Abstract:The great success of Physics-Informed Neural Networks (PINN) in solving partial differential equations (PDEs) has significantly advanced our simulation and understanding of complex physical systems in science and engineering. However, many PINN-like methods are poorly scalable and are limited to in-sample scenarios. To address these challenges, this work proposes a novel discrete approach termed Physics-Informed Graph Neural Network (PIGNN) to solve forward and inverse nonlinear PDEs. In particular, our approach seamlessly integrates the strength of graph neural networks (GNN), physical equations and finite difference to approximate solutions of physical systems. Our approach is compared with the PINN baseline on three well-known nonlinear PDEs (heat, Burgers and FitzHugh-Nagumo). We demonstrate the excellent performance of the proposed method to work with irregular meshes, longer time steps, arbitrary spatial resolutions, varying initial conditions (ICs) and boundary conditions (BCs) by conducting extensive numerical experiments. Numerical results also illustrate the superiority of our approach in terms of accuracy, time extrapolability, generalizability and scalability. The main advantage of our approach is that models trained in small domains with simple settings have excellent fitting capabilities and can be directly applied to more complex situations in large domains.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is the lack of scalability and out - of - sample generalization ability in existing physics - informed neural networks (PINNs) when solving partial differential equations (PDEs). Specifically, the paper proposes a new discrete method - physics - informed graph neural network (PIGNN) to solve forward and backward nonlinear spatio - temporal PDEs. ### Core Problems and Solutions in the Paper 1. **Limitations of Existing Methods**: - **Poor Scalability**: Many PINN - based methods are difficult to handle large - scale or complex scenarios. - **Limited Out - of - Sample Generalization Ability**: These methods usually perform well only within the training data range and perform poorly for unseen time scales, spatial resolutions, or computational domains. - **High Computational Cost**: A large number of collocation points are required to calculate the PDE residuals, especially in high - dimensional spaces, resulting in huge training costs. 2. **Proposed Solutions**: - **Introduce the PIGNN Framework**: Combine graph neural networks (GNNs), physical equations, and the finite - difference method to solve forward and backward PDEs more efficiently. - **Applicable to Irregular Grids**: By converting irregular grids into graph structures, PIGNN can handle arbitrary spatial resolutions, varying initial conditions (ICs), and boundary conditions (BCs). - **Reduce Computational Cost**: Use the finite - difference method and the least - squares technique to approximate differential operators, thereby significantly reducing the computational complexity. - **Enhance Generalization Ability**: After being trained in a small domain, the model can be directly applied to larger and more complex domains and maintain good fitting ability. ### Specific Contributions 1. **Efficiently Solve Forward and Reverse PDEs**: - For forward problems, the physical - informed loss can be calculated without labeled data. - For reverse problems, through a small amount of labeled data and a regularization term constrained by the underlying PDE, the unknown parameters are quickly learned and physical consistency is ensured. 2. **Propose the Combination of the Finite - Difference Method and the Least - Squares Technique**: - Used to approximate differential operators on graphs and discuss their approximation errors. 3. **Ability to Handle Irregular Domains**: - Directly constructed from irregular grids, able to handle irregular geometries and support adaptive resolutions. 4. **General Rules for Selecting Appropriate Features**: - Propose selection rules for node and edge features, which help the model generalize in out - of - sample scenarios. 5. **Numerical Experiment Verification**: - A series of numerical experiments are carried out on multiple typical PDEs, demonstrating the superior performance of PIGNN in terms of accuracy, time extrapolation, generalization ability, and scalability. 6. **Extension to Multi - GPU Systems**: - Apply PIGNN to distributed multi - GPU systems to solve large - scale problems, demonstrating its scalability. ### Conclusion Through the above improvements, PIGNN not only improves the efficiency and accuracy of solving PDEs but also shows strong generalization ability and applicability in handling the simulation of complex physical systems.

Combining physics-informed graph neural network and finite difference for solving forward and inverse spatiotemporal PDEs

PhyGNNet: Solving spatiotemporal PDEs with Physics-informed Graph Neural Network

Physics-informed graph neural Galerkin networks: A unified framework for solving PDE-governed forward and inverse problems

A Spatial-Temporally Adaptive PINN Framework for 3D Bi-Ventricular Electrophysiological Simulations and Parameter Inference

RBF-MGN:Solving spatiotemporal PDEs with Physics-informed Graph Neural Network

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

Generalized conditional symmetry enhanced physics-informed neural network and application to the forward and inverse problems of nonlinear diffusion equations

An Efficient Framework for Solving Forward and Inverse Problems of Nonlinear Partial Differential Equations Via Enhanced Physics-Informed Neural Network Based on Adaptive Learning

Hybrid Finite Difference with the Physics-informed Neural Network for solving PDE in complex geometries

A High-Efficient Hybrid Physics-Informed Neural Networks Based on Convolutional Neural Network

PPINN: Parareal physics-informed neural network for time-dependent PDEs

Physics-Informed neural network solver for numerical analysis in geoengineering

MRF-PINN: A Multi-Receptive-Field convolutional physics-informed neural network for solving partial differential equations

PhyCRNet: Physics-informed convolutional-recurrent network for solving spatiotemporal PDEs

Physics-informed neural networks combined with polynomial interpolation to solve nonlinear partial differential equations

Distributed physics informed neural network for data-efficient solution to partial differential equations

Deep fuzzy physics-informed neural networks for forward and inverse PDE problems

Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equations

General-Kindred Physics-Informed Neural Network to the Solutions of Singularly Perturbed Differential Equations

Physics-informed neural networks with domain decomposition for the incompressible Navier–Stokes equations

Data-Guided Physics-Informed Neural Networks for Solving Inverse Problems in Partial Differential Equations