormal law. In the more general case of several nuclides, the unbiased estimator has been derived still thanks to Ref. 1 in the case of a matrix exponential function ~that represents the depletion operator! and was found to be @we recall Eq. ~9!#

170

DUMONTEIL and DIOP

N< NB ~t ! ⫽ N< B ~t !cos~ sjt T ! . This expression can be very easily integrated in matrix exponential solvers since Taylor series development of the cosine uses the same matrices as the one of the exponential. This estimator has been applied on a very simple example and for one time step, and the numerical value of the bias estimated with a Monte Carlo simulation was found to be consistent with the first-order development given by formula @Eq. ~13!#. This work should open new perspectives in the field of Monte Carlo burnup simulations, since the implementation of Eq. ~9! allows one to have an unbiased result, to measure the statistical errors ~from several estimations of this unbiased result!, and to fully parallelize the whole code ~not only the transport part of it!. Obviously, all this work supposes the flux to be constant in a given volume, and this hypothesis is obviously not valid anymore for strong burnup cases. The more general study of any probability distribution of the flux is currently under investigation by the authors. ACKNOWLEDGMENTS The authors wish to thank sincerely F. Malvagi and A. Lokhov for their help and advice, and AREVA and EDF who have partially supported this work. REFERENCES 1. A. STEFANSKI, J. NOVICK, and V. DEVANARAYAN, “Estimating a Nonlinear Function of a Normal Mean,” Biometrika, 92, 3, 732 ~2005!. 2. H. L. GRAY, T. A. WATKINS, and W. R. SCHUCANY, “On the Jackknife Statistic and Its Relation to UMVU Estimators in the Normal Case,” Commun. Statis., A, 2, 285 ~1973!. 3. E. DUMONTEIL and C. M. B. DIOP, “Unbiased Minimum Variance Estimator of a Matrix Exponential Function. Application to Boltzmann0Bateman Coupled Equations Solving,” Proc. Int. Conf. Mathematics Computational Methods and Reactor Physics, Saratoga Springs, New York, May 3–7, 2009, American Nuclear Society ~2009!. 4. J. P. BOTH, Y. K. LEE, Y. PÉNÉLIAU, O. PETIT, and B. ROESSLINGER, “Tripoli4. A Three Dimensional Polykinetic Particle Transport Monte Carlo Code,” presented at SNA’2003, Paris, France, September 22–24, 2003. 5. Y. K. LEE, B. ROESSLINGER, and A. TSILANIZARA, “Tripoli-4 Pepin Depletion Code and Its First Numerical Benchmarks for PWR High-Burnup UO2 and MOX Spent Fuel,” Proc. Mathematics and Computation, Supercomputing, Reactor Physics and Nuclear and Biological Applications, Avignon, France, September 12–15, 2005. 6. H. R. TRELLUE, “Development of Monteburns: A Code That Links MCNP and ORIGEN2 in an Automated Fashion

for Burnup Calculations,” LA-13514-T, Los Alamos National Laboratory ~1998!. 7. W. HAECK and B. VERBOOMEN, “An Optimal Approach to Monte Carlo Burnup,” Nucl. Sci. Eng., 156, 2, 180 ~2007!. 8. M. L. FENSIN, J. S. HENDRICKS, and S. ANGHAIE, “The Enhancements and Testing of the MCNPX Depletion Capability,” Proc. Int. Congress on Advances in Nuclear Power Plants (ICAPP ’08), Anaheim, California, June 8–12, 2008, American Nuclear Society ~2008!. 9. G. S. CHANG, “MCWO-Linking MCNP and ORIGEN2 for Fuel Burnup Analysis,” Proc. Topl. Mtg. Monte Carlo, Chattanooga, Tennessee, April 17–21, 2005, American Nuclear Society ~2005!. 10. N. A HANAN, A. P. OLSON, R. B. POND, and J. E. MATOS, “A Monte Carlo Burnup Code Linking MCNP and REBUS,” ANL-TD-CP-97492, Argonne National Laboratory ~1998!. 11. J. CETNAR, W. GUDOWSKI, and J. WALLENIUS, “User Manual for Monte-Carlo Continuous Energy Burnup ~MCB! Code—Version 1,” draft report, Royal Institute of Technology, Department of Nuclear and Reactor Physics, Stockholm, Sweden ~2005!. 12. D. J. KELLY, “Depletion of a BWR Lattice Using the RACER Continuous Energy Monte Carlo Code,” Proc. Int. Conf. Mathematics and Computation, Reactor Physics, and Environmental Analyses, Portland, Oregon, April 30–May 4, 1995, American Nuclear Society ~1995!. 13. C. TIPPAYAKUL, K. IVANOV, and S. MISU, “Improvements of MCOR: New Monte Carlo Depletion Code System for Fuel Assembly Reference Calculations,” Proc. PHYSOR 2006 Conf., Vancouver, Canada, September 10–14, 2006, American Nuclear Society ~2006!. 14. E. M. GELBARD and R. PRAEL, “Computation of Standard Deviations in Eigenvalue Calculations,” Prog. Nucl. Energy, 24, 237 ~1980!. 15. E. DUMONTEIL and T. COURAU, “Dominance Ratio Assessment and Monte Carlo Criticality Calculations: Dealing with High Dominance Ratio Systems,” Nucl. Technol., 172, 120 ~2010!. 16. A. N. KOLMOGOROV, “Unbiased Estimates,” Izv. Akad. Nauk SSSR, Ser. Mat., 14, 303 ~1950!. 17. S. J. BALL and R. K. ADAMS, “MATEXP, a General Purpose Digital Computer Program for Solving Ordinary Differential Equations by the Matrix Exponential Method,” ORNLTM-1933, Oak Ridge National Laboratory ~1967!. 18. J. CETNAR, “General Solution of Bateman Equations for Nuclear Transmutations,” Ann. Nucl. Energy, 33, 640 ~2006!. 19. T. TAKEDA, N. HIROKAWA, and T. NODA, “Estimation of Error Propagation in Monte-Carlo Burnup Calculations,” J. Nucl. Sci. Technol., 36, 9, 738 ~1999!. NUCLEAR SCIENCE AND ENGINEERING

VOL. 167

FEB. 2011

ARTICLE

Systematic Uncertainty Due to Statistics in Monte Carlo Burnup Codes: Application to a Simple Benchmark with TRIPOLI-4-D Emeric BRUN*, Eric DUMONTEIL and Fausto MALVAGI Commissariat à l’Energie Atomique, Centre de Saclay, Service d’Etude des Réacteurs et de Mathématiques Appliquées (DEN/DANS/DM2S/SERMA/LTSD), F91191 Gif-sur-Yvette CEDEX, France

