Automatic Wafer Defect Classification Based on Decision Tree of Deep Neural Network

Abstract

The most widely adopted approach for defect analysis in the semiconductor manufacturing plant (fab) is the automatic defect classification (ADC) that uses images taken by optical microscopy or scanning electron microscopy (SEM) to classify defects. The state-of-art ADC methods are based on Convolutional Neural Network (CNN) but are expensive in revising or expanding defect categories, and low in classification accuracy. In this paper, we propose a novel method for ADC based on Deep Neural Network (DNN) with two innovations. 1) We use a decision tree of DNNs to classify each image into successively refined categories. In contrast to a single CNN/DNN, the benefit of a decision tree of DNNs is that the latter is significantly smaller in total size and faster in training time. 2) We create a mechanism of self-learning by reporting images whose classification confidences are below a threshold as “Unknown”. Once the unknown images are manually labeled, the cases are sent back for a quick re-training. This is possible since the decision tree of DNNs permits the re-training of one or a few DNNs instead of the entire system. Experiment results show that the proposed approach achieves 100% classification accuracy, in which 2% are classified as “Unknown” and require manual classification which will be used to re-train the DNNs. The re-training time of our ADC based on decision tree DNNs is about 60 times faster than ADCs based on a single CNN/DNN.