Open set recognition of wafer bin map defect patterns

Abstract

The wafer bin map (WBM) is a spatial map expressed according to a pass/fail verification test for a chip on a wafer. There are various studies that automatically classify defect patterns using machine learning to identify the causes of process defects. However, as the processes for producing high-quality semiconductors evolve and become more complex, the likelihood of new defects and defect patterns occurring during the process increases. Because of this, the general machine learning assumption that the information in the training data is sufficient for real-world situations may not be satisfied, so existing research may not be suitable for real-world processes. In order to design a machine learning classifier that is more suitable for real-world processes, we use an open set recognition methodology to design a classifier that considers all possible combinations of patterns. This is because classifiers designed with supervised learning focus on extracting information for classification from the data, whereas unsupervised learning can be combined to extract additional information that the classifier does not consider. By using the added information, it is possible to utilize the property that the reconstruction model has higher reconstruction errors of new patterns than that of trained patterns. However, there is a phenomenon that caused this property to not be satisfied in the case of WBM data. This phenomenon is mainly caused by two reasons. First, the reconstruction model generalizes well. To solve this problem, a memory module is added to the reconstruction model to make the extracted features closer to the features of the training pattern, further increasing the reconstruction error of the new patterns. Another reason is that the reconstruction errors are greatly affected by the type of defect pattern. In this case, in order to recognize a new defect pattern, the defect error according to the type of pattern was compared and solved using the information obtained from the prediction information of the classifier. In this work, based on the classification-reconstruction-based model, the proposed model is developed to perform open set recognition of WBM defect patterns. In addition, it is a single model with an end-to-end method and makes no assumptions about new defect patterns during training. Therefore, unlike the methodology of making a classifier according to each pattern, it can be said that it is a methodology suitable for the actual manufacturing process even if the number of new defect patterns continues to increase.