Semi-Supervised Learning Under Extreme Class Imbalance in Wafer Bin Map Defect Pattern Classification

The semiconductor industry faces significant challenges in wafer quality assessment due to the increasing complexity of device fabrication and the rise in inline defects and circuit failures. As manufacturing processes advance, conventional manual analysis of wafer bin maps (WBMs) has become impractical and unreliable, unable to keep pace with the industry’s demands. While convolutional neural networks (CNNs) are widely used in automated systems for classifying WBM defect patterns, their dependence on supervised learning limits their effectiveness, as they cannot utilize the vast amounts of unlabeled WBM data. Additionally, class imbalances in WBM datasets often result in models biased toward majority classes, leading to suboptimal performance. In this study, we produce ClimEx, a semi-supervised learning framework that addresses the extreme class imbalance problem commonly found in WBM datasets through consistency regularization and pseudo-labeling. Our method maintains dynamic per-class thresholds to correct the imbalance in pseudo-labels caused by biased predictions, ensuring balanced representation across classes even under extreme imbalance. On a publicly available benchmark, ClimEx demonstrates improved predictive performance compared to state-of-the-art methods. We also provide a comprehensive analysis of the key components of our method from multiple perspectives.