摘要
针对融合了物理控制方程,尤为适用于物理场预测的新兴神经网络方法——物理神经网络(PINN),开展深入的文献调研,形成对面向流体力学的物理神经网络方法发展趋势的研判。首先,对神经网络融合物理信息的思路进行溯源;其次,介绍当前物理神经网络基本架构,针对全连接型物理神经网络,从间断问题的高精度预测研究、偏微分方程(PDE)植入形式、流场重建问题、损失函数形式、多精度数据及多尺度问题以及训练控制等方面进行文献综述;再次,对于基于卷积神经网络(CNN)和其他新兴网络架构的物理神经网络进行文献梳理;最后,形成面向流体力学的物理神经网络发展趋势与思考。通过对2017年至2023年间近百篇文献的研究及相关数值实验可知,针对强间断的高分辨率预测是面向高速流动问题的物理神经网络研究中需要解决的重要问题;基于全连接网络的物理神经网络拥有无网格化的优势,可用于各类流动问题的求解;基于卷积网络的物理神经网络具备与已有传统数值方法深度融合的优势,可有效利用已有的流场图像、物理量云图等结构化数据,进行复杂流动问题的求解。
An in-depth literature review was carried out for the new Physics-Informed Neural Network(PINN)method,which combines physical control equations and is especially suitable for physical field prediction,to acquire the knowledge on the development trend of PINN methods oriented to fluid dynamics.Firstly,origin of the idea of informing the neural network of physical information was investigated.Secondly,the basic architecture of the current physical neural network was introduced.And review of literatures on fully connected PINN was conducted in perspectives ranging from the high-precision prediction of discontinuities,the implantation form of the controlling Partial Differential Equations(PDE),the flow field reconstruction problem,the form of loss function,multi-precision data and multi-scale problems,to network training control.Again,literatures on PINNs based on Convolutional Neural Network(CNN)and other emerging network architectures were also examined.Finally,the development trend and thinkings of PINN for fluid dynamics were formed.Based on the research of nearly a hundred articles from 2017 to 2023 and related numerical experiments,it can be seen that high-resolution prediction for strong discontinuities is an important topic in the research of PINNs for high-speed flow problems;PINN has the advantage of being mesh free and can be used to solve various flow problems;PINNs based on convolutional network architecture has the advantage of deep integration with existing numerical methods,and can effectively use existing flow field images,contours and other structured data to solve complex flow problems.
作者
田松岩
黄鑫格
段焰辉
陈洪波
陈文秀
TIAN Songyan;HUANG Xinge;DUAN Yanhui;CHEN Hongbo;CHEN Wenxiu(School of Systems Science and Engineering,Sun Yat-sen University,Guangzhou Guangdong 510275,China;Department of Mechanics and Aerospace Engineering,Southern University of Science and Technology,Shenzhen Guangdong 518055,China)
出处
《计算机应用》
CSCD
北大核心
2024年第S01期133-141,共9页
journal of Computer Applications
关键词
流场预测
物理神经网络
损失函数
偏微分方程
间断问题
flow field prediction
Physics-Informed Neural Network(PINN)
loss function
Partial Differential Equation(PDE)
discontinuity problem
