Scientific Imaging: Materials and Connectomics

Scientific imaging domains such as electron microscopy produce exceptionally rich data, but extracting quantitative insight from these images often requires substantial expert effort. This theme focuses on machine learning and image analysis methods for microscopy-based scientific discovery across both materials science and biomedicine. On the materials side, we study morphology, texture, and multiscale structure in electron microscopy images to support characterization of complex materials and nuclear materials workflows. On the biomedical side, we develop methods for connectomics from electron microscopy, where dense ultrastructural imagery must be converted into reliable segmentations and reconstructions before neural circuits can be studied. Across both areas, the common challenge is the same: scientifically valuable images contain fine-scale structure that is too complex and too voluminous for manual analysis alone.

Methodologically, this work emphasizes scalable, structure-aware learning for difficult imaging regimes. We develop segmentation methods for dense microscopy data, representation-learning approaches that capture morphology across magnifications and imaging conditions, and robust modeling strategies that can transfer across instruments, samples, and acquisition settings. In connectomics, this includes learning from limited supervision and exploiting the hierarchical structure of segmentation problems in large EM volumes. In materials imaging, it includes automated morphology analysis, multiscale classification, and domain-adaptive modeling for robust characterization pipelines. The broader goal is to turn expert-intensive microscopy analysis into reproducible, scalable, and quantitatively reliable computation that can accelerate discovery in both the physical and biological sciences.