Wafer Defect Classification Algorithm With Label Embedding Using Contrastive Learning

Abstract

Classifying wafer defects in the wafer manufacturing process is increasingly critical for ensuring high-quality production, optimizing processes, and reducing costs. Most existing methods for wafer map defect classification primarily rely on images alone for model training and prediction. However, these approaches often lack interpretability, which can hinder process improvement and problem-solving efforts. In other words, existing methods only calculate the probability of a specific image belonging to each class, making it difficult to visually judge why the image belongs to a particular class. Additionally, these methods make it challenging to assess the distance of new images from each class. Furthermore, it is difficult to obtain representative images of each class. To address these limitations, we propose a novel approach for wafer defect classification using contrastive learning with label embedding. The proposed method aims to map label information and wafer defect images into a shared latent space through contrastive learning using label embedding. This not only facilitates defect class prediction from images but also enhances interpretability by visualizing relationships between images and defects (labels) and providing representative defect images. Moreover, compared to previous methods, our approach demonstrates better classification performance and computational efficiency, even in situations with imbalanced labels. This method also shows significant potential in identifying unseen defects not defined in the original classification tasks. Consequently, the proposed approach extends its applicability beyond wafer map defect patterns, showing promising potential for use in various domains.