Geometry-aware localization evaluation of grad-CAM for wafer map defect classification

This study investigated the localization accuracy of Grad-CAM as an Explainable AI (XAI) technique for industrial inspection, specifically evaluating it on the WM811K semiconductor wafer defect dataset - a large-scale, publicly available collection of 25,519 labeled wafer maps. Localization performance was measured using the Top-10% Intersection over Union (IoU) metric on a held-out test sample ([Formula: see text]), yielding a baseline accuracy of 0.129. A consistent radial attribution bias was observed across both Grad-CAM and XGrad-CAM, with both methods producing structurally similar bias profiles, suggesting the bias reflects a property of the model's learned representations rather than a gradient-weighting artifact. To address the radial attribution bias, the authors implemented a radial suppression method, which improved the average Top-10% IoU by 19.4% (from 0.129 to 0.155), verified through a two-tailed paired t-test ([Formula: see text]). Class-level analysis revealed significant improvements for edge-ring defects while other defect types remained stable. Finally, an oracle-based refinement demonstrated that geometric representation characteristics account for a substantial portion of residual localization errors, establishing an upper bound of 0.278 IoU.