A new burnup code, TRIPOLI-4-D, has been developed at CEA Saclay. It relies on the use of TRIPOLI-4 for the Monte Carlo neutron transport part, and on MENDEL nuclide depletion solver for the burnup calculation. Its main characteristics are pointwise energies, probability tables for the unresolved resonances range, potential fine multigroup calculation of reaction rates, the fourth order Runge-Kutta to solve Bateman's equations, predictor/corrector schemes, novice and expert levels (either launch the calculation with TRIPOLI-4 data set or potential C++ scripting to interact with the coupling macro, through the use of CINT, the ROOT C++ interpreter), post-processing with ROOT, possibility to perform parallel calculations, restart capability. This code owns a special parameter that allows the user to calculate the statistical uncertainty on any physics quantity using the Central Limit Theorem (this very classical Monte Carlo method is also often referred to as “brute force” or “independent replicas” since it consists in replicating a given simulation many times using different random sequences). This technique is particularly adapted to massive parallelism, but can also be set up sequentially on personal computers. In this paper we will apply it on a simple case study, a Pressurized Water Reactor (PWR) cell, and we will analyze different quantities with their associated statistical error bars. Comparing those results to reference simulations, we will try to highlight systematic errors, inherent to Monte Carlo burnup codes, and we will show that a particular systematic error, due to the number of simulated particles, has to be considered while performing Monte Carlo burnup simulations. KEYWORDS: TRIPOLI-4-D, Monte Carlo, burnup, TRIPOLI-4, uncertainty propagation, parallelism. 1

I. Introduction

Monte Carlo burnup codes are of quite recent interest for the neutronic community. Those codes are all based on a linking between Monte Carlo transport codes (e.g. MCNP,1) MCNPX,2) TRIPOLI-43)...) and burnup capabilities (ORIGEN,4) MENDEL5)...). Most of them are based on MCNP(X), like MONTEBURNS,6) MCNPX,7) ALEPH,8) VESTA,9) MURE10) but this year the first version of TRIPOLI-4-D11,12) has been released. Formally, all those codes solve the coupled Boltzmann/Bateman equations. Doing that with a proper Monte Carlo (MC) code would require to store all the isotopes and collision sites to be able to perform a true microscopic depletion of the fuel. Since this is obviously not possible (nowadays at least) this non linear problem is solved by successive iterations over a transport part (with fixed compositions) and a depletion part (with fixed fluxes or reaction rates). But doing so, systematic errors of a priori unknown magnitude can arise. Moreover, most codes do not propagate uncertainties (in fluxes, reaction rates and concentrations) from the first step to the last, but they limit themselves to providing step by step values. In this paper we will first present the most general fea*Corresponding author, E-mail: [email protected] © Atomic Energy Society of Japan

tures of TRIPOLI-4-D. We will then focus on the uncertainty sources (arising from both statistical and systematic errors) in Monte Carlo codes, emphasizing the treatment of statistical uncertainties in TRIPOLI-4-D, based on independent simulations thanks to its parallelism capabilities. Finally, we will define a simple but realistic benchmark (a Pressurized Water Reactor (PWR) cell). This benchmark will be studied to analyze a particular systematic error due to the number of neutrons, foreseen by Dumonteil and Diop13,14) and by Z. Xu et al15) on simple analytical or numerical benchmark with ad hoc Monte Carlo codes. In this paper we analyze and quantify the effect of the number of simulated neutrons on the final biais of the calculation. And this on a realistic case (a PWR pin cell) and with an industrial code (TRIPOLI-4-D). We show that it is something of worrying importance.

II. TRIPOLI-4-D TRIPOLI-4-D is made of two C++ interfaces that wrap the functionalities of both TRIPOLI-4 and MENDEL. The first is a set of generic methods for MC neutrons transport, allowing to run a simulation, or to dynamically reload compositions or the geometry for example. The second provide methods to perform material depletion calculation with the fourth order Runge-Kutta method. Both interfaces are linked

to the C++ interpreter CINT16) allowing to build advanced coupling schemes between transport and burnup which are either run interactively or compiled and executed with coupling scripts. This interpreter belongs to the ROOT17) libraries developed at CERN and under LGPL public license. ROOT is linked to TRIPOLI-4-D offering numerous functionalities in terms of data treatment, analysis and visualisation. From a user point of view its main characteristics are as follows:  Use of CINT to launch transport/depletion sequences either interactively or in batch mode.  Time irradiation performed with either fluxes or reaction rates, or both of them at the same time, on an isotope by isotope basis. In case fluxes are calculated, GENDF data (processed by NJOY18)) are used to calculate reaction rates afterward.  Simulations with either a fixed flux, fixed source intensity or a fixed power defined by the user. In this last case, the power can be automatically calculated with TRIPOLI-4 deposited energy tally, or with provided kerma data.  Three different time schemes: Euler explicit and two predictor-corrector (order 2) time schemes (midpoint and mean). Those two schemes are described in Ref. 19).  Automatic statistical error propagation calculation using independent simulations.  Full parallelisation with high scalability of the whole transport/depletion procedure.  Use of ROOT and PROOF17) (its parallelisation engine) to provide and speed up both the visualisation and the analysis of the results.

III. Uncertainties Propagation in MC Burnup Codes Since Monte Carlo burnup codes make more assumptions than simple Monte Carlo transport codes, this section details the different sources of uncertainties inherent to burnup code, focusing on their treatment in TRIPOLI-4-D. (1) 1. Statistical Uncertainties In standard Monte Carlo transport codes, each calculated tally is provided together with an associated statistical uncertainty (the standard deviation of the probability density associated to the tally). In most Monte Carlo burnup codes, the coupling script iterates between a parallelized transport step and a mono-processed depletion step. At the end of the procedure, the statistical uncertainty of the transport step is thrown away and not propagated to the burnup step nor to the next transport. At the end, no confidence intervals are provided on isotopic inventories. One of the solutions to propagate the statistical uncertainties through the different burnup steps is to perform independent simulations (often referred to as independent replicas or brute force20)) of this whole procedure: each transport/depletion iteration over all the time steps (each with a different sequence of random numbers) has to be re-

peated many times, to be able to compute a probability distribution for each tally. This procedure was implemented in TRIPOLI-4-D, which gives the user a reliable propagated statistical uncertainty. On top of that TRIPOLI-4-D uses this independent simulations mechanism to implement its parallelism ability; each processor handles its own set of independent MC burnup simulations, and the results are gathered at the end. (2) (3) 2. Systematic Error Once the statistical uncertainties correctly computed, they can be used to detect systematic errors: the statistical uncertainty can indeed be decreased by increasing the number of simulated neutrons, allowing finding potentially small systematic effects. As explained in the introduction, Monte Carlo transport does very few assumptions, and the simulations are therefore often considered as reference ones because of this. Monte Carlo burnup codes are however really different, since at least three causes of systematic errors add to the statistical uncertainties:  Fluxes and reaction rates are considered as stationary over the different step of the time scheme.  Isotope densities are considered as uniform in each depletion volume/region (while in reality it is continuously varying over space).  Bateman equation are non linear with respect to the fluxes and this induce a systematic error due to the number of neutrons (the lower the number of neutrons in each simulation, the bigger the bias). This last bias has been extensively theoretically studied and can eventually be corrected.13,14) It is caused by the fact that Bateman equation (that we call g ) being non linear function of the flux, the real concentration of an isotope (called N R ) is a function of the real flux  R but differs from the expectation of Bateman equation solving with different estimated flux

:

N R  g (R )  E ( g ( )) .

(1)

In the following section of this paper we will try to quantify it with TRIPOLI-4-D on a very simple case study: a PWR cell.

