Answers are grounded in this paper and cite their sources.
BackgroundThe accurate solution of the neutron diffusion equation to determine the effective multiplication factor (keff) is fundamental to nuclear reactor design and safety analysis. However, the performance of a standard single-network Physics-Informed Neural Networks (PINN) is often limited when applied to practical multi-material problems, exhibiting significant accuracy degradation, particularly at the interfaces between disparate materials where steep neutron flux gradients arise.PurposeThis study aims to propose an algorithm for solving the neutron diffusion equation based on the PINN source iteration method and introduces multiple optimization strategies to address the limitations in complex multi-region problems, thereby achieving high-accuracy and robust solutions to the neutron diffusion eigenvalue problem.MethodsFirst, the correctness of the source iteration algorithm was verified on a standard single-region homogeneous example. For multi-region problems with relatively small differences in material cross-sections, two optimization methods, Residual Adaptive Resampling (RAR) and Weight Adaptive, were employed to enhance the expressive capacity of the model. Furthermore, for multi-region problems with steep neutron flux gradients at region interfaces, which made these optimization methods ineffective, the Extended Physics-Informed Neural Network (XPINN) framework was applied and validated using the TWIGL benchmark.ResultsThe results show that, for multi-region problems with relatively small cross-section differences, the unoptimized PINN produces a keff prediction error of 9.89×10⁻³. Both RAR and Weight Adaptive optimization strategies improve prediction accuracy and convergence performance, with the Weight Adaptive method being the most effective, reducing the error to 6.3×10⁻⁴. In contrast, for the TWIGL benchmark problem involving steep neutron flux gradients at material interfaces, the single-network PINN and its associated optimizations fail, yielding errors exceeding 1×10⁻². With the application of the XPINN framework, however, the keff prediction error is reduced to just 2.5×10⁻⁴, representing more than an order-of-magnitude improvement and exhibiting excellent convergence stability.ConclusionsThe optimization framework for solving neutron diffusion equation proposed in this study substantially enhances the performance of standard PINN algorithm, exhibiting notable advantages in prediction accuracy and training stability, and thereby providing a more effective solution to such challenges.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: