Wafer Map Classifier using Deep Learning for Detecting Out-of-Distribution Failure Patterns

Abstract

Pattern analysis of wafer maps in semiconductor manufacturing is critical for failure analysis aspects or activities that increase yield. As deep learning becomes more popular than ever, research on the wafer map classification is active. However, more accurate pattern classification and data processing methods are required for the accuracy of commonality analysis to find suspected facilities using wafer map classification. It is difficult to represent all types of wafer maps in dozens of forms, and the frequency of wafer map shapes that vary with yield changes also requires the processing of undefined pattern wafer map data. We define out-of-distribution data of wafer map data that does not identify in the pattern classifier and suggest a network that uses the convolutional neural network (CNN) with residual units and training methods to classify it efficiently. We employ 15,436 real wafer map data for pattern classification and detection of out-of-distribution data.