This work proposes a semi-supervised deep learning approach for the classification of defect patterns in Wafer Bin Maps (WBM). In semiconductor manufacturing, wafer is a thin slice of semiconductor substance used for fabricating integrated circuits. After fabrication, each wafer is tested to ensure that it is defect-free. A WBM is a two-dimensional grid showing the test results. Correct classification of WBM defect patterns can help to identify the root causes, and hence, the overall manufacturing process. Manual identification of these defects by a human expert is expensive and time consuming. Therefore, various automated approaches have been proposed; Most of them use supervised approaches, which require labeled data for training. As obtaining large amount of labeled data for training is a tedious process, some attempts were explored to use semi-supervised learning, which make use of the available unlabeled data in addition to the labeled data for improving the classification performance. We propose a semi-supervised Convolutional Neural Network (CNN) based approach for the classification of WBM. In our approach each unlabeled sample is weighted based on how well it is predicted, and the CNN is updated in an end-to-end manner by using both of these weighted samples and the original labeled data. Unlike other semi-supervised approaches proposed for WBM, we use sample weighting, and apply label-smoothing to reduce the number of samples which are incorrectly classified with high confident. We show improved classification performance by our proposed approach compared to the considered baselines, for example, when only 10% of the label data is considered for training, our approach gives an F1 score of 0.86 compared to the F1 score of 0.84 obtained by its fully supervised counterpart.

Please refer the attached file for more details and the results.