基于深度神经网络的高阶非线性激波判别式  

Higher-order nonlinear shock wave discriminant based on deep neural network

作　　者：杨华 陈树生 李猛 庞超 高正红[1] 向星皓 YANG Hua;CHEN Shusheng;LI Meng;PANG Chao;GAO Zhenghong;XIANG Xinghao(School of Aeronautics,Northwestern Polytechnical University,Xi’an 710072,China;AVIC The First Aircraft Institute,Xi’an 710089,China;State Key Laboratory of Aerodynamics,Mianyang 621000,China)

机构地区：[1]西北工业大学航空学院,西安710072 [2]中航工业第一飞机设计研究院,西安710089 [3]空气动力学国家重点实验室,绵阳621000

出　　处：《空气动力学学报》2023年第7期56-63,I0001,共9页Acta Aerodynamica Sinica

基　　金：中国科协青年人才托举工程(2022QNRC001)。

摘　　要：超声速流动数值模拟中准确识别激波位置对求解飞行器气动特性十分关键,传统的人工构造激波识别方法严重依赖于经验参数,在实际应用中存在一定局限性。为了建立更加通用的激波识别方法,采用深度神经网络构建了高阶非线性激波判别式。在生成了训练数据集与测试数据集的基础上,构建深度前馈神经网络并训练该模型使其满足激波识别预测精度,并通过一维Lax、Sod激波问题及二维超声速无黏圆柱绕流算例验证了该判别式的准确性。结果表明,本文构建的高阶非线性激波判别式相较于人工构造的激波识别器能更加精准地判别光滑区域和激波区域,在不同来流马赫数、不同网格分布条件下均能准确识别出激波位置,但激波识别精度依赖于网格分辨率。In the numerical simulation of supersonic flow,accurately identifying shock wave position is essential to solve an aircraft's aerodynamic and thermal characteristics.The traditional artificial shock identification method seriously depends on empirical coefficients,which has some limitations in practical application.A higher-order nonlinear shock wave discriminant is constructed by using a deep neural network to establish a more general shock identification method.Based on the training data set and the test data set,the depth Feedforward Neural Network is constructed,and the model is trained to meet the prediction accuracy of shock wave identification;it is applied to one-dimensional Lax and Sod shock problems and two-dimensional supersonic inviscid circular cylinder to verify the accuracy of the discriminant.The results show that this paper's high-order nonlinear shock discriminant can more accurately distinguish the smooth region and shock region than the artificial shock discriminant.Secondly,the shock discriminant can accurately identify the shock position under different Mach numbers and grid distributions,but the accuracy of shock identification depends on the grid resolution.

关 键 词：激波判别式 神经网络 超声速 计算流体力学 机器学习 

分 类 号：V211.3[航空宇航科学与技术—航空宇航推进理论与工程]