现代应用物理2025，Vol.16Issue(1)：144-150,7.DOI:10.12061/j.issn.2095-6223.202412031
深度学习方法求解反应堆物理中子燃耗方程
Solving Neutron Burnup Equations in Reactor Physics Via Deep Learning
摘要
Abstract
To address the vanishing gradient problem caused by extremely small values such as nuclide disappearance rates,reaction rates,and decay rates in neutron burnup equations,adaptive optimization strategies including time scale transformation,equation equivalence transformation,and weight initialization are introduced.The deep learning method is utilized to solve the neutron burnup equations for five nuclides(U-235,U-236,U-237,U-238,and U-239)in reactor physics.During the training process,through time-scale transformation,the physical quantities are appropriately scaled to alleviate numerical instability caused by extremely small values;equation equivalence transformation is applied to enhance the stability of the computation;and adaptive weight initialization is introduced to improve the training efficiency and convergence of the network.Hyperparameters such as the number of training iterations,the weight of physical constraint loss,and the weight of initial condition loss are adjusted,and a fully connected neural network architecture is used for numerical validation.The numerical validation results show that the adaptive optimization strategies effectively solve the vanishing gradient problem caused by extremely small values in the neutron burnup equations,demonstrating the effectiveness and adaptability of physics-informed neural networks(PINN)combined with these strategies in solving multi-nuclide burnup equations.关键词
深度学习/物理信息神经网络/中子燃耗方程/梯度消失/适应性优化策略Key words
deep learning/physics-informed neural network(PINN)/neutron burnup equation/vanishing gradient/adaptive optimization strategies分类
能源科技引用本文复制引用
向迪,郭凤晨,刘东,潘俊杰,江勇,郑鹏德,张斌斌..深度学习方法求解反应堆物理中子燃耗方程[J].现代应用物理,2025,16(1):144-150,7.基金项目
四川省重大科技专项项目(2024ZDZX0006) （2024ZDZX0006）
四川省揭榜挂帅项目(2023YFG0373) （2023YFG0373）
稳定支持基础科研项目(WDZC-2023-05-03-05) （WDZC-2023-05-03-05）
中核核能软件与数字化反应堆工程技术研究中心项目(Z404202301002) （Z404202301002）
