A Study on Detection Method Using 2-Class Classifiers for Defective Wafer Maps

In semiconductor manufacturing, a pattern of chips with electrical failures in the wafer is usually used to identify failure factors. Wafers with similar in-plane trends are likely to have the same defect factors, and clustering techniques are often used to identify defect factors. It is, however, difficult for clustering approaches to make a cluster of infrequent unknown patterns. As a result, it will occur missing defect patterns. We discussed the method to detect infrequent unknown patterns and accurately classify frequent known defect patterns. We tried to make the proposed scheme with three strategies. As the first approach, VGG16 and SVM were used as the feature extractor and a classifier, respectively. The second approach is Convolutional Auto Encoder (CAE). We constructed CAEs for each known class, and the CAEs were trained to reconstruct the input images. When we take the above strategy, the constructed CAE cannot reconstruct the same image when the input image does not belong to the same class. It will be helpful to judge whether the given image belongs to the same class. The third approach uses the difference degree between the given image and typical images. The calculated value of the difference is used for distinguishing using thresholds. The experimental results show that the classification accuracy of known classes is 75.9%, the detection rate of unknown classes is 62.5%, and unknown clusters containing 35.7% of unknown classes are successfully created. To improve yield, it is essential to detect unknown defects at an early stage. If we can generate clusters that are mostly composed of unknown classes, it will be possible to recognize the occurrence of unknown defects. Since the proposed method was able to generate clusters in which unknown classes account for about 35%, we believe that it is sufficient to detect the occurrence of unknown defects.