Abstract:In many cutting-edge applications, high-fidelity computational models prove to be too slow for practical use and are therefore replaced by much faster surrogate models. Recently, deep learning techniques have increasingly been utilized to accelerate such predictions. To enable learning on large-dimensional and complex data, specific neural network architectures have been developed, including convolutional and graph neural networks. In this work, we present a novel encoder-decoder geometric deep learning framework called MAgNET, which extends the well-known convolutional neural networks to accommodate arbitrary graph-structured data. MAgNET consists of innovative Multichannel Aggregation (MAg) layers and graph pooling/unpooling layers, forming a graph U-Net architecture that is analogous to convolutional U-Nets. We demonstrate the predictive capabilities of MAgNET in surrogate modeling for non-linear finite element simulations in the mechanics of solids.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to construct an efficient surrogate model on high - dimensional data to accelerate prediction in nonlinear finite - element simulations. Specifically, the authors propose a novel graph U - Net architecture named MAgNET, which is designed to handle any graph - structured data, especially for mesh - based simulations. This method extends the capabilities of convolutional neural networks, enabling them to adapt to complex data structures, thereby significantly improving computational efficiency while maintaining the required accuracy. This is particularly important for application scenarios that require real - time feedback or control, such as in fields like computer - aided surgery or soft robotics. ### Main contributions of the paper: 1. **Proposing the MAgNET framework**: This is a graph U - Net architecture that combines multi - channel aggregation (MAg) layers and graph pooling / unpooling layers and can effectively handle any graph - structured data. 2. **Innovative graph pooling / unpooling operations**: These operations allow the model to compress data during the encoding stage and restore the original fine - mesh structure during the decoding stage, thereby achieving efficient information encoding and decoding. 3. **Applicability to high - dimensional relationships**: It is especially suitable for cases where the input and output dimensions are large, such as experimental full - field measurement data or synthetic mesh data generated from finite - element simulations. 4. **Verification and application**: The effectiveness and prediction ability of the MAgNET framework are verified through multiple 2D and 3D nonlinear finite - element simulation benchmark tests and compared with the traditional CNN U - Net architecture. ### Problems solved: - **Computational efficiency problem**: Although high - fidelity numerical simulations are accurate, they are computationally expensive and difficult to meet the requirements of real - time applications. MAgNET significantly improves the prediction speed by constructing an efficient surrogate model. - **Data structure limitation**: Traditional convolutional neural networks (CNNs) can only handle mesh - structured data, which limits their use in many practical applications. MAgNET overcomes this limitation by introducing the concept of graph neural networks (GNNs) and can handle any graph - structured data. ### Application prospects: - **Engineering and physics**: Especially in fields that require fast and accurate predictions, such as materials science, fluid mechanics, etc. - **Medical imaging**: For example, in computer - aided surgery, real - time feedback and control are required. - **Soft robotics**: It needs to respond quickly and adapt to environmental changes. In conclusion, the MAgNET framework provides a new solution for efficient surrogate modeling on high - dimensional data and has broad application prospects.

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

Convolution, aggregation and attention based deep neural networks for accelerating simulations in mechanics

Learning Mesh-Based Simulation with Graph Networks

Enhancing Graph U-Nets for Mesh-Agnostic Spatio-Temporal Flow Prediction

Scientific Computing Algorithms to Learn Enhanced Scalable Surrogates for Mesh Physics

Topology-Agnostic Graph U-Nets for Scalar Field Prediction on Unstructured Meshes

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

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

AMGNET: multi-scale graph neural networks for flow field prediction

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

MAGNet: Multi-agent Graph Network for Deep Multi-agent Reinforcement Learning

Mesh Neural Networks Based on Dual Graph Pyramids

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

MagNet: A Neural Network for Directed Graphs

MeshingNet: A New Mesh Generation Method based on Deep Learning

NeuralMAG: Fast and Generalizable Micromagnetic Simulation with Deep Neural Nets

A graph neural network-based framework to identify flow phenomena on unstructured meshes

A generative learning and graph-based framework for computing field variables in Finite Element simulations

Developing a novel structured mesh generation method based on deep neural networks

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

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