Semi-automatic wafer map pattern classification with convolutional neural networks

In semiconductor manufacturing, the defect patterns of wafer maps provide crucial information to identify the root causes of wafer defects. Recently, convolutional neural networks (CNNs) have been actively applied to automatic wafer map pattern classification. As a requirement for real-world application, a CNN must be as accurate as a process engineer. Existing studies have attempted to improve the training phase of a CNN to make it more accurate. However, they often fail to achieve the near-perfect accuracy requirement in practice. To sidestep the difficulty, we focus on improving the inference phase of an imperfect CNN with the aid of a process engineer. In this paper, we propose a semi-automatic wafer map pattern classification method that selectively utilizes the CNN for classifying new wafer maps. Given a query wafer map, we decide whether to use the CNN by quantifying its predictive uncertainty on the wafer map. If the predictive uncertainty is sufficiently low, the wafer map is classified using the CNN. Otherwise, the wafer map is subject to manual classification by a process engineer. In the experiments using the WM-811k dataset, the proposed method attains an accuracy of over 99% with a CNN coverage of 93%.