Precise segmentation is fundamental to quantitative microstructural analysis, a cornerstone of establishing processing-structure-property relationships in materials science. However, the generalization of deep learning models remains a key challenge, with performance often degrading significantly when applied to unseen material systems or imaging conditions. This limitation underscores the urgent need for a universal segmentation model. Herein, we propose SAMM, a general-purpose segmentation model for material micrographs, developed by fine-tuning the Segment Anything Model 2 (SAM2) on a large-scale, diverse dataset. Our approach integrates a full-parameter fine-tuning strategy, a cross-scale feature fusion module for enhanced detail resolution, and a hybrid loss function to ensure both pixel-level accuracy and structural integrity. Consequently, SAMM demonstrates superior generalization across 13 diverse material datasets. For instance, it achieves a mean Intersection-over-Union (mIoU) of 76.63% on challenging nickel-based superalloy micrographs, outperforming the next-best baseline by 10.3%. In segmenting additive manufacturing powders, SAMM delivers near-perfect results (98.13% mIoU). Crucially, SAMM exhibits exceptional zero-shot generalization, achieving up to 97.41% mIoU on datasets entirely unseen during training. This work not only presents a robust framework for universal microstructure segmentation but also provides a comprehensive, publicly available dataset to foster further research in this domain.
Tu et al. (Sun,) studied this question.