Deep Learning-Based Silicon Wafer Defect Classification: A Performance Comparison of Pretrained Networks

Semiconductor processing technology heavily relies on defect inspection to enhance yield by identifying surface defects in the manufacturing process. However, manual inspection is prone to errors and can be a tedious process, which necessitates automated methods to replace human eyes. Deep learning techniques, such as convolutional neural networks (CNNs), are promising for automated wafer defect classification. In this study, a comparative analysis is performed on different pretrained deep learning networks to identify the most accurate and efficient network architecture for wafer defect classification. Five pretrained deep learning models, including GoogleNet, MobileNet-v2, ResNet-18, ResNet-50, and ShuffleNet, are trained and evaluated. Simulation results show that MobileNet-v2 outperforms four other pretrained networks in terms of accuracy, recall, precision, and F1-score values. The findings observed from current study can provide useful insights into the effectiveness of pretrained deep learning networks in wafer defect classification. It is believed that this study can be beneficial for manufacturing companies to improve the quality control process of silicon wafer production, leading to higher yield and better product quality.