GPU-Optimized CMFD-Based Transient Fixed Source Solution for STREAM-3D

Abstract

Accurate transient reactor analysis requires solving the time-dependent neutron transport equation, which provides both temporal and spatial fidelity. However, modelling complex heterogeneous reactor geometries imposes substantial computational cost on memory and execution time. With the advancement of Graphics Processing Unit (GPU) technology, offloading the most computationally intensive portions of reactor physics codes to GPUs has become one of the most effective acceleration strategies for whole-core transport simulations. This research develops a GPU-Optimized time-dependent multi-group Coarse Mesh Finite Difference (CMFD) solver based on the Transient Fixed Source Problem (TFSP) formulation. It solves the neutron diffusion equation, with correction factors and homogenized cross-sections, which dynamically updated from the underlying transport solution. The approach leverages the assumption that reactor conditions remain constant during each transient time step, enabling efficient use of the CMFD-based TFSP (CTFSP) technique. The MOC/DD solver is further employed to calculate CMFD cell cross sections and correction factors for GPU-enabled CTFSP. Subsequently, the 3D CMFD-transient solver computes the global neutron flux distribution, significantly reducing computational cost while preserving physical accuracy. Upon completion of the GPU enabled CTFSP, the transport solution is iteratively refined for improved consistency. The STREAM transport code, developed at UNIST, is coupled with three thermal- hydraulic solvers for multi-physics feedback analysis. In this work, the GPU-optimized CTFSP performance is examined both with and without multi-physics coupling. Several benchmark problems are solved to validate the proposed methodology, confirming that the approach maintains solution accuracy and stability. The results demonstrate that smaller time steps enhance transient accuracy, though at the cost of increased computational time and memory. The GPU-enabled CMFD-transient solver effectively mitigates this trade-off, offering significant performance improvements. The study also evaluates parallel scalability, and thermal-hydraulic feedback effects, with results compared against other deterministic solvers and experimental data. Finally, a rod-ejection accident in a typical pressurized water reactor (PWR) is analyzed to illustrate the solver’s capability for large-scale reactors (SPERT III, Watts bar benchmark, APR-1400, and OPR-1000), high-fidelity transient simulations, fulfilling the primary objective of this research. KEYWORDS: Transient analysis, CTFSP, Multi-physics, Rod Insertion Analysis, GPU.