About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Accelerating Large Multiscale Composite Simulations with a GNN/LSTM Microscale Surrogate |
| Author(s) |
Joshua A. Stuckner, Trenton Ricks, Brandon Hearley, Steven Arnold |
| On-Site Speaker (Planned) |
Joshua A. Stuckner |
| Abstract Scope |
We introduce a neural network architecture containing graph convolutional network (GCN) and long short-term memory (LSTM) layers to dramatically accelerate a multiscale physics-based composite modeling simulation by acting as a fast and accurate surrogate for the microscale. The GCN layers model microstructure while the LSTM models damage and loading history. An initial LSTM-only model was deployed within a multiscale simulation containing both the microscale neural network surrogate and a physics-based classical lamination theory model to handle the macroscale. The multiscale model was 145 times faster with a coefficient of determination of 0.98. The final model generalized to varying microstructure using GCN layers and predicts elements of the stiffness matrix with a coefficient of determination of 0.985 to 0.999. We also report on our open-source library to interface FORTRAN simulations with TensorFlow and work to integrate the final microscale surrogate with commercial finite element software to perform large multiscale analysis. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, Composites |