Lawrence Berkeley National Laboratory

Abstract REBCO coated conductors are of significant interest for high-field superconducting applications owing to their exceptional critical current retention under high magnetic fields. However, depending on the fabrication route, microstructural inhomogeneities such as porosity, Cu x O precipitates, and a-axis oriented grains can emerge at various length scales, disrupting current transport and limiting the critical current density. This study investigates and characterizes such defects in commercial REBCO conductors of varying specifications. Top-view SEM images of the REBCO layer were acquired following chemical etching of Cu and Ag layers to expose the microstructure for analysis. Conventional image analysis and segmentation techniques prove insufficient for reliably quantifying these defects, while manual identification remains prohibitively labor-intensive. To overcome these limitations, a machine learning approach is explored to enable rapid, automated, and accurate defect detection. Specifically, an open-source computer vision model, Mask R-CNN, is fine-tuned on domain-specific SEM image data. The fine-tuned model achieved a validation mean average precision of 45.23% and enabled quantitative defect analysis across 63 tapes. Partial correlation analysis, controlling for confounding between defect types, revealed independent associations with critical current density that varied in strength and sign across defect types and measurement conditions. These findings motivate further targeted characterization to establish the microstructural origins of these relationships and inform conductor optimization.