An Analog Wafer Map Clustering Model with Deep-Learning Based Data Augmentation and Feature Extraction

Abstract

Process deviations during semiconductor manufacturing have significant influence on wafer yield. Tasks such as failure type recognition, outlier detection and etc. can be accomplished by checking wafer maps. In this study, unsupervised feature clustering on unlabeled analog wafer maps is our primary concern. Each wafer map indicates a test parameter and the task is finding cluster of parameters sharing a similar pattern. Convolutional variational auto-encoder is used to extract features and the features are clustered by density-based algorithm to replace K-means in previous works. Besides, rather than supervised data augmentation such as policy-search system, we construct a deep convolutional generative adversarial network for unsupervised data augmentation. Performance of the study is tested on real-world and synthetic wafer maps and the proposed method improves silhouette coefficient by 31 percent compared to previous studies.