About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
H-10: A Framework for Foundation Models in Materials Sciences: Application to 3D Polycrystalline Materials |
| Author(s) |
Andreas E. Robertson, Michael Buzzy, Peng Chen, Surya R. Kalidindi |
| On-Site Speaker (Planned) |
Andreas E. Robertson |
| Abstract Scope |
Foundation Models are generalist deep learning models capable of (1) providing high quality representation for large diverse datasets (e.g., polycrystalline microstructures from many material systems) and (2) easy repurposing to tackle new challenges outside of their training scope. Foundation models enable scientists to rapidly experiment by providing a strong – and, importantly, general – starting point. Unfortunately, developing such models is challenging in materials science because (1) data is highly limited and (2) a clear mechanism enabling repurposing has not been identified. In this poster, we address these limitations and present PolyMicros, a foundation model for 3D Polycrystalline Microstructure. Notably, utilizing PolyMicros, we tackle several inverse problems without any specialized training or datasets. To achieve this, we present a statistical framework for generating high quality synthetic data. We also develop a deep diffusion model framework for scientific foundation models that enables repurposing by using Bayes rule. |
| Proceedings Inclusion? |
Planned: |