嵌入物理约束的可压缩多介质流神经网络模型  

Neural Network Models of Compressible Multi-Medium Flows Embedded with Physical Constraints

作　　者：刘子岩 许亮[1] LIU Ziyan;XU Liang(China Academy of Aerospace Aerodynamics,Beijing 100074,China)

机构地区：[1]中国航天空气动力技术研究院,北京100074

出　　处：《计算物理》2023年第6期761-769,共9页Chinese Journal of Computational Physics

基　　金：国家自然科学基金(11872351)资助项目。

摘　　要：研究模拟可压缩多介质流的机器学习建模方法,利用神经网络实现多介质Riemann解的回归预测。为使训练结果更加符合流动物理,根据流场间断关系构建神经网络额外的物理约束层。建立神经网络代理模型,应用于实用虚拟流体方法(PGFM)。通过各种典型的一维与二维多介质流动问题,对不同规模神经网络训练得到的代理模型进行数值验证。研究发现:嵌入物理约束后,神经网络模型的结果更符合真实情况,而且较简单的神经网络模型即可满足计算需求。机器学习建模方法具有较高的计算精度和计算效率,具备发展潜力。A machine learning method for simulating compressible multi-medium flows is studied.The regression prediction of multi-medium Riemann solution is realized by using neural network.In order to make the training results more consistent with the physical flow,an additional physical constraint layer is constructed according to the discontinuity relationship of the flow field.A neural network model is established and applied to practical ghost fluid method(PGFM).Through a variety of typical one-dimensional and two-dimensional multi-medium flow problems,the surrogates trained by neural networks of different sizes are verified numerically.It is found that the results of neural network model are more consistent with the real situation after embedding physical constraints.In addition,the relatively simple neural network model can meet the computing requirements.Machine learning method has high computational accuracy and efficiency,and has potential development.

关 键 词：可压缩多介质流 虚拟流体方法 多介质Riemann问题 神经网络 物理约束 

分 类 号：O359[理学—流体力学]