A hybrid system for detecting semiconductor wafer defects using modified MobileNet with multi-head attention

Identifying defective semiconductor wafers is a crucial and complicated aspect of the manufacturing process. Developed systems often confront difficulties in capturing intricate defect patterns, as well as the long-range relationships between defects and small defects within wafers. To address these issues, this paper proposes a hybrid system that combines a modified MobileNet structure with an error-correcting output code (ECOC)- based support vector machine (SVM) classifier. In the proposed method, for feature extraction purposes, the MobileNet architecture is modified 1) by using the Swish function instead of the ReLU function in the depth-wise separable convolution block of the MobileNet structure for detecting complex defective patterns and 2) by incorporating the multi-head attention mechanism in the MobileNet architecture for capturing long-range dependencies on defective wafers. To solve the dataset class imbalance issue, the ECOC-SVM classifier is used in the proposed method. For enhancing the visibility of minor defects, the histogram equalization technique is applied to wafer images. The real-world semiconductor wafer dataset WM-811K is used in this study. This paper presents an ablation study to validate the necessity of each modification to the proposed system. The proposed system achieved a superior testing accuracy of 98.55% and better average values of AUC (99.74%), recall (93.34%), precision (95.64%), and F1-score (94.42%) compared with the original version of MobileNet. Also, Comparative analyses of the proposed system with other developed systems and state-of-the-art models are given in this paper.