Mixed-type wafer defect recognition using optimized YOLOv8 model

Abstract As semiconductor manufacturing continues to advance, efficient and accurate defect detection becomes increasingly vital to ensure high-quality production standards. This research investigates the application of the You Only Look Once (YOLO) model family for classifying mixed-type wafer defects during the wafer testing stage. These defects include a mixture of eight basic types: Centre (C), Donut (D), Edge-Loc (EL), Edge-Ring (ER), Loc (L), Near-Full (NF), Scratch (S), and Random (R). The novelty of this work lies in the development of a lightweight YOLO-based model specifically tailored for wafer defect classification, targeting both improved computational efficiency and high detection accuracy. This research introduces several YOLOv8 backbone optimization incorporating Depthwise Convolution (DWConv) and GhostBottleneck modules. The proposed model achieves substantial improvements over the baseline YOLOv8n model, including a reduction up to 79.40% in file size, 8.72% in MFLOPs, 81.88% in model parameters, along with an increase of up to 15.14% in inference speed. Moreover, the optimized model achieves 99.88% accuracy, 97.78% precision, 97.74% recall and 97.73% F1-score on the evaluated dataset. This work contributes to the semiconductor industry’s shift toward automated, high-throughput defect inspection systems, demonstrating that developing lightweight models does not necessarily require compromising on performance.