A Self-Supervised Foundation Model for General-Purpose Microstructure Representation from EBSD Data

Electron backscatter diffraction (EBSD) provides spatially resolved crystallographic orientation maps encoding a wealth of microstructural information, including grain morphology, texture, phase distribution, and deformation state. However, the dominant analysis paradigm still relies on hand-crafted descriptors that discard most of this spatial and orientational context. We address this limitation by training a self-supervised foundation model directly on raw EBSD orientation data, producing dense, transferable representations without requiring microstructural labels. Our approach adapts a large-scale vision foundation model architecture, initialized from natural image weights and fine-tuned on a corpus of EBSD patches, enabling effective representation learning with a relatively small domain-specific dataset. The frozen backbone produces per-patch embeddings that support a broad range of downstream tasks via lightweight probes. Quantitative analysis demonstrates the model’s effectiveness in predicting microstructural characteristics such as Kernel Average Misorientation (KAM), Grain Average Misorientation (GAM), and Grain Orientation Spread (GOS). Notably, this approach outperforms supervised CNN benchmarks, particularly in data-constrained regimes. Qualitative analysis of attention maps and patch embeddings reveals that the model emergently develops sensitivity to grain boundaries, orientation gradients, and defects. The framework performs well on out-of-distribution data, demonstrating robustness across diverse material systems. This positions the approach as a material-agnostic, general-purpose feature extractor for quantitative microstructure analysis, with potential applications spanning alloy design, deformation monitoring, and automated materials informatics pipelines.