A review of graph neural network applications in mechanics-related domains

Yingxue Zhao, Haoran Li, Haosu Zhou, Hamid Reza Attar, Tobias Pfaff, Nan Li

Introduction

Mechanics-related tasks, spanning solid and fluid dynamics, often present unique challenges in achieving accurate geometric and physical representations. Traditional numerical simulations, such as Finite Element Analysis (FEA) and Computational Fluid Dynamics (CFD), are the gold standard for solving these complex, non-linear problems. However, these methods are computationally expensive, often requiring significant time and resources for high-fidelity results.

While traditional machine learning models like Multi-Layer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) have been explored as surrogates to accelerate these tasks, they face fundamental limitations. CNNs, for instance, require regular, pixel-like grids, which struggle to capture the irregular geometries typical of mechanical structures—such as the sharp tip of a crack or the complex curvature of an airfoil. Graph Neural Networks (GNNs) have emerged as a powerful solution, natively processing the irregular, non-Euclidean data structures (like meshes) that define the field of mechanics.

The Machine Learning Pipeline in Mechanics

Applying machine learning to mechanics involves a structured workflow that integrates domain knowledge with data-driven techniques. This process typically begins with the definition of the mechanics problem and the design of experiments or simulations to generate training data.

The general workflow for applying machine learning models to solve mechanics-related problems, from problem definition to model deployment and retraining.

As shown in the pipeline above, the data generated from computational or experimental results undergoes processing to form datasets for training, validation, and testing. The selection of the machine learning model is critical, with GNNs increasingly favored for their ability to handle the complex connectivity of mechanical systems. Once trained and evaluated, these models can be deployed for real-time prediction of mechanical properties, classification of load levels, or inverse design of mechanical metamaterials.

GNN Architectures and Methodologies

GNNs represent a specialized subset of deep neural networks designed to operate on graph-structured data. In the context of mechanics, nodes in a graph can represent material points, particles, or mesh vertices, while edges represent physical interactions or spatial proximity.

Hierarchy of artificial neural networks highlighting the relationship between shallow networks, deep networks, and the specialized domain of Graph Neural Networks.

Several GNN architectures have become prominent in the literature:

  • Message Passing Neural Networks (MPNNs): These are widely used for spatial-temporal rollouts, such as tracking the movement of fluid particles or the propagation of cracks over time.
  • Graph Convolutional Networks (GCNs): These are frequently employed for static field predictions, such as determining stress distributions within a structural beam.
  • Graph Attention Networks (GATs) and GraphSAGE: These models handle heterogeneous connectivity by dynamically weighing the importance of neighboring nodes, allowing the network to focus on critical regions of a structure.

To overcome computational bottlenecks in large-scale simulations, researchers have developed advanced techniques like multi-scale graphs, which allow information to "shortcut" across massive meshes, preventing the "oversmoothing" common in deep GNNs.

Key Insights and Advantages

The adoption of GNNs in mechanics offers several transformative advantages:

  1. Natural Fit for Lagrangian Mechanics: GNNs are uniquely suited for Lagrangian frameworks, where material points or particles are tracked as they move. This makes them significantly more effective than grid-based CNNs for simulating large deformations and fluid-structure interactions.
  2. Mesh Independence: Unlike models that require fixed-size inputs, GNNs can process meshes of varying resolutions and topologies. This adaptability allows models to generalize across different geometries without retraining.
  3. Physics-Informed Training: By integrating physical laws (e.g., conservation of mass and momentum) directly into the loss function, GNNs can achieve higher accuracy and better generalizability, reducing the risk of producing unphysical results.

Performance and Real-World Impact

The review highlights that GNN surrogate models routinely achieve massive speedups, often operating orders of magnitude faster than traditional numerical solvers. For example, some models have predicted microcrack propagation up to 25 times faster than traditional FEA. Despite this speed, models like MeshGraphNet (MGN) maintain high fidelity, matching the accuracy of traditional simulators even in complex scenarios like turbulent wakes or shockwave predictions.

Furthermore, GNNs demonstrate impressive generalizability. A model trained on simple 2D geometries can often successfully predict the dynamics of much more complex 3D structures with significantly more nodes. To support future research, the authors have compiled a repository of open-source datasets and codebases, covering domains from hyper-elastic plates to cardiac mechanics.

Future Directions and Challenges

Despite their promise, several hurdles remain for the industrial deployment of GNNs:

  • Error Accumulation: In time-dependent simulations, small errors in each step can accumulate, eventually leading to unphysical states. Future work is focusing on hybrid GNN-RNN architectures to mitigate this.
  • Material Non-linearity: While linear elastic models are well-studied, extending GNNs to handle complex elastoplastic behaviors and non-linear material properties is an ongoing area of research.
  • Experimental Integration: Most current models rely on simulation data. Incorporating real-world experimental data, such as from Particle Image Velocimetry (PIV), is essential for validating these models in practical engineering contexts.

In conclusion, GNNs represent a significant leap forward in the intersection of machine learning and mechanics, offering a path toward real-time, high-fidelity simulations of the physical world.

Summary updated

Abstract

Mechanics-related tasks often present unique challenges in achieving accurate geometric and physical representations, particularly for non-uniform structures. Graph neural networks (GNNs) have emerged as a promising tool to tackle these challenges by adeptly learning from graph data with irregular underlying structures. Consequently, recent years have witnessed a surge in complex mechanics-related applications inspired by the advancements of GNNs. Despite this process, there is a notable absence of a systematic review addressing the recent advancement of GNNs in solving mechanics-related tasks. To bridge this gap, this review article aims to provide an in-depth overview of the GNN applications in mechanics-related domains while identifying key challenges and outlining potential future research directions. In this review article, we begin by introducing the fundamental algorithms of GNNs that are widely employed in mechanics-related applications. We provide a concise explanation of their underlying principles to establish a solid understanding that will serve as a basis for exploring the applications of GNNs in mechanics-related domains. The scope of this paper is intended to cover the categorisation of literature into solid mechanics, fluid mechanics, and interdisciplinary mechanics-related domains, providing a comprehensive summary of graph representation methodologies, GNN architectures, and further discussions in their respective subdomains. Additionally, open data and source codes relevant to these applications are summarised for the convenience of future researchers. This article promotes an interdisciplinary integration of GNNs and mechanics and provides a guide for researchers interested in applying GNNs to solve complex mechanics-related tasks.
BibTeX
@inproceedings{zhao2024review,
  author = {Yingxue Zhao and Haoran Li and Haosu Zhou and Hamid Reza Attar and Tobias Pfaff and Nan Li},
  title = {{A review of graph neural network applications in mechanics-related domains}},
  booktitle = {Artif. Intell. Rev},
  year = {2024},
  url = {https://openreview.net/forum?id=xvblqTkTEQ},
}
Figure