欢迎光临《计算力学学报》官方网站!
 
基于物理信息的神经网络求解曲面上对流扩散方程
Physics-Informed Neural Networks for Solving Convection-Diffusion Equations on Surfaces
投稿时间:2021-10-21  修订日期:2021-12-21
DOI:
中文关键词:  机器学习  自动微分  Laplace-Beltrami算子  物理模型  曲面  
英文关键词:machine learning  automatic differentiation  Laplace-Beltrami operator  physical model  surfaces  
基金项目:国家自然科学基金项目,国家自然科学基金项目(面上项目,重点项目,重大项目)
作者单位邮编
汤卓超 河海大学力学与材料学院 211100
傅卓佳 河海大学力学与材料学院 211100
摘要点击次数: 15
全文下载次数: 0
中文摘要:
      本文引入基于物理信息的神经网络(Physics-informed Neural Networks,简称PINNs)并将其用于求解曲面对流扩散方程。区别于传统的神经网络模型,PINNs在建立模型过程中引入了自动微分技术,并将物理信息即偏微分方程信息编译其中,通过定义损失函数得到关于该模型中神经网络参数即权重和偏置的优化目标,随后利用已有的优化算法进行求解。显而易见,PINNs通过添加额外的物理信息约束放宽了对于数据量的要求,对于一个确定性模型显示出更好的鲁棒性。本文基于曲面微分算子与欧氏空间下标准微分算子的解析关系,引入两种曲面微分算子处理技术,即非本征技术和嵌入技术,并结合PINNs针对定义在高维复杂曲面上的对流扩散方程进行求解,多个数值算例证明了该方法的有效性、鲁棒性以及其在求解此类问题的潜力。
英文摘要:
      This paper introduces Physics-informed Neural Networks and applies them to surface partial differential equations. Different from the traditional Neural Network model, PINNs introduce automatic differentiation technology in the process of establishing the model and encode the physical information, that is, the partial differential equation information into it. By defining the loss function, the optimization goal of the neural network parameters in the model including weights and biases can be obtained explicitly and solved by the existing optimization algorithms. Obviously, PINNs reduce the requirements for the amount of data by adding additional physical information constraints, and show better robustness for some deterministic models. Based on the analytical relationship between surface differential operators and standard differential operators in Euclidean space, this paper introduces two techniques including the extrinsic way and embedding way and combines them with PINNs to solve the convection-diffusion equations defined on high-dimensional complex surfaces. Numerous numerical examples prove the effectiveness, robustness and potential of this method in solving such problems.
  查看/发表评论  下载PDF阅读器
您是第9452609位访问者
版权所有:《计算力学学报》编辑部
本系统由 北京勤云科技发展有限公司设计