IV. Application 1. Description of the Case Study This case study is a 4.9% enriched UOX PWR cell depleted until 60 GWd/t. The fuel pin (Fig. 1) is divided into six annular depletion regions to take into account spatial flux distribution (spatial self-shielding effect). Reflection boundary conditions are attached to the 6 faces of the cell. JEFF3.1 nuclear data were used to run the simulations, with an optimized isotopic depletion chain made of 161 isotopes. For the sake of simplicity, the first order Euler depletion scheme was used. The irradiation period was divided into 45 steps. For each step, a 11,500 multigroup flux is estimated by TRIPOLI-4. From this flux and GENDF cross sections, reaction rates are built for each isotope except for U-238 which reaction rates are directly estimated by TRIPOLI-4 so

Fig. 1

Geometry of the PWR cell case study

Fig. 2 Concentration of U-235 versus burn-up in the inner ring of the fuel cell

as to properly capture self-shielding phenomenon. To avoid source convergence problems while using a low number of neutrons, the simulations were performed in fixed source mode, with a uniformly distributed source. Two types of flux normalization were considered in this study: normalization to the power and fixed source intensity. The latter case was also studied to get rid of the normalization procedur used in the more realistic case of the normalization to the power since it may induce a bias propagating from step to step. 2. Results To quantify the bias, different TRIPOLI-4-D simulations were run, each with a different number of neutrons per simulation (either 11, 55, 1010, 2020, 100100, 1001,000 or 1,0001,000 neutrons, the first number being the number of cycles and the second the number of neutrons per cycle). However, the overall number of independent simulations in each case was adapted in order to have similar standard deviations. The reason why such small numbers of neutrons are compared is that it provides a hint of what really happens when one simulate a full core. Indeed, when a full core of 100 axial zones and 200 assemblies is simulated, each being made of 1717 pin cell, which are themselves made of 6 burnup regions, then a total number of roughly 108 volumes have to be explored by the neutrons. With one billion neutrons for the whole simulation the average number of neutrons per volume is therefore of only 10. On top of that, this mean number even does not tell that some regions will have a good statistics of neutrons (e.g. outer rings of the pin cell) while some other will systematically see very few neutrons (e.g. inner zones), worsening the situation. The following results are presented first for fixed power simulations, then for fixed source intensity. Each of these normalization methods induces a different expression of the renormalization factor (RF) used to settle the flux and reaction rates level before depletion calculations. (1) Fixed Power In this case, a power level is specified by the user (Puser). The flux level given by TRIPOLI-4 is arbitrary (it depends

Fig. 3 Zoom on the concentration of Pu-239 for the last burn-up steps in the inner ring of the fuel cell

essentially on the neutron source intensity) and corresponds to the power PT4 (calculated as the sum on all isotopes of the fission and capture rates multiplied by their corresponding energy release). The expression of RF is then:

RF  Puser / PT 4 . Ten independent simulations of the case with 1,000 cycles of 1,000 neutrons have been performed. Since this case should be the less biased and that statistical fluctuations between the 10 simulations remains negligible this case represents the reference. Figure 2 gives the U-235 concentration in the inner fuel cell as a function of the burnup for different number of neutrons per independent simulation. The one-sigma error bars are represented on top of the points but are small and do not appear at first sight. Only the 1x1 case seems to behave differently from the others. A detailed analysis of numerical values shows that the 55, 1010 and 2020 cases are also outside the statistical uncertainties, and biased of as much as 7%. Figure 3 gives also the concentration versus burnup, but for Pu-239. Pu-239 is produced while U-235 disappears, and therefore the Pu-239 concentrations given by the different simulations are growing values of the burnup. Like U-235,

Fig. 4

Renormalization factor versus burn-up

this figure indicates that the bias is a decreasing function of the number of neutrons. The two last figures also show that the systematic error gets bigger with increased burnups. This is confirmed by a detailed numerical analysis that shows that for Pu-239, in the case of 55 neutrons, the bias goes up to 15%. The mechanisms explaining the propagation and the amplification of this phenomenon from one step to the other do not rely only on the biased concentrations. Indeed, once the concentrations are biased different parameters can link one burnup step to the other, like the normalization factor, and become biased (Figs. 4 and 5). Obviously, once the concentrations are biased the transport sequences start from wrong material properties and the flux is also biaised (Fig. 6). To characterize the bias more precisely Table 1 indicates for three selected families of isotopes (U, Pu and Xe) both

Fig. 5

Zoom on the renormalization factor versus burn-up

Fig. 6

Flux in the inner ring of the fuel cell versus burn-up

Table 1 Test of compatibility between selected isotopes concentration distributions and 100 independent replicas of 1001,000 in the fixed power case. (Diff is the relative difference between means : Diff = (m-mref)/mref ; p is the significance of the Student’s t test) 5*5 Isotope

Diff %

p

20*20 Diff %

p

100*100 Diff % p

PU236 PU238 PU239 PU240 PU241 PU242 U232 U233 U234 U235 U236 U237 U238 XE130 XE131 XE132 XE133 XE134 XE135

1.40 -0.82 14.00 11.00 5.60 -4.10 -0.07 31.00 4.90 6.40 0.38 15.00 -0.19 -3.30 -2.20 0.26 1.40 -0.56 15.00

0.91 0.58 1.e-30 6.e-12 1.e-4 2.e-5 0.99 0.04 0.03 3.e-27 0.30 0.07 1.e-27 3.e-05 0.03 0.45 0.15 2.e-4 3.e-13

-0.097 0.11 0.58 -0.025 0.77 -0.58 -0.88 0.36 0.10 0.38 -0.055 -0.39 -0.0064 -0.25 -0.087 -0.023 -0.14 -0.049 0.79

0.97 0.72 0.01 0.94 0.02 0.03 0.68 0.82 0.79 1.e-3 0.47 0.84 0.04 0.15 0.70 0.74 0.48 0.12 0.06

0.27 -0.13 0.024 -0.11 0.16

0.58 0.03 0.58 0.11 0.02

-0.14

3.e-3

-0.98 -0.54 -0.20 0.052 -0.027 -0.41 -8.6e-4 -0.06 0.086 -0.014 -0.04 7.9e-3 0.16

0.02 0.06 0.005 0.02 0.05 0.31 0.16 0.06 0.05 0.32 0.30 0.21 0.04

XE136

-2.4

1.e-28

-0.15

1.e-3

7.8e-3

0.39

Fig. 7 Concentration of U-235 versus burn-up in the inner ring of the fuel cell

the difference in concentrations at the final burnup step for the inner ring between 55, 2020, 100100 neutrons simulations and 100 independent replicas of 1001,000 (instead of the reference case 1,0001,000 to ensure the robustness of the Student’s t test) and the significance p of the Student’s t test.21) The significance p is the probability that the difference between the means of two distributions could be this large just by chance for distributions with equal means. Therefore, a small significance (0.05 or 0.01) means that the observed difference is statistically very significant. The results below 1% probability are indicated in bold, and red. As one can see, in the 55 case, half of the chosen isotopes concentrations differ from the reference with pretty important differences. For the 2020 case this number falls to two isotopes over the 20 considered (U-235 and Xe-136), while only one remains different in the 100100 case (Pu-242). All those observations clearly tell that there is a bias depending on the number of neutrons in each independent simulation. (2) Fixed Source Intensity In this case, the flux level is imposed by the neutron source intensity defined by the user. Thus, RF=1. In the case of fixed source intensity almost no other parameters than concentrations couple a burnup step to the next one. As one can see on the same figures studied before, the bias stays roughly unchanged with respect to the fixed power simulations (Figs. 7 to 9 and Table 2). Table 2 even indicates that the results are worse in the fixed source intensity case. Indeed, the comparison of simulations with 55, 2020 and 100100 neutrons to the reference case shows that 7, 9 and 2 isotopes respectively over 20 fail to the Student’s t test. These numbers can be compared to 10, 2 and 1 isotope over 20 that fail to this test in the case of fixed power. One can conclude that the calculation bias showed by simulations with an insufficient number of particles is really linked to the burn-up calculation, and not to the normalization procedure.

