Physical World & Real-World ImpactKrishnamoorthy et al. (multi-institution: India, Malaysia, Iraq)
Self-supervised multi-resolution learning of thin-film SEM defect morphology
Learned representations of thin-film SEM defect morphology with self-supervised multi-resolution learning — no defect labels used in training at all. By the authors' measurement it had the best silhouette score, 0.50, but its NMI, which measures agreement with the actual classes, was 0.203, lower than classic GLCM texture features (0.21–0.29)
Why it matters
Grounds for a workflow that clusters defect morphologies with zero labels and picks only representative samples for labelling. At the same time, a counterexample showing that internal clustering metrics and agreement with the actual classes can diverge
Source — Primary sources first; secondary reports are labelled “Press”
Papernature.comhttps://www.nature.com/articles/s41598-026-46947-3Added in Week 35, 2026 · See that week