Abstract:This study introduces a two-scale Graph Neural Operator (GNO), namely, LatticeGraphNet (LGN), designed as a surrogate model for costly nonlinear finite-element simulations of three-dimensional latticed parts and structures. LGN has two networks: LGN-i, learning the reduced dynamics of lattices, and LGN-ii, learning the mapping from the reduced representation onto the tetrahedral mesh. LGN can predict deformation for arbitrary lattices, therefore the name operator. Our approach significantly reduces inference time while maintaining high accuracy for unseen simulations, establishing the use of GNOs as efficient surrogate models for evaluating mechanical responses of lattices and structures.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper aims to address the high-cost nonlinear finite element simulation problem of 3D lattice structures. Specifically, the authors propose a two-scale graph neural operator named LatticeGraphNet (LGN) as an efficient surrogate model to predict the deformation of 3D lattice structures under compression. Traditional physical experiments and high-fidelity numerical simulation methods can accurately describe the mechanical properties of lattice structures but are costly and time-consuming. For example, simulations using the Incremental Potential Contact (IPC) method can take 48 hours to 10 days, depending on mesh size, topological characteristics, and nonlinear factors such as buckling and self-contact.
### Main Contributions
1. **Two-Scale Graph Neural Network Architecture**:
- **LGN-i**: Learns a simplified (beam) representation of the lattice to predict coarse displacements.
- **LGN-ii**: Maps the simplified displacements predicted by LGN-i back to the full 3D mesh to predict detailed displacements.
2. **Efficient Predictive Capability**:
- By training the model with 108 high-fidelity IPC simulation data, LGN can complete inference in 10 to 30 minutes, whereas traditional FEM simulations require thousands of hours.
- For unseen simulations, LGN's point-to-point average error remains within 1.67% of the structure size.
3. **Application Prospects**:
- This method provides a baseline for using Graph Neural Networks (GNN) as surrogate models to simulate the mechanical response of lattices and structures.
- In practical applications such as 3D printing, LGN can significantly accelerate high-fidelity analysis, promoting the broader application of complex structures.
### Method Overview
- **LGN-i**: Learns a simplified representation of the lattice through three components: encoder, processor, and decoder. Inputs include features of nodes and edges, and outputs include displacement increments and stress changes of nodes.
- **LGN-ii**: Based on the predictions of LGN-i, it further predicts the displacements of the full 3D mesh. It encodes spatial information by constructing subgraphs and updates the features of nodes and edges using a message-passing mechanism.
### Results and Discussion
- **Deformation Prediction**: LGN can accurately predict the deformation of elastic lattice structures, including buckling phenomena. The point-to-point error distribution is concentrated at lower values, with most points having an error within 1 millimeter.
- **Force Prediction**: Although LGN cannot directly predict reaction forces, they can be approximately calculated through homogenization methods. Despite some errors, the overall trend is consistent with IPC simulation results, indicating that LGN can capture the topological characteristics of the structure.
### Conclusion
LatticeGraphNet (LGN) provides an efficient method for predicting the response of 3D lattice structures. This method not only significantly reduces inference time but also shows high accuracy on unseen geometries. Future work can further improve the model's accuracy by expanding the training dataset.