An FEA surrogate model with Boundary Oriented Graph Embedding approach

Xingyu Fu, Fengfeng Zhou, Dheeraj Peddireddy, Zhengyang Kang, Martin Byung-Guk Jun, Vaneet AggarwalarXiv 2021 · arXiv

Abstract

Finite Element Analysis (FEA) is a critical tool in engineering design but is often computationally prohibitive for iterative processes. This paper introduces the Boundary Oriented Graph Embedding (BOGE) approach, which utilizes Graph Neural Networks (GNNs) as high-speed surrogate models. By implementing "shortcuts" in the graph adjacency matrix that connect internal elements directly to boundary conditions, BOGE overcomes the message-passing limitations of traditional GNNs. Our results show that BOGE can predict stress fields and topology optimization outcomes with high precision in milliseconds, offering a speedup of up to seven orders of magnitude over conventional solvers.

Introduction

Engineering design cycles often require thousands of simulations to optimize a structure's performance. Traditional FEA solvers, while accurate, are slow and create significant bottlenecks. While deep learning models like Convolutional Neural Networks (CNNs) have been explored for acceleration, they are generally restricted to regular grids. Graph Neural Networks (GNNs) offer a more flexible alternative for irregular meshes but suffer from "over-smoothing" and slow information propagation across large graphs. In a standard GNN, a force applied at one end of a mesh must traverse many layers of neighbors to influence distant nodes. This paper addresses these challenges through a novel graph embedding strategy that prioritizes boundary information.

Methods

The BOGE approach enhances standard GNN architectures through three primary innovations:

  1. Rich Vertex Embedding: Each node in the graph represents a mesh element and is initialized with a feature vector $f_v$ containing its spatial coordinates, material properties (Young’s modulus, Poisson’s ratio), and geometric edge data (normal vectors and distances to boundaries).
  2. Boundary Shortcuts: To bypass the "many-hop" limitation of standard message passing, the model introduces direct connections between internal elements and elements located on the boundaries where fixtures or external forces are applied.
  3. Multi-hop Local Connectivity: The graph structure is further augmented with 2-hop and 3-hop local connections, providing elements with broader contextual information and helping to smooth the resulting physical fields.
BOGE Shortcut Mechanisms: Local multi-hop connectivity (left) and global boundary shortcuts (right) facilitate rapid information transfer from boundary conditions to internal elements.

The backbone of the system is a 3-layer DeepGCN, which is robust against over-smoothing even with these dense shortcut connections.

Experimental Results

The model was evaluated using a synthetic dataset of cantilever beams with varying dimensions and internal elliptical holes, generated via ABAQUS.

Accuracy and Performance

The BOGE model demonstrated exceptional fidelity in predicting von Mises stress fields and material density distributions for topology optimization. For stress prediction, the model achieved a Mean Squared Error (MSE) of 0.011706 (~2.41% MAPE). In topology optimization tasks, the MSE was even lower at 0.002735.

Comparison of FEA Ground Truth and Surrogate Model Results: The model accurately predicts stress fields for complex geometries like cantilever beams with elliptical holes.

Computational Efficiency

The most striking result is the inference speed. While a traditional ABAQUS simulation for topology optimization might take nearly 500 seconds, the BOGE-based GNN completes the same task in approximately 0.012 ms on a GPU. This represents a speedup of roughly 40 million times, enabling real-time interactive design and massive-scale generative optimization.

Discussion and Conclusion

The BOGE approach demonstrates that physics-informed graph connectivity—specifically the use of global shortcuts to boundary conditions—is a powerful tool for building GNN surrogates. By allowing a shallow network to "see" global boundary conditions simultaneously, the model mimics the physical equilibrium of a structure more effectively than deep, purely local networks.

While currently demonstrated on 2D geometries, the principles of BOGE are extensible to 3D meshes and other physics domains such as fluid dynamics. Future work will focus on optimizing the sensitivity of GNNs to input tensor shapes and exploring generalization across diverse classes of boundary value problems. This research paves the way for a new generation of "instant" engineering simulation tools.

Summary updated

Abstract

In this work, we present a Boundary Oriented Graph Embedding (BOGE) approach for the Graph Neural Network (GNN) to serve as a general surrogate model for regressing physical fields and solving boundary value problems. Providing shortcuts for both boundary elements and local neighbor elements, the BOGE approach can embed structured mesh elements into the graph and performs an efficient regression on large-scale triangular-mesh-based FEA results, which cannot be realized by other machine-learning-based surrogate methods. Focusing on the cantilever beam problem, our BOGE approach cannot only fit the distribution of stress fields but also regresses the topological optimization results, which show its potential of realizing abstract decision-making design process. The BOGE approach with 3-layer DeepGCN model \textcolor{blue}{achieves the regression with MSE of 0.011706 (2.41\% MAPE) for stress field prediction and 0.002735 MSE (with 1.58\% elements having error larger than 0.01) for topological optimization.} The overall concept of the BOGE approach paves the way for a general and efficient deep-learning-based FEA simulator that will benefit both industry and design-related areas.
BibTeX
@misc{fu2021fea,
  author = {Xingyu Fu and Fengfeng Zhou and Dheeraj Peddireddy and Zhengyang Kang and Martin Byung-Guk Jun and Vaneet Aggarwal},
  title = {{An FEA surrogate model with Boundary Oriented Graph Embedding approach}},
  year = {2021},
  eprint = {2108.13509},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.13509},
}
Figure