About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Novel Strategies for Rapid Acquisition and Processing of Large Datasets From Advanced Characterization Techniques
|
| Presentation Title |
Enabling Data Starved Microstructural Segmentation With Foundation Models as First-Pass Segmentors in Low-Contrast Al-Si Solidification |
| Author(s) |
Nicholas Amano, Aramanda Kiran, Ashwin Shahani, Elizabeth A Holm |
| On-Site Speaker (Planned) |
Nicholas Amano |
| Abstract Scope |
Microstructural image segmentation is a critical tool in materials science, enabling quantitative analysis of phase behavior and solidification evolution, keys in understanding material properties. Image foundation models enable the extraction of valuable information from common object images but are often limited in their capture of microstructural information. While segmentations produced by models like meta’s SAM 2 are often insufficient for comprehensive microstructural analysis, a limited number of accurate frames can be leveraged for further model training. The training frames that capture representative features of the underlying structures enable the training of supervised models from a small subset of reasonably segmented examples. In this study, we demonstrate that rough, unsupervised segmentations generated by SAM 2 can be used to bootstrap a strong supervised segmentation model. We apply this approach to detecting Al-Si solidification fronts in extremely low-contrast optical imaging conditions, showing significant improvements in segmentation performance and robustness. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Characterization, Computational Materials Science & Engineering, Solidification |