基于混合神经网络的非定常流场预测方法  

An unsteady flow-field prediction method based on hybrid neural network

作　　者：孔德天 董义道 张来平 邓小刚 KONG De-Tian;DONG Yi-Dao;ZHANG Lai-Ping;DENG Xiao-Gang(College of Computer Science,Sichuan University,Chengdu 610065,China;Tianfu Engineering-oriented Numerical Simulation&Software Innovation Center,Sichuan University,Chengdu 610065,China;College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China;Institute of Defense Science and Technology Innovation,Academy of Military Sciences,Beijing 100071,China;Institute of Systems Engineering,Academy of Military Sciences,Beijing 100082,China)

机构地区：[1]四川大学计算机学院,成都610065 [2]四川大学天府工程数值模拟与软件创新中心,成都610065 [3]国防科技大学空天科学学院,长沙410073 [4]军事科学院国防科技创新研究院,北京100071 [5]军事科学院系统工程研究院,北京100082

出　　处：《四川大学学报（自然科学版）》2024年第6期135-143,共9页Journal of Sichuan University(Natural Science Edition)

基　　金：国家重大专项(GJXM92579);四川省科技计划资助(2023YFG0158);国防科技大学科研计划项目(ZK21-08)。

摘　　要：在众多复杂物理系统的模拟中,基于网格离散化求解偏微分方程是一项关键而耗时的任务.为了克服这一挑战,本文提出了一种创新的复合神经网络,称为GRNet,它结合了图神经网络(GNN)和循环神经网络(RNN).GNN模型被训练以学习由Navier-Stokes方程控制的网格节点之间的物理规律.RNN网络被训练以揭示网格节点的时间依赖性.本文所提出的模型可以有效地利用高分辨率网格的多尺度优势,仅需少量的起始帧,便能快速精确地预测后续流场.我们广泛探究了GRNet在多个复杂的多尺度流场预测任务中的性能,例如圆柱和翼型.与传统的数值模拟结果相比,我们的模型不仅保持了出色的准确性,而且运行速度令人印象深刻.与基准模型(GN)相比,GRNet在最小化累积预测误差方面表现出明显优势.Solving partial differential equations based on mesh discretization is a critical and time-consuming task in the simulation of numerous complex physical systems.To overcome this challenge,this paper pro⁃poses an innovative hybrid neural network,called GRNet,which combines graph neural network(GNN)and recurrent neural network(RNN).The GNN model is trained to learn the law of physics between mesh nodes governed by the Navier-Stokes equations.The RNN network is trained to uncover the temporal depen⁃dence of mesh nodes.The proposed model can effectively utilize the multi-scale advantages of high-resolution meshes to quickly and accurately predict subsequent flow fields based on just a few starting frames.We exten⁃sively explore the performance of GRNet in several complex flow field prediction tasks,such as cylinders and airfoils.Compared with conventional numerical simulation results,our model not only maintains exceptional accuracy but also operates at an impressively high speed.In contrast to the baseline model(GN),GRNet ex⁃hibits a substantial advantage in minimizing cumulative prediction errors.

关 键 词：混合神经网络 图神经网络 循环神经网络 流场预测 非定常 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]