A Reliable Defect Detection Method for Patterned Wafer Image Using Convolutional Neural Networks with the Transfer Learning

Abstract

In the semiconductor manufacturing process, wafer inspection images have valuable information on defects and wafer yield. It is worthy to analysis these images and there are many algorithms for this. Whenever new product designs are introduced, the algorithms are used to detect defects. Most new product designs have new types of defects. Hence, it is important to update the algorithms to cope with new types of defects. To update the algorithms, engineers collect the data and adjust a parameter values, such as threshold, to detect defects. It is time-consuming to collect enough data and only a knowledgeable engineer can find an appropriate parameter value. However, there is always a lack of engineer resources and time in the manufacturing industry. Due to these reasons, many algorithms can't be used reliably and have disappeared. Therefore, we propose the advanced method to update an algorithm easily using a deep learning model. Deep learning models achieve high accuracy with less engineer knowledge than traditional algorithms. But it also needs a lot of data to train the model, so we apply an advanced method of updating with a small amount of data in a short time. This is called as transfer learning. Once we train a deep learning model with high accuracy, we can update the model with a small amount of data. Transfer learning uses the information gained during the training. As a result, we make a model that is easy to update with high accuracy. In the experiments, we obtained 99% accuracy with sufficient training data. To update the model for the new data, we did transfer learning and got 97% accuracy. It was lower, but we only used 10% of the training data and 5% of the training time. Also, the accuracy we have obtained is enough to be used for defect analysis. Our approach outperformed and executed faster than traditional algorithms.