Mixed-defect wafer map separation and detection based on single-defect wafer map

Abstract

The difficulty in labeling wafer maps of mixed-defect has affected the development of deep learning-based detection models. In the case of zero labeled samples, this paper proposes a mixed-defect wafer map separation and detection (MDWMSD) method based on single-defect wafer map. First, mixed-defect wafers are generated using different categories of single-defect wafers. Then, a mixed-defect separation model was proposed based on a residual neural network with U-net structure to separate mixed-defect wafer map into several single-defect wafer maps. Finally, the separated single-defect wafers are identified using the trained single-defect classifier. During the validation process, two mixed-defect wafer map separation models were developed using single-defect wafer maps from the MIR-WM811k dataset and the MixedWM38 dataset, respectively. The developed model was then tested on mixed-defect wafer map in the MixedWM38 dataset. The results show that the detection accuracy of mixed-defect wafer under zero samples condition can reach 95% and 85.38% in two cases, which proves the effectiveness of the proposed method.