Date of Publication :11th August 2023

Abstract: The detection of defects in semiconductor wafers is of paramount importance in guaranteeing the quality and dependability of silicon wafers. However, AI based detection of defects in wafers is challenging due to imbalanced datasets, where the number of defective samples is considerably lower than non-defective wafer bin maps. The study introduces SM-CapsNet, an approach that tackles the issue of class imbalance by utilizing SMOTE (Synthetic Minority Over-sampling Technique) alongside a modified capsule network. This combination enhances the classification model's ability to localize defects effectively. The methodology is evaluated on the widely used WM-811K dataset, which contains bin-map images of semiconductor wafers captured under various conditions. The experiments show that the proposed method outperformed standard neural networks based on accuracy and F1-score metrics. It is also noted that using SMOTE for data generation requires less time and resources compared to traditional data augmentation techniques like GANs and CAEs. Thus, SM-CapsNet depicts potential for accurate semiconductor wafer defect detection to improve silicon wafer production yields in the in the semiconductor industry.

Reference :