Segmentation approaches in niche materials domains, such as in nuclear applications, can be hindered by limited data for neural network-based methodologies and inconsistent manual or algorithmic methods. In contrast, foundation models like Meta’s Segment Anything Model (SAM) offer potential zero-shot solutions. In this study, we evaluate SAM’s performance on three publicly available datasets relevant to nuclear materials research, spanning multiple microstructural features (Pt nanoparticles, dislocation loops with black dots, and cavities) that were imaged with varying electron microscopy modalities. Results from the SAM ViT-H model show a systematic tendency toward over-segmentation, which creates an opportunity for domain-informed segmentation filter development. Dataset-specific post-processing , including morphological filtering for complex features, enables segmentation results with F1 scores between 0.63 and 0.86, underscoring the role of domain knowledge in optimizing the segmentation transferability of this foundation model. This work identifies the opportunities and challenges of applying foundation models to microstructural image analysis in nuclear materials specifically and recommends the development of standardized filtering workflows to support broader community adoption. • SAM shows dataset-specific microstructural feature detection without fine-tuning. • SAM over-segmentation is corrected with post-processing and morphology filtering. • SAM transferability is assessed across electron microscopy imaging modalities and feature type. • SAM segmentation quality, with or without post-processing, affects property analysis.
Sibley et al. (Mon,) studied this question.