Imbalanced Wafer Map Dataset Classification with Semi-Supervised Learning Method and Optimized Loss Function

Abstract

Wafer is the crucial raw material of semiconductor devices. In wafer production, impurities cannot be removed entirely, which will cause various wafer map defects. Quickly and precisely classifying wafer map defects can help engineers track failures in the semiconductor manufacturing process. However, different wafer map defect patterns occur randomly and irregularly, and labeling work is labor-intensive and time-consuming. Therefore, the wafer map dataset is usually imbalanced and consists of many unlabeled data. In this paper, we utilize unlabeled data by using semi-supervised learning methods and alleviate the imbalance problem by optimizing the loss function to increase the accuracy of wafer map classification. The performance of the proposed method is illustrated with the WM-811K dataset which consists of real-world wafer maps.