基于流场反演和图神经网络的翼型分离流动预测方法  

An airfoil separation flow prediction method based on field inversion and graph neural network

作　　者：邹远洋 董义道 张来平 邓小刚 ZOU Yuan-Yang;DONG Yi-Dao;ZHANG Lai-Ping;DENG Xiao-Gang(College of Computer Science,Sichuan University,Chengdu 610065,China;National Key Laboratory of Fundemental Algorithms and Models for Engineering Simulation,Chengdu 610207,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]工程数值模拟基础算法与模型全国重点实验室,成都610207 [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)》2025年第2期359-368,共10页Journal of Sichuan University(Natural Science Edition)

基　　金：国家重大专项(GJXM92579);国防科技大学科研计划项目(ZK21-08)。

摘　　要：传统的RANS模型采用布辛涅斯克近似,假设湍流雷诺应力和平均速度梯度张量之间呈线性关系,这一假设适用于简单的剪切流动,但很难推广应用于复杂分离流动问题.本文基于流场反演和机器学习FIML方法框架,针对目前该方法框架内普遍采用的多层感知机网络对于湍流空间相关性表征不足的缺陷,通过图神经网络的引入,对工程应用较为广泛的SA一方程湍流模型生成项进行了修正.在此基础上,结合流场分离特征设计了一种加权函数,改进了图神经网络的消息传递机制.针对大攻角、高雷诺数S809翼型分离流动的实验结果表明,与现有的多层感知机网络相比,图神经网络在不同攻角、不同网格上预测得到的升力系数和实验值更加接近,且新的消息传递机制能够进一步提升图神经网络预测精度.The traditional Reynolds-Averaged Navier-Stokes(RANS)model uses the Boussinesq approximation and assumes a linear relationship between the turbulent Reynolds stress and the average velocity gradient tensor.While this assumption applies to simple shear flows,its extension to intricate separated flow scenarios poses challenges.To address the limitation of the multilayer perceptron network commonly utilized in the Field Inversion and Machine Learning(FIML)framework for representing turbulence spatial correlation,this paper proposes a graph neural network to modify the production term of the Spalart-Allmaras oneequation turbulence model.Furthermore,a weighted function is devised utilizing the characteristics of flow field separation to augment the message passing mechanism of the graph neural network.The experimental results for separated flow over the S809 airfoil at high angles of attack and high Reynolds numbers indicate that,compared with the existing multilayer perceptron networks,the utilization of graph neural network results in lift coefficient predictions that closely align with experimental values at various angles of attack and grid configurations.Moreover,the incorporation of a novel message passing mechanism in the graph neural network enhances its predictive accuracy.

关 键 词：湍流建模 流场反演 图神经网络 计算流体力学 数值模拟 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]