Abstract: Microstructure, the internal arrangement of a material’s components, is a key determinant of its performance. Accurate microstructure analysis is essential for understanding material properties and identifying defects. Traditional manual methods are often time-consuming, while current machine learning models for microstructure segmentation require extensive labeled data and often struggle to generalize across different materials. To address these challenges, we propose the Self-annotated and Distilled U-Net (SDU-Net) framework, a versatile approach for microstructure segmentation that doesn’t require pre-existing labeled data. SDU-Net utilizes the Segment Anything Model (SAM) as a teacher model to generate initial pseudo labels for microstructure image elements through a two-step clustering process. These labels are then refined, enabling SDU-Net to learn general features and handle tasks with well-defined labels. Model distillation and preprocessing techniques within SDU-Net provide high-quality pseudo labels that guide and accelerate the training of a U-Net student model. This approach overcomes the data specificity limitations of the teacher model, significantly improving both accuracy and inference speed. By delivering precise and efficient segmentation across diverse microstructure tasks, SDU-Net offers a new paradigm for using foundation model distillation in image segmentation and demonstrates immense potential for advancing material characterization and personalized microstructural analysis.

Key words: Deep learning, Microstructural analysis, Artificial intelligence, Image segmentation, Electron microscopy