This thesis aims to develop and assess GPU-based deterministic neutronics solvers for unstructured meshes, enabling efficient and accurate multigroup neutron transport analysis of advanced nuclear reactors. The increasing complexity of advanced reacto...
This thesis aims to develop and assess GPU-based deterministic neutronics solvers for unstructured meshes, enabling efficient and accurate multigroup neutron transport analysis of advanced nuclear reactors. The increasing complexity of advanced reactor designs, including fast reactors, molten salt systems, and microreactors, necessitates transport solvers that can handle irregular geometries and strong material heterogeneities while exploiting modern heterogeneous computing platforms. Traditional methods, which rely on structured meshes or simplified physics, often fall short in capturing such complexities. To address these challenges, a unified computational code, NuDEAL (Neutronics using Deterministic Finite Element Algorithm), has been developed. NuDEAL supports deterministic transport methods that are geometry-agnostic, multigroup-compatible, and GPU-accelerated. The framework is designed to accommodate tightly coupled multiphysics simulations, and its architecture allows seamless integration with external physics solvers.
NuDEAL incorporates three deterministic solvers: The Method of Characteristics with axial Discontinuous Galerkin (DGMOC), the Discrete Finite Element Method with SN angular discretization (DFEM-SN), and a hybrid approach combining planar MOC with a 3D Hybrid Finite Element Method (HFEM). All methods adopt finite element discretization in space and leverage a common set of multigroup cross sections generated by the GPU-based continuous-energy Monte Carlo code PRAGMA. The DGMOC method introduces a finite element expansion in the axial direction, enhancing accuracy while preserving radial MOC structure. The DFEM-SN method is designed for unstructured geometries and demonstrates robust angular convergence. Both DGMOC and DFEM-SN are GPU-parallelized formulations of methodologies originally developed for CPU-based production codes, which have been adapted here for high-performance execution on modern heterogeneous architectures. In contrast, the MOC/HFEM approach represents a novel methodology proposed in this work, in which a planar MOC solver is coupled to a 3D HFEM framework operating on coarse meshes, with GPU acceleration implemented from the ground up. The HFEM method, based on the variational nodal method, is well-suited for coarse meshes, allowing for computationally efficient 3D calculations. Consistency between MOC and HFEM is not achieved through a closure term but is mediated by CMFD, which ensures equivalence in scalar flux and reaction rates between the fine and coarse domains. All the proposed deterministic methodologies are inherently developed for GPU execution, rather than being ported from existing CPU-based algorithms. Each implementation is tailored with dedicated GPU optimization techniques, including fine-grained parallelism over spatial domains, energy groups, and angular quadrature sets, to exploit the massive parallelism offered by GPU hardware fully.
The proposed methods are validated against both classical and realistic reactor problems. Numerical simulation results on the C5G7 core problem demonstrate the numerical accuracy and computational efficiency of each solver. Applications to advanced reactor models, such as the ABTR sodium fast reactor, the Empire heat pipe microreactor, and the MSRE molten salt reactor, highlight their practical effectiveness. The validation results confirm that key physical quantities, including eigenvalue, power distribution, and scalar flux profiles, are in strong agreement with multigroup Monte Carlo reference solutions. Each solver has distinct advantages and disadvantages. The MOC-based solvers, namely MOC/HFEM and DGMOC, are inherently formulated for extruded reactor geometries. While DGMOC exhibits improved convergence stability and delivers accurate solutions if sufficiently fine axial discretizations are assumed, MOC/HFEM achieves significantly reduced computation times by leveraging a nodal formulation with coarse spatial discretization. In contrast, the DFEM-SN method offers the highest geometric generality and delivers the most spatially detailed solution. However, the absence of geometric simplifications and the use of a sequential solution algorithm result in substantially higher computational cost and memory pressure, with limited GPU parallel scalability.
The framework is further validated through coupled multiphysics simulations of the KRUSTY experiment, in which temperature-dependent cross sections and flux mapping are handled via volumetric projection algorithms and tightly integrated with OpenFOAM and ANLHTP. The multiphysics coupling is realized through a manager-worker architecture based on MPI dynamic process management. A custom-developed mesh intersection tool is employed to compute tetrahedron-level overlap volumes, which enables accurate scalar field transfer between physics domains. These results confirm that NuDEAL delivers a flexible, accurate, and scalable neutronics solution framework suitable for designing and analyzing advanced reactors on heterogeneous computing platforms. The demonstrated capability to perform high-fidelity, cross-code coupled simulations marks a step forward in practical deterministic analysis for advanced reactor systems on heterogeneous computing platforms.