Wafer Bin Map Recognition With Autoencoder-Based Data Augmentation in Semiconductor Assembly Process

Semiconductor manufacturers use the wafer bin map recognition (WBMR) system to identify failure modes in processing. This study proposes an WBMR system embedded with three modules: data preprocessing, region classification, and systematic pattern recognition. After using a revised Jaccard index to separate random patterns from systematic patterns, we compare three data augmentation techniques, particularly autoencoder-based, to find the best augmented method that addresses any data imbalance problems between the defect classes. We propose an adaptive algorithm to determine the amount of generated data. We describe the two tools, t-distributed stochastic neighbor embedding (t-SNE) and earth mover's distances (EMD) we use to quantify and visualize the information content of the augmented dataset. Finally, we use an inception architecture of convolutional neural network (CNN) to improve the WBMR system's recognition accuracy. An empirical study of the semiconductor assembly manufacturer and a public dataset validate that our proposed WBMR system effectively recognizes different types of defective patterns.