Wafer map defect recognition with few shot learning based on hybrid self-attention mechanism and prototype network

Semiconductor manufacturing is a very complex dynamic manufacturing process. Wafer manufacturing is the most critical part of it. Abnormalities in any process in wafer production may cause wafer defects. Accurately identifying various defect patterns on the wafer surface can help find and adjust abnormal factors in the online manufacturing process and improve the yield of integrated circuit production. Aiming at the very small number of wafer defect patterns that appear in the wafer manufacturing process, this paper proposes a small sample wafer defect pattern classification method based on a hybrid self-attention mechanism and a prototype network. This method introduces a hybrid self-attention module into the feature embedding network that helps to automatically learn the global dependence of channels and positions in the wafer map to obtain more discriminative features. The metric classifier is used for classification to realize the pattern recognition of wafer defects in small samples. A large number of experiments were conducted on the WM-811K data set and the MixedWM38 data set. The recognition accuracy of this method in the 3-way 1-shot, 3-way 5-shot, 5-way 1-shot and 5-way 5-shot tasks achieved average classification accuracy rates of 89.99%, 94.96%, 85.77%, and 92.32%, respectively. The experimental results show that compared with the traditional deep learning method, the method in this paper can learn new defect pattern features and realize automatic classification only through a small number of wafer samples.