Amélioration des méthodes de calcul de cœurs de réacteurs nucléaires dans APOLLO3 : décomposition de domaine en théorie du transport pour des géométries 2D et 3D avec une accélération non linéaire par la diffusion

摘要

This thesis is devoted to the implementation of a domain decomposition method applied to the neutron transport equation. The objective of this work is to access high-fidelity deterministic solutions to properly handle heterogeneities located in nuclear reactor cores, for problems’ size ranging from colorsets of assemblies to large reactor cores configurations in 2D and 3D. The innovative algorithm developed during the thesis intends to optimize the use of parallelism and memory. The approach also aims to minimize the influence of the parallel implementation on the performances. These goals match the needs of APOLLO3 project, developed at CEA and supported by EDF and AREVA, which must be a portable code (no optimization on a specific architecture) in order to achieve best estimate modeling with resources ranging from personal computer to compute cluster available for engineers analyses. The proposed algorithm is a Parallel Multigroup-Block Jacobi one. Each subdomain is considered as a multi-group fixed-source problem with volume-sources (fission) and surface-sources (interface flux between the subdomains). The multi-group problem is solved in each subdomain and a single communication of the interface flux is required at each power iteration. The spectral radius of the resolution algorithm is made similar to the one of a classical resolution algorithm with a nonlinear diffusion acceleration method: the well-known Coarse Mesh Finite Difference. In this way an ideal scalability is achievable when the calculation is parallelized. The memory organization, taking advantage of shared memory parallelism, optimizes the resources by avoiding redundant copies of the data shared between the subdomains. Distributed memory architectures are made available by a hybrid parallel method that combines both paradigms of shared memory parallelism and distributed memory parallelism. For large problems, these architectures provide a greater number of processors and the amount of memory required for high-fidelity modeling. Thus, we have completed several modeling exercises to demonstrate the potential of the method: 2D full core calculation of a large pressurized water reactor and 3D colorsets of assemblies taking into account the constraints of space and energy discretization expected for high-fidelity modeling.