Response Matrix-based Artificial Neural Network of Nuclear Fissile Systems

Sammanfattning

Advanced nuclear reactor analysis requires computational solvers that are both fast and physically reliable. While High-Fidelity Monte Carlo simulations are accurate, their computational cost is prohibitive for many applications. This thesis develops a Physics-Informed Neural Network (PINN) as an ultrafast surrogate model for the Response Matrix Method (RMM), bypassing the Monte Carlo bottleneck. A dataset of local subresponse matrices was generated using OpenMC for a 17x17 Pressurized Water Reactor (PWR) fuel assembly across various operational states. The PINN architecture, based on a Multi-Head ResNet, enforces neutronic and thermodynamic constraints such as neutron conservation and negative Doppler feedback. The model was benchmarked against an Unconstrained Deep Feed Forward Neural Network (FFNN) and a Random Forest (RF) regressor within a 3D coupled mini-core solver. Results show that the PINN accelerates global spatial evaluations by a factor exceeding $10^5$, reducing inference time from minutes to approximately one millisecond. Error decomposition analysis revealed that the Random Forest achieved artificial accuracy through unphysical error cancellation. In contrast, the PINN demonstrated superior robustness, maintaining a 1.47\% MAPE in spatial power accuracy while strictly honoring physical limits. This work validates PINNs as a safe and efficient foundation for reactor surrogate modeling.