基于神经网络的CFD粗网格模拟优化方法研究  

Optimization method of CFD coarse grid numerical simulation based on neural network

作　　者：靳爽 刘晓晶[1] 程旭[1] JIN Shuang;LIU Xiaojing;CHENG Xu(School of Nuclear Science and Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学核科学与工程学院,上海200240

出　　处：《核技术》2021年第6期75-81,共7页Nuclear Techniques

基　　金：国家自然科学基金(No.11922505)资助。

摘　　要：为缓解使用计算流体力学(Computational Fluid Dynamics,CFD)程序进行反应堆热工水力数值模拟中计算效率和计算精度之间的突出矛盾,以实现在较高计算效率下得到较高精度的计算结果,研究借助机器学习(Machine Learning,ML)中的前馈神经网络(Feedforward Neural Network,FNN)方法,通过对一定数量物理过程类似但具体参数不同的工况在粗网格、细网格两套网格下数值模拟计算结果差异的对比,得到针对同类工况具有推广适用性的误差函数表达式,对粗网格的数值模拟计算结果进行优化。通过比较优化前后相关物理量的计算结果与细网格下计算结果之间的均方根误差(Root Mean Square Error,RMSE),对优化效果进行评价。结果表明:粗网格计算结果在经由前馈神经网络得到的误差函数优化后,与细网格相关物理量的均方根误差显著降低。因此,研究建立的基于神经网络的粗网格数值模拟优化技术,可以有效提高反应堆粗网格数值模拟的计算精度,为在较高计算效率下实现较高精度的数值模拟提供了方法借鉴。[Background]There is a prominent contradiction between calculation efficiency and calculation accuracy when computational fluid dynamics(CFD)program is used in the thermal-hydraulic numerical simulation of reactors.[Purpose]This study aims to alleviate the contradiction to achieve higher calculation results with higher calculation efficiency by finding a method to effectively optimize the results of the coarse-grid numerical simulation.[Methods]First of all,the feedforward neural network(FNN)machine learning(ML)method was employed to compare two sets of numerical simulation results under a certain number of similar working conditions under coarse and fine grids.Then,an error function expression with general applicability for similar conditions was obtained to optimize the numerical simulation calculation results of the coarse grid.Finally,the optimization effect was evaluated by comparing the root mean square error(RMSE)between the calculation results of relevant physical quantities before and after optimization and the calculation results under fine grid.[Results]The research results show that after the coarse grid calculation result is optimized by the error function obtained by the FNN method,the RMSE of the physical quantity related to the fine grid is significantly reduced.[Conclusions]The coarse-grid numerical simulation optimization technology based on neural network established in this study can effectively improve the calculation accuracy of the reactor’s coarse-grid numerical simulation,and provide a method reference for realizing highprecision numerical simulation with high computational efficiency.

关 键 词：神经网络 粗网格 数值模拟 优化 

分 类 号：TL334[核科学技术—核技术及应用]