Unsupervised Learning for Wafer Surface Defect Pattern Recognition

With the development of the semiconductor industry, the demand for wafer production has gradually increased. Wafer manufacturing is a very complicated process, and any abnormal fluctuations in each process in this process may cause surface defects on the wafer. Accurate and rapid identification of wafer defect patterns can promptly reflect abnormal problems in the production process. Aiming at such tasks, this paper proposes a pattern recognition method for wafer surface defects based on unsupervised learning. The method is divided into two parts: unsupervised pre-training and classification fine-tuning. In the unsupervised pre-training stage, in order to improve the model’s surface defect feature extraction capability, we propose a new wafer-oriented unsupervised sampling method (WaUSM). Our wafer surface defect pattern recognition method uses an unsupervised pre-training model as the initialization to set up a classification model downstream, and fine-tune the classification model through a small number of label images with wafer defect patterns, so that the classification model can more accurately identify various types of defects pattern. The method we proposed was successfully verified by wafer defect pattern recognition on the public dataset WM-811K, and the experimental results fully proved the effectiveness and industrial applicability of the method we proposed.