| About this Abstract |
| Meeting |
TMS Specialty Congress 2025
|
| Symposium
|
The 7th International Congress on 3D Materials Science (3DMS 2025)
|
| Presentation Title |
Graph Neural Networks for Generalizable Machine Learning of 3D Microstructure--Property Relationships in Polycrystals |
| Author(s) |
Guangyu Hu, Gyu-Jang Sim, Myoung-Gyu Lee, Marat I. Latypov |
| On-Site Speaker (Planned) |
Marat I. Latypov |
| Abstract Scope |
3D characterization provides unique datasets for understanding process--structure--property relationships in alloys. Graphs offer a reduced-order representation of the 3D polycrystalline microstructure that enables efficient machine learning of these relationships with graph neural networks (GNNs). In this contribution, we will present our recent advances in data-efficient training of GNNs that can generalize predictions well beyond their training sets. Specifically, we will present GNNs that predict (i) anisotropic mechanical properties of textured polycrystals in new loading directions; (ii) grain-level fatigue indicator parameters in microstructure volume elements order of magnitude larger than those used for training; and (iii) properties of polycrystals for metals and alloys not included in the training set. This contribution intends to spark 3DMS community's interest in the GNN computational platform that could significantly benefit from experimental datasets for model training and fine-tuning. |
| Proceedings Inclusion? |
Undecided |