Advancing Semiconductor Fabrication: A CNN-Based Wafer Defect Detection with XAI Insights

Semiconductor fabrication encounters substantial obstacles in defect identification, adversely affecting yield and quality of the product. Conventional manual classification techniques are ineffective, attaining accuracy rates below 45%. This study develops and evaluates deep learning classifiers, including ResNet50, a custom CNN, Inception v3, and a machine learning classifier SVM, utilizing the WM-811K dataset, which comprises nine distinct wafer defect classes, to improve classification accuracy and minimize class imbalance. ResNet50 had exceptional performance, attaining 99.93% accuracy for "Near-full" defects and 95.77% for "Local" defects. Furthermore, the incorporation of Gradient-weighted Class Activation Mapping (Grad-CAM), an Explainable Artificial Intelligence (XAI) technique, with ResNet50 enhanced model interpretability by emphasizing significant defect areas. These automated classification method provides a reliable and efficient approach for detecting defects, thereby optimizing the semiconductor manufacturing process and improving industry growth by enhancing yield and quality.