Pattern Classification for Small-Sized Defects Using Multi-Head CNN in Semiconductor Manufacturing

Abstract

To improve the quality of semiconductor manufacturing, defects need to be detected and their root causes controlled. Because the root causes can vary depending on defect patterns, classifying the patterns accurately is important. Several recent studies have investigated automatic defect classification using a convolutional neural network (CNN) with wafer map images. CNNs are excellent tools for classifying images of different shapes and sizes. However, the detection of small-sized defects that have small clusters and linear patterns is difficult. Therefore, this study focuses on patterns that are difficult to detect. We propose three steps for pattern classification. First, modified median filtering is used to preserve the original shapes of patterns. Second, a rotated defects (RoD) transform is performed by applying the rotational properties of wafer maps. The RoD transform augments the defect proportion and improves the detection of small-sized defects. Third, a multi-head CNN is used to extract and combine the features from the original and transformed maps. The combined features are then used to classify the defect patterns. Overall classification performance of defects can be improved by accurately classifying small clusters and linear patterns. The proposed model was evaluated using WM-811K wafer maps, and small-sized defects were accurately classified. Such an accurate defect classification model will enable effective root cause analysis and quality improvement in semiconductor manufacturing.