V. Conclusion This paper aimed at giving a brief overview on uncertainties treatment in MC burnup code, and in particular in TRIPOLI-4-D. Two different kinds of uncertainties have to

Fig. 8 Concentration of Pu-239 versus burn-up in the inner ring of the fuel cell

Fig. 9

Flux in the inner ring of the fuel cell versus burn-up

be taken into account, and propagated through the burnup steps. Statistical uncertainties (propagation) have to be taken into account in MC burnup code, and the formal approach to deal with them is by using independent replicas. This was implemented in TRIPOLI-4-D and allowed to work on the systematic uncertainties in burnup codes. Those may arise from at least three sources: the discretization in time, in space, or in number of neutrons (i.e. the time/burnup step structure, the spatial binning, and the number of neutrons). This last source of systematic errors has been studied on a PWR cell with TRIPOLI-4-D. It has been shown that there was indeed a potentially strong bias depending on the number of neutrons in independent simulations. This bias has to be dealt with because it may affect the whole burning procedure in case the depletion regions have too few neutrons. This may happen in case of full core modelling, because to take into account the spatial gradient of the flux, the core has to be studied on a pin by pin basis. If one considers 6 depletion volumes in each pin cell, 1717 pin cells per assembly, roughly 200 assemblies in the core and 100 axial zones, then even for 100100 neutrons per volume the total number of simulated neutrons should be of at least 1012 to get rid of the effect. If the brute force approach is used to treat statistical uncertainties, the user have to make sure that not only the overall number of neutrons is big

Table 2 Test of compatibility between selected isotopes concentration distributions and 100 independent replicas of 1001000 in the fixed source intensity case. (Diff is the relative difference between means : Diff = (m-mref)/mref ; p is the significance of the Student’s t test) 5*5

20*20 Diff % Isotope

100*100 Diff % p

0.22 0.27 2.e-7 3.e-5 0.06 1.e-3 0.13 0.02 0.48 7.9e-11 0.71 0.96 1.e-09 0.93 0.12 0.03 0.77 1.e-5 0.53

-2.60 0.70 0.53 0.86 0.79 1.20 -3.00 -1.30 -0.41 0.048 5.4e-3 -0.06 -3.7e-3 -0.47 -0.018 -0.039 8.3e-3 -0.061 -6.3e-3

0.29 0.11 4.e-6 1.e-6 8.e-3 1.e-4 0.22 0.47 0.52 7.e-11 0.87 0.97 3.e-11 2.e-4 0.75 0.25 0.98 0.007 0.97

0.74 0.39 -0.021 0.062 -0.19 -0.087 0.51 0.39 0.12 -1.5e-3 -4.2e-5 0.40 6.7e-5 0.023 0.022 -0.011 0.13 0.0029 -2.4e-3

0.12 1.e-4 0.31 0.11 0.003 0.11 0.23 0.22 0.27 0.34 0.99 0.2 0.52 0.41 0.02 0.06 0.09 0.53 0.95

1.e-16

-0.18

2.e-06

5.5e-3

0.45

Isotope

Diff %

p

PU236 PU238 PU239 PU240 PU241 PU242 U232 U233 U234 U235 U236 U237 U238 XE130 XE131 XE132 XE133 XE134 XE135

68.00 3.00 3.70 4.70 -3.30 -5.10 90.00 1.1e+2 2.90 0.31 0.085 -0.33 -0.021 -0.076 -0.42 -0.43 -0.48 -0.64 0.58

XE136

-1.80

enough, but also the number of neutrons per independent simulations. In all cases, the natural scale for fixing the number of neutrons is the uncertainty on nuclear data (which are not specific to MC burnup code but to MC transport codes more generally). Those uncertainties may affect the results of several percent on final quantities like concentration, reaction rates or fluxes. Further developments concerning TRIPOLI-4-D will concern the treatment of this bias using recently proposed methods,13,14) and the quantification of other systematic errors (e.g. the smooth variation of macroscopic cross sections).

Acknowledgment

5) 6)

7)

8) 9) 10)

The authors want to acknowledge EDF and AREVA for partial financial support, as well as C.M. Diop and A. Zoia for their fruitful advises. References 1) X-5 Monte Carlo Team, MCNP – A General Monte Carlo N-Particle Transport Code, Version 5, LA–UR-03-1987, Los Alamos National Laboratory (LANL) (2003). 2) J. S. Hendricks et al., MCNPX 2.6.0 Extensions, LA-UR-08-2216, Los Alamos National Laboratory (LANL) (2008). 3) C. M. Diop, E. Dumonteil, F. X. Hugot, C. Jouanne, Y. K. Lee, F. Malvagi, A. Mazzolo, O. Petit, J. C. Trama, “An Overview on the Monte Carlo Particle Transport Code TRIPOLI-4,” Trans. Am. Nucl. Soc., 97, 694-695 (2007). 4) M. J. Bell, ORIGEN – The ORNL Isotope Generation and De-

11) 12)

13)

14)

pletion Code, ORNL-4628 UC-32, Oak Ridge National Laboratory (ORNL) (1973). A. Tsilanizara, S. Lahaye, E. Brun, CEA Internal Report, (2009) H. R. Trellue, Development of Monteburns: a code that links MCNP and ORIGEN2 in an automated fashion for burnup calculations, LA-13514-T, Los Alamos National Laboratory (LANL) (1998). M. L. FENSIN, J. S. Hendricks, S. Anghaie, "The Enhancements and Testing of the MCNPX Depletion Capability," 2008 International Congress on Advances in Nuclear Power Plants, Anaheim, CA, (2008). W. Haeck, B. Verboomen, “An optimal approach to Monte Carlo Burnup,” Nucl. Sci. Eng., 156, 180-196 (2007). W. Haeck, VESTA User’s manual, IRSN DSU/SEC/T/2008-331 Indice A (2008). O. Meplan et al. "MURE : MCNP Utility for Reactor Evolution – Description of the method, first applications and results," ENC2005 European Nuclear Conference. Nuclear Power for the XXIst Century: From basics research to high-tech industry, Versailles, France (2005). E. Brun, E. Dumonteil, CEA Internal report (2010). R. Reyes-Ramírez, C. Martín-del-Campo, J.-L. François, E. Brun, E. Dumonteil, F. Malvagi, “Comparison of MCNPX-C90 and TRIPOLI-4-D for fuel depletion calculations of a Gas-cooled Fast Reactor,” Ann. Nucl. Energy, 37, 1101-1106 (2010). E. Dumonteil, C. M. B. Diop, “Unbiased minimum variance estimator of a matrix exponential function. Application to Boltzmann/Bateman coupled equations solving,” Proc. Int. Conf. on Mathematics Computational Methods & Reactor Physics, New York, May 3-7, 2009 (2009). E. Dumonteil, C. M. Diop, “Biases and error bars in Monte Carlo burnup calculations: an unbiased stochastic scheme to

