CowSSL: contrastive open-world semi-supervised learning for wafer bin map

In the semiconductor industry, wafer bin maps (WBMs) refer to image data that reveal the defect of each chip positioned on that wafer. The WBMs provide crucial information that can facilitate the identification of underlying causes for any defects present on a wafer. With the advent of artificial intelligence (AI), a significant amount of research has been conducted leveraging machine learning and deep learning techniques to automatically classify wafer bin map defects. Although there have been various attempts to enhance performance by using both unlabeled and labeled data, current research is constrained by its narrow focus on improving the detection of known defect patterns. However, in the real world, multiple novel patterns frequently arise that have not been previously encountered. Hence, AI models must exhibit the capacity to detect not only existing known defect patterns but also newly emerging defect patterns, while ensuring effective classification of these new patterns among themselves. In this study, we propose the contrastive open-world semi-supervised learning that can classify multiple novel patterns in WBMs simultaneously. We introduce a contrastive loss function to address the challenges associated with the existence of significantly fewer new defect patterns than existing patterns in the WBM problem. We confirm that the proposed methodology effectively detects and classifies diverse new patterns separately in real-world open data, WM-811 K. Moreover, we demonstrate that the proposed method outperforms other existing open-world semi-supervised learning in WBM classification.