A novel training-free approach to efficiently extracting material microstructures via visual large model

Abstract

The precise quantitative description of material microstructures is essential for deeply exploring the relationship between material composition and property. This significant understanding efficiently enables composition design, process optimization, and property enhancement. Traditionally, the analysis of material microstructures has relied heavily on professional expertise. Even with machine /deep learning (ML/DL)-based analysis methods, substantial expert annotation is required for training, and the trained models often suffer from weak generalizability and poor recognition of new images. This study proposed MatSAM (Materials Segment Anything Model), a novel training-free approach for efficient material microstructure extraction based on the Segment Anything Model (SAM), a type of visual large model (VLM). Integrating region marking and microscopy-adapted points, an automated point-based prompt strategy was developed to achieve accurate and efficient material microstructure recognition. Without any manual annotations, MatSAM precisely identified 11 kinds of metallic material microstructures obtained through various characterization methods. Compared to optimal conventional rule-based methods that do not involve a learning process (non-ML/DL), MatSAM achieved an average relative improvement of 35.4 % in metrics combining the adjusted Rand index (ARI) and Intersection over Union (IoU), outperforming the original SAM by an average of 13.9 %. On four public microstructure segmentation datasets, the IoU of MatSAM showed an average improvement of 7.5 % over corresponding specialist deep models requiring annotations. Meanwhile, MatSAM satisfied the generalization capability of a single model for various microstructures, including grain boundaries, phases, and defects. This approach significantly reduces the labor and computational costs of quantitatively characterizing material microstructures, further accelerating the development of advanced materials.