Abstract
Mechanics-related problems 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 problems. 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. The scope of this paper covers 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.
The Challenge of Irregular Geometries in Mechanics
Traditional numerical simulations, such as Finite Element Analysis (FEA) and Computational Fluid Dynamics (CFD), are the gold standard for engineering but are computationally expensive and slow. While standard Machine Learning—specifically Convolutional Neural Networks (CNNs)—has been used to accelerate these processes, CNNs are inherently "grid-bound." They struggle with the irregular meshes, varying resolutions, and Lagrangian (particle-tracking) representations essential to mechanics.
As shown in the figure above, CNNs operate on fixed, regular grids, which limits their ability to represent complex engineering geometries. In contrast, GNNs naturally handle irregular, relational data by defining neighborhoods based on connectivity rather than spatial proximity on a grid.
GNN Architectures and Methodologies
The literature identifies several fundamental GNN "building blocks" currently transforming the field:
- Message Passing Neural Networks (MPNN): The most popular framework in mechanics, where nodes (like mesh vertices) exchange information with neighbors to update their physical states.
- Graph Networks (GN): A generalized framework that includes global attributes (like gravity or total energy) alongside node and edge features.
- Graph Convolutional Networks (GCN) & Attention (GAT): Used for localized feature aggregation and weighting the importance of specific neighbors, such as focusing on high-stress areas.
- Multi-scale/Hierarchical Approaches: Techniques like MAgNET or MeshGraphNet use pooling and unpooling to capture long-range physical dependencies that a single-layer GNN might miss.
A significant trend is the shift toward Physics-Informed GNNs, where loss functions are constrained by physical laws (e.g., Navier-Stokes equations) to ensure results remain physically plausible.
Applications and Performance Gains
GNNs have been applied across a wide range of mechanics sub-domains, from continuum structures to fracture mechanics.
Key results from the literature include:
- Speed: In fracture mechanics, the MicrocrackGNN was found to be 25 times faster than traditional FEA simulations while maintaining high accuracy.
- Accuracy: In biomechanics, the PhysGNN model achieved an absolute displacement error of under 1mm for 97% of nodes in soft tissue simulations, meeting strict requirements for neurosurgery.
- Robustness: GNN models like MeshGraphNet demonstrated the ability to handle "rollouts" (long-term time-step predictions) for complex contacts and self-collisions that would typically cause standard neural networks to diverge.
Key Insights and Future Outlook
GNNs are a "natural fit" for Lagrangian mechanics because nodes can represent moving particles or mesh points, handling deformation and flow much more intuitively than pixel-based CNNs. Furthermore, GNNs excel at generalization; they can often predict the behavior of entirely new shapes because they learn local physical rules rather than global patterns.
However, challenges remain. Scaling to high-fidelity, industrial-scale 3D meshes is computationally taxing, and error accumulation in time-dependent simulations requires further integration with Recurrent Neural Networks (RNNs). Future research is also needed to capture complex material non-linearities and to ground synthetic models in real-world experimental data.
Summary updated