Data-efficient semantic segmentation of fracture surfaces using foundation vision models

Quantitative fractography using scanning electron microscopy (SEM) is essential for failure analysis and defect-based life assessment. However, fracture surface evaluation remains largely manual, slow, subjective, and difficult to scale. This is especially true when analyzing Chloride Stress Corrosion Cracking (CSCC) in 316L stainless steel, where distinguishing mixed-mode quasi-cleavage regions requires significant expert effort. Although recent studies have explored using machine learning to automate this process, most still rely on large, task-specific annotated datasets. In practice, producing detailed, pixel-level annotations is the biggest bottleneck in deploying these ML models. To address this bottleneck, we propose a data-efficient machine learning framework that adapts the Segment Anything Model (SAM) for automated semantic segmentation and operates reliably in strict low-data regimes, requiring only a handful of expertly annotated micrographs. We introduce a novel, expert-annotated dataset of 316L fracture surfaces and benchmark the adapted foundation model against state-of-the-art supervised architectures. Our results demonstrate that the proposed framework generalizes robustly from minimal supervision, establishing a new benchmark for few-shot learning in fractography. We show that the foundation model requires significantly fewer annotated examples to surpass the peak performance of traditional supervised networks, thereby effectively addressing the data-scarcity bottleneck. Furthermore, we conduct a systematic ablation of fine-tuning strategies, revealing that Parameter-Efficient Fine-Tuning (PEFT) via Multi-Layer Perceptron (MLP) adapters achieves superior segmentation accuracy compared to other fine-tuning techniques while significantly reducing computational overhead. By validating that a lightweight adaptation of frozen encoders yields better boundary adherence and textural understanding than fully trained baselines, this work establishes a practical path toward high-fidelity automated fractography accessible to laboratories with limited annotated data and computational resources. The code and dataset are made publicly available to support further research.