15)

16) 17)

18)

solve Boltzmann/Bateman coupled equations,” Nucl. Sci. Eng., 167, 165-170 (2011). Z. Xu, J. Rhodes, K. Smith, “Statistical implications in Monte Carlo depletions,” Proc. of PHYSOR2010, Pittsburgh, Pennsylvania, USA, May 9-14, 2010 (2010). M. Goto, “C++ interpreter - cint,” CQ publishing (1997). I. Antcheva, M. Ballintijn, B. Bellenot et al., “ROOT - A C++ framework for petabyte data storage, statistical analysis and visualization,” Comp. Phys. Comm., 180, 2499-2512 (2009). R. E. MacFarlane, D. W. Muir, The NJOY Nuclear Data Processing System Version 91, LA-12740, Los Alamos National Laboratory (LANL) (1994).

19) M. L. Fensin, J. Hendricks, H. Trellue, S. Anghaie, Incorporation of a Predictor-Corrector Methodology and 1-group reaction rate reporting scheme for the MCNPX depletion capability, LA-UR-06-3925, Los Alamos National Laboratory (LANL); Trans. Am. Nucl. Soc., 95, 317 (2006). 20) N. Garcia-Herranz et al., “Propagation of statistical and nuclear data uncertainties in Monte Carlo burnu-up calculations,” Ann. Nucl. Energy, 35, 714-730 (2008). 21) W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Numerical Recipes in C, Second Edition, Cambridge University Press, (1992).

Joint International Conference on Supercomputing in Nuclear Applications and Monte Carlo 2013 (SNA + MC 2013) La Cité des Sciences et de l’Industrie, Paris, France, October 27-31, 2013

i

DEPLETION CALCULATIONS BASED ON PERTURBATIONS. APPLICATION TO THE STUDY OF A REP-LIKE ASSEMBLY AT BEGINNING OF CYCLE WITH TRIPOLI-4®. Cyril DIEUDONNE, Eric DUMONTEIL, Fausto MALVAGI and Cheikh M’Backé DIOP Commissariat à l’Energie Atomique et aux Energies Alternatives, Centre de Saclay, Service d’Etude des Réacteurs et de Mathématiques Appliquées (DEN/DANS/DM2S/SERMA/LTSD) F91191 Gif-sur-Yvette CEDEX, France Corresponding Author, E-mail: [email protected]

For several years, Monte Carlo burnup/depletion codes have appeared, which couple Monte Carlo codes to simulate the neutron transport to deterministic methods, which handle the medium depletion due to the neutron flux. Solving Boltzmann and Bateman equations in such a way allows to track fine 3-dimensional effects and to get rid of multi-group hypotheses done by deterministic solvers. The counterpart is the prohibitive calculation time due to the Monte Carlo solver called at each time step. In this paper we present a methodology to avoid the repetitive and time-expensive Monte Carlo simulations, and to replace them by perturbation calculations: indeed the different burnup steps may be seen as perturbations of the isotopic concentration of an initial Monte Carlo simulation. In a first time we will present this method, and provide details on the perturbative technique used, namely the correlated sampling. In a second time the implementation of this method in the TRIPOLI-4® code will be discussed, as well as the precise calculation scheme a meme to bring important speed-up of the depletion calculation. Finally, this technique will be used to calculate the depletion of a REP-like assembly, studied at beginning of its cycle. After having validated the method with a reference calculation we will show that it can speed-up by nearly an order of magnitude standard Monte-Carlo depletion codes. KEYWORDS: neutronic, Monte Carlo simulation, burnup/depletion calculation, perturbation, correlated sampling method, TRIPOLI-4®, ROOT

1. THE PERTURBED MONTE CARLO BURNUP METHOD 1.1 Context In neutronic simulations of reactor cores, several numerical methods can be used to solve the neutron transport equation. Among them, Monte Carlo codes make very few approximations or physics hypotheses. They model physical phenomena rigorously and simulate the particle-matter interactions in complex 3-dimensional geometries. They can also provide confidence intervals on computed tallies. Those codes can then be coupled with deterministic burnup solvers to reproduce the burning of materials under irradiation. The resulting codes, called Monte Carlo burnup codes, are suitable for reactor core calculations because they handle burnup effects while preserving an accurate modeling of physical phenomena and a flexible handling of geometries. However, compared to deterministic solvers, Monte Carlo burnup codes have an important counterpart in term of computation time, since they iterate over successive transport calculations. In order to handle such calculation times and to speed up Monte Carlo burnup codes, a new method has been recently developed. It consists in considering all burnup steps as consecutive perturbation of an initial Monte Carlo simulation. This “factorization” of the transport sequences may thus be of great interest in terms of

calculation time [1, 2]. To evaluate the impact of small input parameters changes, most Monte Carlo codes offer a perturbation option, often based on Taylor series expansion (e.g. MCNP6). 1.2 The correlated sampling A different approach is used by TRIPOLI-4®, which proposes the correlated sampling method [3-5] for perturbation studies. This method uses the neutrons collisions histories of an unperturbed calculation and adjusts the particle weights by considering the difference of reactions probabilities to solve the perturbed problem. Let us now describe the weight adjustment calculation of the correlated sampling method. Considering a set of heterogeneous media M0 and an initial neutron source distribution S, the solution of the Boltzmann equation at phase space point P is noted ψ0(P) and can be expressed by a Neumann series expansion as: !!

K !! S P

ψ! P =

(1)

!!!

where K0 represents the transition kernel, which can be written as the product of a collision kernel C0 and a transport kernel T0. Let M1 be a set of heterogeneous media given by an isotopic concentrations perturbation of M0. To compute

the neutron collision density ψ1(P) from tracks simulated in M0, the correlated sampling method uses several weight adjustments noted ω!" . Those corrections correspond to the ratio of the perturbed probability over the unperturbed probability of each event between M1 and M0: !!

!!

K!! ! K S P K !! !

K!! S P =

ψ! P = !!!

!!

!!!

(2)

ω!" K !! S P

= !!!

More specifically, two weights adjustments have to be considered. The first one concerns the inter-collision distance sampling. The probability for a neutron to be transported in a homogeneous medium Mm on a distance dP and to have a collision in phase space position P is ! !!! ! ! !!! ! !! . Thus, the weight is after collision, in the perturbed median, is modified by the factor: !

T! Σ!! P e!!! ! !" P + dP = ! ! T! Σ! P e!!! ! !" Σ!! P ! !!! ! !!!! ! = ! e Σ! P

!"

