Physics-proxy self-supervised representation learning for unsupervised degradation-state clustering of palladium membrane SEM images

2026-08-06

PRE

Sangkeum Lee

Gihun Gil

Younghwan Im *

DOI：10.1016/j.measurement.2026.122725

Original

Original request for help

Save

Abstract

En 中文

• Image-computed proxies provide task-specific supervision without manual SEM-task annotations. • Unsupervised clustering yields a three-level degradation-severity structure from SEM patches. • Physics-proxy embedding recovers the clusters at 67.4% accuracy under leakage-safe GroupKFold. • Imbalance-aware metrics reported: macro-F1 0.66, per-class precision/recall, confusion matrix. • Fully reproducible pipeline with released code, fixed seed, and split manifests.

Keywords:

Self-supervised representation learning

Scanning electron microscopy

Unsupervised clustering

Degradation-severity assessment

Leakage-safe cross-validation

Palladium hydrogen separation membranes

Physics-proxy self-supervised representation learning for unsupervised degradation-state clustering of palladium membrane SEM images

Abstract

Journal

Organization

Cited Papers

Citing Papers