Abstract:Engineering components must meet increasing technological demands in ever shorter development cycles. To face these challenges, a holistic approach is essential that allows for the concurrent development of part design, material system and manufacturing process. Current approaches employ numerical simulations, which however quickly becomes computation-intensive, especially for iterative optimization. Data-driven machine learning methods can be used to replace time- and resource-intensive numerical simulations. In particular, MeshGraphNets (MGNs) have shown promising results. They enable fast and accurate predictions on unseen mesh geometries while being fully differentiable for optimization. However, these models rely on large amounts of expensive training data, such as numerical simulations. Physics-informed neural networks (PINNs) offer an opportunity to train neural networks with partial differential equations instead of labeled data, but have not been extended yet to handle time-dependent simulations of arbitrary meshes. This work introduces PI-MGNs, a hybrid approach that combines PINNs and MGNs to quickly and accurately solve non-stationary and nonlinear partial differential equations (PDEs) on arbitrary meshes. The method is exemplified for thermal process simulations of unseen parts with inhomogeneous material distribution. Further results show that the model scales well to large and complex meshes, although it is trained on small generic meshes only.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is how to quickly and accurately solve unsteady and nonlinear partial differential equations (PDEs) on any grid, especially in the simulation of complex thermal processes encountered in engineering component design, material systems, and manufacturing processes. Specifically, the paper proposes a new method - Physics - Informed MeshGraphNets (PI - MGNs), aiming to overcome the problem of high computational cost in traditional numerical simulations and improve the generalization ability of the model for unseen geometries and material properties. ### Main problems: 1. **Computational efficiency**: Traditional numerical simulation methods (such as the Finite Element Method - FEM) are accurate but time - consuming in calculation, especially in iterative optimization. 2. **Generalization ability**: Existing data - driven machine - learning methods require a large amount of expensive training data and are difficult to handle unseen geometries and material properties. 3. **Unsteady and nonlinear problems**: Many practical engineering problems involve unsteady (time - varying) and nonlinear physical phenomena, which pose challenges to existing methods. ### Solutions: - **Combination of Physics - Informed Neural Networks (PINNs) and MeshGraphNets (MGNs)**: By introducing physics - informed training (i.e., directly using physical equations as training targets), the dependence on a large amount of labeled data is reduced, and the generalization ability of the model is improved. - **Solving unsteady and nonlinear PDEs on arbitrary grids**: PI - MGNs can efficiently solve unsteady and nonlinear PDEs on arbitrary grids, which is suitable for complex engineering problems. - **Temporal bundling and global features**: By introducing temporal bundling and global features, the performance and stability of the model are further enhanced. ### Application scenarios: - **Thermal process simulation**: The paper shows the application of PI - MGNs in heat conduction problems with different complexities, including two - dimensional linear heat diffusion, two - dimensional convex polygons with nonlinear heat sources, and the heating problem of hollow cylinders under three - dimensional mixed boundary conditions. - **Generalization ability for large - scale unseen grids**: The generalization ability of PI - MGNs on unseen large - scale grids is verified, indicating that it can be applied to a wider range of engineering scenarios. ### Conclusion: By introducing physics - informed training and an improved network structure, PI - MGNs not only significantly improves computational efficiency but also can maintain high prediction accuracy on unseen geometries and material properties, providing a new solution for engineering simulation. ### Formula summary: - **Unsteady PDE**: \[ \frac{\partial T}{\partial t}=\alpha\nabla^{2}T + q(T),\quad\forall x\in\Omega \] where \(T = T(x, t)\) represents a physical quantity (such as temperature), \(\alpha=\alpha(x, t)\) is the diffusion coefficient, and \(q(T)\) is the nonlinear source term. - **Weak form of FEM**: \[ \int_{\Omega}\left(\frac{\partial T}{\partial t}\phi+\alpha\nabla T\cdot\nabla\phi - q(T)\phi\right)dV-\int_{\partial\Omega_N}\alpha h_N\phi dA = 0 \] - **FEM discretization**: \[ \sum_{v = 1}^{N_{\phi}}\Phi\left(\frac{T_v^{n + 1}-T_v^n}{\Delta t},T_v^{n + 1},\phi_v,\phi_m\right)=0 \] - **Loss function**: \[ L_n^{\text{FEM}}=\frac{1}{N_{TB}N_{\phi,F}}\sum_{\tilde{n}=1}^{N_{TB}}

Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes

Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes

Physics-informed Neural Networks to Solve Lumped Kinetic Model for Chromatography Process.

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

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

PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems

Mesh-based GNN surrogates for time-independent PDEs

Physics-Informed Neural Networks for Solving 2-D Magnetostatic Fields

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

M‐PINN: A mesh‐based physics‐informed neural network for linear elastic problems in solid mechanics

Physics-Informed Graph-Mesh Networks for PDEs: A hybrid approach for complex problems

PIG: Physics-Informed Gaussians as Adaptive Parametric Mesh Representations

Up-sampling-only and Adaptive Mesh-based GNN for Simulating Physical Systems

Learning Mesh-Based Simulation with Graph Networks

Deep Learning-based surrogate models for parametrized PDEs: handling geometric variability through graph neural networks

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

Scientific Computing Algorithms to Learn Enhanced Scalable Surrogates for Mesh Physics

Graph Neural Networks for Mesh Generation and Adaptation in Structural and Fluid Mechanics

PINN-MG: A Multigrid-Inspired Hybrid Framework Combining Iterative Method and Physics-Informed Neural Networks

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

Physics-informed attention-based neural network for solving non-linear partial differential equations