A variational autoencoder for a semiconductor fault detection model robust to process drift due to incomplete maintenance

Abstract

In the semiconductor manufacturing field, few studies on fault detection (FD) models have considered process drift due to incomplete maintenance. Process drift refers to the shift in sensor measurements over time due to tool aging, and it leads to defective production when it is severe. Tool maintenance is conducted regularly to prevent defects. However, when it is performed improperly, tool aging accelerates, and the drift increases. In this paper, we propose an FD model robust to process drift by modeling process drift with a variational autoencoder (VAE). Because process drift is characterized by time-varying information, the proposed model encodes some time-varying information through separate hidden layers. By adopting a strategy that combines information separately encoded in a feature vector, the proposed model successfully models process drift. With actual chemical vapor deposition process data, we were able to generate many virtual datasets that incorporate process drift with various drift characteristics, such as patterns, degrees, and speeds. The proposed model outperformed four comparison FD methods on these datasets.