Physics-Informed Machine Learning for Nuclear Core Simulations

Sammanfattning

High-fidelity nuclear core simulators are essential tools for reactor analysis and fuel management, but their computational cost limits their use in workflows requiring many repeated evaluations, such as screening and optimization of core designs. This thesis investigates the extent to which machine-learning surrogate models can approximate selected outputs of a high-fidelity core simulator for boiling water reactor applications and how known physics can be incorporated to improve robustness. A modular surrogate-model framework was developed to predict burnup, void fraction, two-group neutron flux, and power distribution on a three-dimensional nodal core grid over an operational cycle. The models were primarily developed as convolutional neural networks, using reactor-state quantities available before running the reference simulator as input. In addition to data-driven training, selected physics-informed loss terms were investigated, including a monotonicity constraint for void fraction and neutronic loss terms based on diffusion-equation residuals and global neutron balance. The results show that surrogate models can reproduce several spatially distributed reactor quantities with useful accuracy at substantially reduced computational cost. For the neutronic flux and power models, the 90th percentile of the absolute relative error was approximately 4–5% on the test dataset. The surrogate pipeline achieved a total wall-time speedup of approximately 350 times compared with the reference simulator, while the pure neural-network inference time corresponded to a speedup of approximately 16,000 times. The physics-informed monotonicity loss significantly improved the void-fraction model, whereas the neutronic loss terms did not lead to a clear improvement in final predictive accuracy. The findings indicate that surrogate models can be useful as fast approximations of selected core-simulator outputs, particularly for exploratory analysis and design screening.