(4)

The second weight adjustment concerns the collision sampling. In TRIPOLI-4®, the isotope i undergoing a collision is sampled before the interaction, so the correction weight due to a collision depends only on the isotope sampling. The probability to sample an isotope in a ! ! homogeneous medium Mm is Σ!,! P Σ!! P , where Σ!,! P is the macroscopic total cross section of the isotope i in Mm, which leads to the following weight adjustment: ! C! Σ!,! P Σ!! P N!! Σ!! P P + dP = ! = C! Σ!,! P Σ!! P N!! Σ!! P

(5)

where N!! is the concentration of the isotope i in the medium Mm. It can be shown that the correlated sampling introduces no approximation or hypothesis and allows the unbiased calculation of any quantity for arbitrary large perturbations. Its only limitation is the resulting degraded variance. Consequently, using the correlated sampling to replace successive Monte-Carlo simulations of a large burnup calculation may lead to shorter computation time but also to a worse estimate of fluxes, reactions rates, concentrations, etc, for high burnup values. However, if an isotope appears at a given steps of the burning procedure, and was not present in the initial Monte-Carlo simulation, the calculation could be biaised. Indeed the isotope appearing has an impact on the flux distribution – and other tallies – that cannot be reproduced by the correlated

sampling since the initial simulation does not contain events implying this isotope. Therefore, in the following section, after having presented the code that enables the perturbed depletion calculation, variance reduction techniques will be discussed to ensure that tallies can be estimate with restrained variances at all burnup steps.

2. IMPLEMENTATION AND COMPUTATION SCHEMES 2.1. The Monte Carlo burnup card of TRIPOLI-4® and the correlated sampling To validate and qualify the performances of this method, it has been implemented in the depletion card of the TRIPOLI-4® code [6, 7]. This card couples TRIPOLI-4® [8] for the transport part and MENDEL [9] for the depletion part - both codes being developed at CEA Saclay- through a C++ interface using the ROOT framework developed at CERN [10]. ROOT is linked to TRIPOLI-4® to build advanced coupling schemes between transport and depletion codes. Those schemes are written in C++ scripts and executed by the ROOT C++ interpreter CINT [11]. ROOT is also linked to TRIPOLI-4® to record the simulated particles histories using ROOT file format persistency. This integrated development library allows setting up high-level calculation schemes, as the one proposed in this present paper. 2.2. Computation schemes The more trivial procedure to run a depletion calculation using perturbations, for a maximum burning time T and n time steps, is the following: • Step 1: o Step 1-a: time t is fixed to 0 o Step 1-b: Initial Monte-Carlo transport at t (and corresponding isotopic concentrations) using TRIPOLI-4® o Step 1-c: Recording of the tracks file containing all neutron histories • Step 2: Calculation of the flux on a fine multigroup mesh for every burned material using the tracks file of Step 1-c • Step 3: Depletion calculation to retrieve concentrations after an irradiation time of T/n using MENDEL, using as an input the fluxes calculated at Step 2 and producing as an output the isotopic concentrations for every burned materials • Step 4: Perturbation calculation using the track file generated at Step 1-c, where the perturbed isotopic concentrations are the ones calculated at Step 3: the weights associated to each neutron track are modified accordingly. The value of t is incremented of T/n. • Step 5: Iteration between steps 2, 3 and 4 until t=T The coupling corresponds to an Euler explicit scheme, which

is a scheme of order 1 in time. It has been implemented and tested. Two main difficulties appeared during the testing: on one side, many isotopes (fission products, poisons, etc) are produced during the first time steps of the depletion, affecting the tallies via the biaising effect discussed in section 1.2). The perturbation method could consequently not begin at the first time step because the associated track file was lacking too many isotopes, and had to be activated only after the beginning of the cycle. On the other side, even when all isotopes were present in the initial track file, the variance was degraded at an exponential rate that cannot be compensated by the degradation of the variance discussed in section 1.2. Indeed, the bigger the value of the total irradiation time, the larger the amplitude of the perturbation, and the more deteriorated the variance. To improve the methodology, we have to enhance the tracks representativeness during the depletion. Therefore, we have to change the isotopic concentrations used to produce the tracks file. Indeed tracks do not need to represent the state of the system at a given burnup step, and can instead be generated with unphysical values of concentrations ensuring small amplitudes of the perturbation. 2.3. Use of variance reduction techniques Those considerations led to the development of adapted variance reduction techniques. The more efficient one consists in doing a pre-calculation (either with a deterministic code, or with a Monte-Carlo burnup code but with less neutrons transported at each time step) allowing to quickly have a rough idea of the concentrations of each isotopes at each time steps. Then, in a second time, this rough knowledge of the concentrations at each time step is used in the standard burnup calculation to minimize the amplitude of the perturbation. The time grid is divided in regular intervals and for each interval a specific tracks file is produced. The size of those intervals, noted ΔT, is a compromise between the burnup of intervals (to minimize the amplitude of the perturbation) and the number of time steps (to maximize the speed-up factor on the computation time). Therefore, the optimized procedure is the following:

o Step 3: Calculation of the flux on a fine multigroup mesh for every burned material using the tracks file of Step 1 o Step 4: Depletion calculation to retrieve concentrations after an irradiation time of T/n using MENDEL, using as an input the fluxes calculated at Step 2. The value of t is incremented of T/n. • Iteration between Step 1 and Loop 1 until t=T

3. APPLICATION TO THE STUDY OF A REP-LIKE ASSEMBLY AT BEGINNING OF CYCLE In this section, we describe the test case used for the validation of this method and then we analyze preliminary results in terms of figure of merit. 3.1. Test case description The test case is shown in Fig. 1 in a (2-D) cut view and is a model of a REP assembly. It is made up of 17x17 cells containing either moderator (i.e. guide tubes) or UOX fuel with an enrichment of 4.9 %. All cells are wrapped in Zr cladding and surrounded by a H2O borated moderator, all materials being defined at a temperature of 300 K. The neutron source is defined uniformly in the assembly with a Watt energy spectrum. And reflective boundary conditions allow the modeling of an infinite lattice of fuel pins. The fuel pin cells are grouped in three depleted regions. The group of each cell is defined by the nature of the eight nearest neighboring cells. The first group gathers cells which have guide tubes as nearest neighbors (sharing a common side) and the third group contains cells surrounded by eight fuel cells. The remaining pin cells belong to the second group. Those groups are used to take into account the spatial variation of the flux spectrum with a limited number of depleted regions (to limit the computation time).

• Step 0 (pre-calculation): o Step 0-a: Monte-Carlo pre-calculation of all isotopes concentrations at each time step (noted {Ni}(t), where the i index stands for isotopes). o Step 0-b: Isotopic concentrations are optimized. For each time step, if an isotope is not present in a depleted region, it is added in the list of isotopes with an arbitrary concentration value o Step 0-c: time t is fixed to 0 • Step 1: Monte Carlo transport using the optimized isotopic concentrations of the pre-calculation at t+ΔT/2 and recording of the tracks file containing all neutron histories • Loop 1 from t equal t to t+ΔT on following steps: o Step 2: Perturbation calculation using the track file generated at Step 1

Figure 1: UOX-like assembly description. The assembly is made of 17*17 cells, each belonging to one of four different categories: guide tubes made of water (white cells), group 1 (red cells), group 2 (orange cells) or group 3 (blue cells). The higher the group number, the less moderated its pin cell.

