Automated Morphology Analysis for Nuclear Materials
Self-supervised and semi-supervised approaches for segmentation and characterization of microstructures in nuclear fuel materials from electron microscopy.
Morphological structure provides a valuable signature for nuclear materials because it reflects the chemistry and processing history that produced the material. In this project, we develop machine learning methods that turn SEM images into quantitative, reproducible evidence for nuclear forensics. Our work began with automated particle segmentation and morphology measurement, reducing the need for labor-intensive manual annotation, and has expanded toward learned representations that classify uranium oxide pathways directly from image data. The broader aim is to make morphology analysis faster, more scalable, and more reliable for identifying provenance and processing route.
Technically, the project combines segmentation, multi-scale representation learning, and robustness to domain shift. We use deep segmentation models to isolate particles for downstream measurement, multi-input CNNs to integrate SEM images across several magnifications, contrastive self-supervised learning to learn transferable morphology-aware encodings, and source-free domain adaptation to handle variation across microscopes and acquisition settings. Together, these methods support automated morphology analysis that is both accurate and practical for real-world nuclear materials applications.
