Learning Microstructure--Property Relationships in Materials with Robust Features from Vision Transformers

Sheila E. Whitman, Guangyu Hu, Marat I. Latypov; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8125-8130

Machine learning of microstructure--property relationships from data is an emerging approach in computational materials science. Most existing machine learning efforts focus on the development of task-specific models for each microstructure--property relationship. We propose utilizing a pre-trained foundational vision model for the extraction of task-agnostic microstructure features and subsequent light-weight machine learning. We demonstrate our approach with a pre-trained DinoV2 model on unsupervised representation of an ensemble of two-phase microstructures and modeling of their overall elastic stiffness. Our results show the potential of foundational vision models for robust microstructure representation and efficient machine learning of microstructure--property relationships without the need for expensive task-specific training or fine-tuning.