And to present the detailed spectral characteristics, 11514 energy groups were used to compute the neutron flux with the collision estimator. In most Monte Carlo burnup codes, the simulation iterates between a parallelized transport step and a mono-processed depletion step. But with this procedure no confidence intervals are provided on isotopic inventories. To overcome this drawback, a new procedure was implemented in the TRIPOLI-4® burnup card, which consists in simultaneously launching independent mono-processed Monte Carlo burnup simulations to compute statistical uncertainties on all tallies. So, in the following results of this paper, we use this capability and all results are provided with statistical uncertainties from 200 independent replica of the depletion. 20 batches of 500 neutrons were simulated using 211 time/burnup steps. This time grid was separated into two parts. The first part contained 11 burnup steps uniformly distributed from 0 to 4 GWd/t and the second part contained the remaining time steps until 14 GWd/t. Therefore the burnup interval between two steps was equal to 400 MWd/t in the first part and 50 MWd/t in the second one. The perturbed scheme is then used only on the second part of the time grid (since one wants to avoid the fast isotopes production phase taking place during the first time steps). The burnup steps structure used for the first part is arbitrary (it has no impact on the comparison between the new method (activated in the second part) and the standard depletion calculation). Therefore large time steps were chosen so as to avoid expensive computation time (the emphasis here is not put on the physics analysis but on the comparison itself).

Figure 2: Collision estimator of the keff. The left axis (blue) gives the variation of the keff as a function of the burnup: the black points are associated to the reference calculation, and the blue points to the perturbed scheme. The right axis (red) gives the difference between the keff calculated by the reference scheme and the perturbed scheme.

3.2. Validation of the perturbed scheme In this section, we begin by testing the validity of our method using a comparison of tallies between the classical burnup scheme (refered to as the reference scheme) and the perturbed one (defined section 2.3). Both schemes require to perform different sets of simulations, containing independent replica of the same simulation to evaluate the variance of the different tallies. The same parameters values were used for both sets (like the number of simulated particles, etc). The variation of the keff (calculated with the collision estimator) is shown Fig. 2 as a function of the burnup between 4 and 14 GWd/t. Results of both simulations are displayed on this figure (with black points for the reference calculation and blue points for the perturbed one) but since results are very close, the differences (in percent) between both schemes are also provided and are shown to never exceed 250 pcm. The same analysis is presented Fig. 3 for isotopic concentrations of different isotopes: 235U, 238U, 239Pu and 241Pu. The first plot shows the variation of isotopic concentrations of both methods as a function of the burnup and the second one displays the difference in percent between those concentrations. Once again, the perturbed results are very close to the reference ones since the differences on the concentrations are never bigger than 0.7%.

Figure 3: Top: concentrations of different isotopes versus burnup calculated by the reference scheme (solid lines) and the perturbed scheme (dashed lines). Bottom: differences between those concentrations.

And to ensure that isotopic concentrations of the perturbed calculation are always close to the reference ones, the maximal value over all burnup steps of the difference between mean isotopic concentrations of both calculation is displayed Fig. 4 for 27 isotopes (those isotopes are selected for their contribution to the the fuel reactivity), exhibiting

differences always smaller than 2% for all concentrations.

Figure 4: Maximum value (for all burnup steps) of the differences between concentrations calculated by the reference scheme and the perturbated scheme, for a selection of isotopes.

For all tallies, the difference of mean values does not increase with the burnup. It fluctuates around zero, what seems to demonstrate a good agreement of the perturbed scheme with the reference one. But the low values of the difference between means results presented on previous figures do not prove that the perturbed scheme is compatible with the reference. Indeed, those measures do not take into account the dispersal of both distributions (done by independent replica, see section 3.1). Thus it is also possible to calculate the t value of the Student t-test of all tallies to determine the compatibility between both schemes. For each tally R, this value is defined as: t[R] =

R !"# − R !"#$% σ!!"# [R] + σ!!"#$% [R]

(6)

where R is the mean value of the tally R and σ2[R] is the variance on the tally R (still keeping in mind that means and variances are calculated by the independent replicas method explained previously). And we consider that the perturbation calculation is compatible with the reference one if the t parameter of all tallies is lower than the quantile of the Student distribution at 99.9%. Since 200 replicas are performed for both simulations, this quantile is equal to 3.111. The maximal value over all burnup steps of this t parameter is noted in the legends of Fig. 2 and 3 for tallies presented on those figures and Fig. 5 provide the same information for the measurements of Fig. 4. Still, all results are in good agreement, except for the isotopic concentrations of 238U. However, a detailed analysis of this tally revealed that its standard deviation is comparable to the numerical error (due to the very low decrease of the 238U concentration).

Figure 5: Maximum value (for all burnup steps) of the t value of the Student t-test to statuate on the compatibility between concentration distributions calculated by the reference scheme and the perturbated scheme, for a selection of isotopes. For the 99.9% quantile, the value of t should not exceed 3.111.

Furthermore, results obtained on other tallies, not presented in the present paper, were checked to be compatible. 3.3. Performance of the method Concerning the performance analysis (which includes the particles simulation time, the tracks/histories storage time, and the perturbation/tallies computation time) the perturbed burnup steps (without recording neutron histories) were measured to be 7 times faster than the reference calculation, while burnup steps with the recording of neutrons histories need 50% overhead. If one defines the maximum speed-up factor that can be obtained on the accumulated computation time (from 4 GWd/t until the last step of the simulation) as: !"#$[!] =

!!"# !!"#

%$(7)

where Tref (respectively Tpertu), then this time gain can not exceed 7 (see Fig. 6 or 7 below). In the meanwhile, the gain in time is somewhere compensated by a loss on the variance of the perturbed scheme, w.r.t. the reference one. Fig. 6 and 7 allow comparing both the gain in time and the degradation of the statistical, since they display the gain on computation time (Eq. 7) and the ratio of the standard deviation on three estimators of keff (Fig. 6) and four isotopic concentrations (Fig. 7) between the perturbation scheme and the reference calculation, as a function of the burnup. The first figure shows that the variance ratio for the keff estimators fluctuate around 1 which means that the variance reduction technique operates very well, while the practical gain in time is of roughly 4.

calculation: !"#$ ! =

Figure 6: Right axis: gain in time versus burnup between the perturbed scheme and the reference scheme (purple curve). Left axis: ratio between standard deviation of the keff between the perturbed and the reference scheme, for different keff estimators.

!"#!"# [!] !!"#$%  ! ! !"#$% [!] = !"#!"#$% [!] !!"#  ! ! !"# [!]

(9)

Fig. 8 and 9, appearing below, display the variation on this speed-up factor on previous keff estimators and isotopic concentrations as a function of burnup steps. As expected, gains on the different estimations of keff fluctuate around the speed-up factor on accumulated computation time (between 3 and 5) because their variances are not deteriorated. In the case of isotopic inventories, the results are strongly dependent on the isotopes considered. An analysis of mean trends seem nevertheless to indicate that during the first perturbed burnup steps, the FOM increase because variances stay close to the reference value while the accumulated gain in time increases quickly. But after some burnup steps, the growing statistical uncertainties of perturbed tallies have an impact on the speed-up factor (see Fig. 7 and 9) and for some isotopic concentrations this effect takes over the time gain. However, this increase of the perturbed variances stays contained and still allow to speed up the

Figure 7: Right axis: gain in time versus burnup between the perturbated scheme and the reference scheme (purple curve). Left axis: ratio between standard deviation on different isotopes concentrations between the perturbated and the reference scheme.

