A fully differentiable GNN-based PDE Solver: With Applications to Poisson and Navier-Stokes Equations

Tianyu Li,Yiye Zou,Shufan Zou,Xinghua Chang,Laiping Zhang,Xiaogang Deng

2024-05-08

Abstract:In this study, we present a novel computational framework that integrates the finite volume method with graph neural networks to address the challenges in Physics-Informed Neural Networks(PINNs). Our approach leverages the flexibility of graph neural networks to adapt to various types of two-dimensional unstructured grids, enhancing the model's applicability across different physical equations and boundary conditions. The core innovation lies in the development of an unsupervised training algorithm that utilizes GPU parallel computing to implement a fully differentiable finite volume method discretization process. This method includes differentiable integral and gradient reconstruction algorithms, enabling the model to directly solve partial-differential equations(PDEs) during training without the need for pre-computed data. Our results demonstrate the model's superior mesh generalization and its capability to handle multiple boundary conditions simultaneously, significantly boosting its generalization capabilities. The proposed method not only shows potential for extensive applications in CFD but also establishes a new paradigm for integrating traditional numerical methods with deep learning technologies, offering a robust platform for solving complex physical problems.

Fluid Dynamics,Mathematical Physics,Computational Physics

What problem does this paper attempt to address?

This paper proposes a new computational framework that combines the finite volume method with graph neural networks (GNN) to address the challenges faced by Physics-Informed Neural Networks (PINNs) in handling different physical equations and boundary conditions. The method utilizes the flexibility of GNN to adapt to various two-dimensional unstructured grids, improving the applicability of the model to different physical scenarios. The core innovation lies in the development of an unsupervised training algorithm that utilizes GPU parallel computing to implement the discretization of fully differential finite volume methods, allowing the model to directly solve partial differential equations (PDEs) during the training process without precomputing data. This approach demonstrates excellent grid generalization capability in handling complex geometries, multiscale phenomena, and multiple physical phenomena, while also being able to handle multiple boundary conditions simultaneously, enhancing the model's generalization performance. The paper also showcases the potential application of this method in fluid mechanics equations such as the Poisson equation and the Navier-Stokes equation, providing a powerful platform for the integration of traditional numerical methods and deep learning techniques to solve complex physical problems.