Explainable Wafer Map Defect Classification Using Lightweight Data Augmentation and Convolutional Neural Network

Wafer map defect classification is a crucial task in semiconductor manufacturing, as early detection of defects improves yield and minimizes production waste. This study proposes an enhanced wafer defect classification framework that integrates a lightweight convolutional neural network (CNN) with data augmentation techniques to improve performance preserving computational effectiveness. The model addresses class imbalance problems in the WM811k dataset by using a convolutional autoencoder for synthetic wafer sample generation. A tailored CNN architecture is designed to extract hierarchical spatial features, ensuring robust classification across nine types of defects. Furthermore, explainability is a key aspect of this work, achieved through Gradient-weighted Class Activation Mapping (Grad-CAM), Integrated Gradient, and Occlusion Sensitivity. These techniques provide visual insights into the model’s decision-making process, fostering transparency and reliability in industrial applications. Experimental results demonstrate that the proposed approach outperforms state-of-the-art models with an outstanding classification accuracy of 99.87%. The combination of a lightweight CNN, data augmentation, and explainability techniques makes this method highly suitable for real-time defect detection in semiconductor manufacturing.