The same results were observed (although not presented here) for all other tallies computed from the transport solver, like fluxes or power tallies. However, the second figure shows that standard deviation on perturbed isotopic concentrations can be larger than reference ones. The variances on those tallies are not degraded at an exponential rate (this is the case when the variance reduction technique is not used) but tend to get bigger than the reference ones for most of isotopic concentrations.

Figure 8: Gain on the keff versus burnup (Eq. 9), for different keff estimators.

So, to take into account the deterioration of the variance in the overall speed up factor, one has to measure the figure of merit (FOM), defined by: !"#[!] =

1 !! ! [!]

(8)

where T is the transport time computation accumulated from 4 GWd/t of burnup and σ2[R] is the variance on the tally R. And the speed-up factor obtained on a tally noted R, noted Gain[R], is now defined as the ratio on the FOM on R between the perturbation scheme and the reference

Figure 9: Gain on the isotopic concentrations versus burnup (Eq. 9), for different isotopes.

estimation of the majority of isotopic concentrations by a factor of 2 at least.

The difference in terms of qualitative behavior of variances estimated by this new depletion scheme between tallies computed by the transport solver (i.e. flux, keff, etc) and tallies computed by the depletion solver (i.e. isotopic concentrations) arises from the variation of the variance of the reference calculation. Indeed the variances of the first tallies remain constant over the different burnup steps, while the seconds are estimated with continuously varying variances over the burnup steps (either increasing or eventually decreasing for some isotopic concentrations estimates). In this last case, since for each interval ΔT where a specific tracks file is used, the variance of isotopic concentrations in the perturbed scheme can only be at least equal to the variance of isotopic concentrations in the middle of the interval ΔT of the reference calculation (see section 2.3), then the variation of the variances in the reference calculation mandatory implies an increase of the variance on isotopic concentrations in the perturbed scheme. This small increase is then propagated and accumulated from each ΔT interval to the other, leading to potentially important increase of variances at the end of the calculation.

4. OUTCOME AND PERSPECTIVES In the present paper, we have shown that the use of a scheme based on concentration perturbation could greatly improve Monte Carlo burnup computation time. Implemented in TRIPOLI-4®, this method allowed improvement in the FOM going up to 5 in the simple case of a PWR assembly. Several test of this method with diverse geometries revealed that the gain on the computation time is linked to the complexity of the geometry: complex geometries require consequent tracking time of the Monte-Carlo codes that can be drastically reduced by the use of perturbation techniques. This leaves hope for a gain of an order of magnitude in the FOM when this method will be applied on complex geometries like the depletion study of a full reactor core. Concerning improvements of the method as it is defined in the present paper, three directions are currently investigated. First, a study of the bias due to the number of particles [12,13] has to be led, to ensure that this new method present no additional penalty compared to standard depletion schemes. Second, a profiling of the code revealed that the use of a universal energy grid for the cross sections [14, 15] could greatly enhance the overall gain on the FOM since the calculation time of the perturbed scheme is mainly due to the computation of perturbed macroscopic cross sections. Finally, optimizations on the variance reduction technique presented here are currently investigated; indeed, a better selection of isotopic inventories during the pre-calculation step has a strong impact on the variance estimation of the different isotopic concentrations.

ACKNOWLEGMENTS TRIPOLI-4® is registered trademarks of CEA. We gratefully

acknowledge EDF and AREVA supports.

References 1) M. Moriwaki, M. Aoyama, “Improvement of Monte Carlo Lattice Burnup Calculation Performance with the Correlated Sampling Method,” Journal of Nuclear Science and Technology, 39[6], 587-596 (2002) 2) M. Moriwaki, “Multi-Assembly Analysis with an Advanced Correlated Sampling Method,” Journal of Nuclear Science and Technology, 40[11], 905-917 (2003) 3) H. Rief, “Generalized Monte Carlo Perturbation Algorithms for Correlated Sampling and a Second-Order Taylor Series Approach,” Annals of Nuclear Energy, 11, 455-476 (1984) 4) CEA Seminar on TRIPOLI-4 code, M&C 2005 Avignon (CD-ROM), organized by NEA/OCDE (2005) 5) G. Dejongue, “Etudes d’Effets Différentiels par la Méthode de Monte Carlo dans le Cadre de l’Equation du Transport. Applications aux Calculs de Protection et de Neutronique,” Ph.D. Thesis, Université Paris XI (1982) 6) E. Brun and E. Dumonteil, “Guide d’Utilisation de TRIPOLI-4 Evoluant Version 1.1,” CEA Internal Report DEN/DM2S/SERMA/LTSD/RT/12-5307/A (2012) 7) R. Reyes-Ramìrez, C. Martìn-del-Campo, J.-L. François, E. Brun, E. Dumonteil and F. Malvagi, “Comparison of MCNPX-C90 and TRIPOLI-4-D for fuel depletion calculations of a Gas-cooled Fast Reactor,” Annals of Nuclear Energy, 37, 1101-1106 (2010) 8) C. M. Diop, E. Dumonteil, F. X. Hugot, C. Jouanne, Y. K. Lee, F. Malvagi, A. Mazzolo, O. Petit and J. C. Trama, “An Overview on the Monte Carlo Particle Transport Code TRIPOLI-4,” Transaction of the American Nuclear Society, 97, 694-695 (2007) 9) P. Bellier, H. Mao and A. Tsilanizara, “Premiers elements de vérification de MENDEL V0 par rapport aux outils actuels DARWIN2 et APPOLLO2,” CEA Internal Report DEN/DANS/DM2S/SERMA/LLPR/RT/12-5422/A (2012) 10) I. Antcheva, M. Ballintijn, B. Bellenot et al., “ROOT – A C++ framework for petabyte data storage, statistical analysis and visualization,” Computer Physics Communications, 180, 2499-2512 (2009) 11) M. Goto, “C++ interpreter – cint,” CQ publishing (1997) 12) E. Dumonteil and C. M. Diop, “Biases and Statistical Errors in Monte Carlo Burnup Calculations: An Unbiased Stochastic Scheme to Solve Boltzmann/Bateman Coupled Equation,” Nuclear Science and Engineering, 167, 165-170 (2011) 13) E. Brun, E. Dumonteil and F. Malvagi, “Systematic Uncertainty Due to Statistics in Monte Carlo Burnup Codes: Application to a Simple Benchmark with TRIPOLI-4-D,” Progress in Nuclear Science and Technology, 2, 879 -885 (2011) 14) W. Haeck, “An Optimal Approach to Monte Carlo

Burnu-up,” PhD Thesis, Ghent University (2007) 15) W. Haeck and B. Verboomen, “An Optimal Approach to Monte Carlo Burn-up,” Nuclear Science and Engineering, 156[2], 180-196 (2007)

ANNEXE B. PUBLICATIONS PRINCIPALES

E. Dumonteil

196

Université Paris-Sud