SAM- and mSAM- Based Inference of Nuclear Materials Processing History from SEM Imagery

Mayolo Valencia Mendoza, Alexei Skurikhin, Judith Cohn, Luther Mcdonald, Kari Sentz; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3614-3621

Particle and microstructure morphology of nuclear materials, as observed in scanning electron microscopy (SEM), is an emerging signature with the potential to assist with the identification of key factors of processing history. We introduce a method for inferring the calcination temperature of nuclear materials without dependence on incomplete or subjective image annotations. Leveraging foundation segmentation models, Segment Anything Model (SAM) and its microscopy-tuned variant, mSAM, we aggregate shape features from all segmented particles in each image. In contrast to previous supervised approaches that trained models to detect only a small set of particles similar to those in the labeled training set, our method exploits the statistical benefits of large sample sizes derived from all segmented particles in the image and enables the extraction of statistical summaries of particle shape characteristics that reflect underlying processing conditions. These statistical summaries are sufficiently distinctive that logistic regression can successfully categorize calcination temperature classes with high accuracy. In evaluations on a dataset of nuclear materials imagery, our approach achieved 92.1% +- 4.3% accuracy in predicting calcination temperatures categories, outperforming a previous alternative method by 8.8%.