Supervised contrastive learning for wafer map pattern classification

Abstract

In the semiconductor manufacturing process, analyzing the defect patterns on a wafer map is crucial for identifying the causes of the defects. The advent of convolutional neural networks (CNNs) has significantly increased the accuracy of automated wafer map pattern classification. Generally, the use of a larger training dataset results in higher classification accuracy. However, collecting a large number of wafer maps and labeling them with their defect categories is expensive and time-consuming. In this paper, we present an improved training method under data insufficiency for wafer map pattern classification. We apply supervised contrastive learning to train a CNN by exploiting the rotational-invariant characteristic of wafer map labeling. The CNN is trained by simultaneously minimizing two loss functions: classification loss and contrastive loss. The first loss function is to classify the rotational variants of wafer maps accurately. The second loss function is to align the representation vectors for the rotational variants of wafer maps with similar labels to be close to each other. Using two benchmark datasets, WM-811K and MixedWM38, we demonstrate that the proposed method enhances classification accuracy compared with existing methods, particularly when the training dataset is small.