Robust process diagnosis using explainable CNNs for highly imbalanced wafer map defect classification in semiconductor manufacturing

Wafer defect pattern classification is essential for yield optimisation in semiconductor manufacturing. However, many existing studies evaluate models under modified class distributions or simplified experimental settings, providing limited evidence of deployment reliability under the severe imbalance observed in production data. This study presents a reproducible CNN-based evaluation framework using all 172,950 labelled samples in the WM-811K dataset, including the dominant normal-pattern class, while preserving the original class distributions of the validation and test sets. Minority-class learning is enhanced through defect-class augmentation and class-weighted focal loss, with the class weights determined by sensitivity analysis. Across 30 independent stratified data partitions, the proposed framework achieved a macro-F1 score of 0.882±0.009, improving Scratch recall from 0.388 to 0.725 relative to the baseline while maintaining stable performance on the normal class. t-SNE and Grad-CAM analyses further indicated that the remaining errors were concentrated in visually subtle patterns, rather than arising solely from majority-class bias. These results support reliable wafer defect classification under realistic manufacturing conditions.