Semiconductor defect classification by using cascaded convolutional neural network

The semiconductor industry is one of the rapidly growing industries globally. To be on par in both quality and quantity, most of the semiconductor industry has an inspection procedure in line with every production process. The dataset defect classes’ main features can be group into two main groups base on their pattern similarity. Previously, most of the studies did not consider the pattern similarity and fed all the input into one classifier, increasing the risk of misclassification into more classes. One classifier does not feed all. Therefore, this study proposed a cascaded convolutional neural network classifier framework for semiconductor defects to enhance the defect pattern analysis. The cascaded convolutional neural network consists of multiple models trained based on various pre-process inputs. The data pre-processing includes a median filter method by using the OpenCV library. The input has to be pre-processed to distinguish the main feature, and the model can classify it based on those features. Comparing the proposed method with the single classifier show a slight increase in accuracy from 86.7% to 87.7%. In the future, further investigation will be carried out, especially on certain defect wafer compromise of multiple defect classes and how does it reduce the capability of the cascaded convolutional neural network model.