Scientific Computing Algorithms to Learn Enhanced Scalable Surrogates for Mesh Physics

Brian R. Bartoldson,Yeping Hu,Amar Saini,Jose Cadena,Yucheng Fu,Jie Bao,Zhijie Xu,Brenda Ng,Phan Nguyen

2023-04-01

Abstract:Data-driven modeling approaches can produce fast surrogates to study large-scale physics problems. Among them, graph neural networks (GNNs) that operate on mesh-based data are desirable because they possess inductive biases that promote physical faithfulness, but hardware limitations have precluded their application to large computational domains. We show that it is \textit{possible} to train a class of GNN surrogates on 3D meshes. We scale MeshGraphNets (MGN), a subclass of GNNs for mesh-based physics modeling, via our domain decomposition approach to facilitate training that is mathematically equivalent to training on the whole domain under certain conditions. With this, we were able to train MGN on meshes with \textit{millions} of nodes to generate computational fluid dynamics (CFD) simulations. Furthermore, we show how to enhance MGN via higher-order numerical integration, which can reduce MGN's error and training time. We validated our methods on an accompanying dataset of 3D $\text{CO}_2$-capture CFD simulations on a 3.1M-node mesh. This work presents a practical path to scaling MGN for real-world applications.

Machine Learning,Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to effectively apply Graph Neural Networks (GNNs), especially mesh - based GNNs (such as MeshGraphNets, MGN), in large - scale physical simulations. Specifically, the paper focuses on the following aspects: 1. **Hardware limitations**: Existing mesh - based GNNs are difficult to be applied to large - scale computational domains due to hardware limitations and memory requirements. One of the goals of the paper is to overcome these limitations, enabling GNNs to handle 3D meshes containing millions of nodes. 2. **Model scalability**: The paper proposes a domain decomposition method. By dividing a large - scale mesh into multiple sub - domains, the extended training of the model can be achieved. This method is mathematically equivalent to directly training on the full mesh, so that large - scale GNNs can be trained under existing hardware conditions. 3. **Model accuracy and efficiency**: The paper explores how to improve the accuracy and training efficiency of MGN through high - order numerical integration methods (such as Heun's method). These methods can reduce the error and training time of MGN, thereby enhancing the overall performance of the model. 4. **Practical applications**: The paper shows how to apply these methods to actual Computational Fluid Dynamics (CFD) simulations, especially in the application of the carbon capture process. By using the 3D CO2 capture CFD simulation data set (C3FD), the effectiveness of the proposed methods is verified. In summary, the main goal of this paper is to make mesh - based GNNs more effectively applicable in large - scale physical simulations through improvement and expansion, while maintaining or improving the accuracy and efficiency of the model.

Scientific Computing Algorithms to Learn Enhanced Scalable Surrogates for Mesh Physics

X-MeshGraphNet: Scalable Multi-Scale Graph Neural Networks for Physics Simulation

Scalable and Consistent Graph Neural Networks for Distributed Mesh-based Data-driven Modeling

Learning Mesh-Based Simulation with Graph Networks

Mesh-based GNN surrogates for time-independent PDEs

Multi-GPU Approach for Training of Graph ML Models on large CFD Meshes

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

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

Bayesian Mesh Optimization for Graph Neural Networks to Enhance Engineering Performance Prediction

Multiscale graph neural networks with adaptive mesh refinement for accelerating mesh-based simulations

Multi-Grid Graph Neural Networks with Self-Attention for Computational Mechanics

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

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

GNN-Surrogate: A Hierarchical and Adaptive Graph Neural Network for Parameter Space Exploration of Unstructured-Mesh Ocean Simulations

On the Scalability of GNNs for Molecular Graphs

Mesh-based Super-Resolution of Fluid Flows with Multiscale Graph Neural Networks

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

Learning CO$_2$ plume migration in faulted reservoirs with Graph Neural Networks

MAgNET: A Graph U-Net Architecture for Mesh-Based Simulations

A Microstructure-based Graph Neural Network for Accelerating Multiscale Simulations