Improving Identification of Defective Wafer Maps by Data Augmentation via Enhanced CycleGAN

In integrated circuit manufacturing, a wafer map represents a pattern of defective dies or chips on a wafer. In order to identify different defective wafer maps or patterns by a deep learning model, it is essential that a balanced number of samples of different patterns of wafer maps is used to train the model. In this paper, the Enhanced CycleGAN generative network is used to generate realistic defective wafer maps for which adequate numbers of samples do not exist. The identification of defective wafer maps is then carried out without and with such a data augmentation. Based on the public domain wafer map dataset of WM-811K, it is shown that our data augmentation improves the overall identification accuracy by 6% and 38% for the two test cases examined compared to the no data augmentation cases.