Using GAN to Improve CNN Performance of Wafer Map Defect Type Classification : Yield Enhancement

Abstract

Semiconductor wafer map data provides valuable information for semiconductor engineers. Correctly classified defect patterns in wafer maps can increase semiconductor productivity. Convolutional Neural Networks (CNN) achieved excellent performance on computer vision and were frequently used method in wafer map classification. The CNN-based classifier of the wafer map defect pattern requires a sufficiently large training set to ensure high performance. However, for the real semiconductor production environment, it is challenging to collect various defect patterns enough. In this paper, we propose a method to supplement the lack of training set using Generative Adversarial Networks (GAN) to improve the performance of the classifier. We measure our performance on the `WM-811k' dataset, which consists of 811K real-world wafer maps. We compare the performance of our classifiers with commonly used augmentation techniques. As a result, we achieved remarkable performance enhancement from 97.0% to 98.3%.