New PINN Method Accelerates Nuclear Reactor Simulation

Accurately determining the effective multiplication factor (keff) and neutron flux distribution is a critical, computationally intensive task in nuclear reactor core analysis. Physics-Informed Neural Networks (PINNs) have emerged as a promising approach, embedding neutron diffusion equations and boundary conditions directly into their training process to solve these problems. To further enhance the efficiency of keff calculations using PINNs, a new method called Joint Initialization Physics-Informed Neural Network (JI-PINN) has been developed. This approach utilizes a low-resolution approximate solution to the K-eigenvalue problem to construct a joint initial state for both the flux network parameters and keff. Both are then simultaneously optimized under physical constraints. The JI-PINN method was validated across several test cases, including 2D and 3D scenarios with varying materials. It consistently reduced total computational time by 25.4% to 49.4% while preserving comparable solution accuracy. Additionally, the method reduced the occurrence of anomalous results where keff significantly deviated from reference values, indicating improved robustness.