Vision-based wafer inspection in semiconductor manufacturing: a case study on scratch defect detection using synthetic data and YOLO models

Scratch-type surface defects are among the most critical in wafer fabrication. Even minor scratches can result in rejected dies, substantial yield loss, and downstream reliability failures, making their early and precise detection essential for cost-effective production and long-term device performance. Given the scale and speed of modern semiconductor manufacturing, manual or traditional inspection methods are insufficient to reliably detect such subtle surface defects, underscoring the urgent need for intelligent, automated inspection systems capable of real-time, high-precision detection. However, developing such systems remains challenged by limited access to diverse, labeled defect datasets and the high cost of manual annotation. To overcome this, we introduce a novel, fully synthetic-data-driven inspection framework that bridges virtual simulation with real-world deployment, without requiring any annotated real-world training data. A high-fidelity 3D simulation pipeline was developed using Blender to procedurally generate thousands of labeled wafer images featuring scratch-type surface defects under varying lighting, texture, and geometric configurations. Concurrently, a physical optical inspection system was constructed to capture wafer images under controlled conditions, serving as the experimental validation platform. YOLOv8m, YOLOv10m, and YOLOv11m object detection models were trained exclusively on the synthetic dataset and directly evaluated on real wafer images obtained from the custom inspection setup. Results demonstrated that YOLOv8m and YOLOv11m achieved detection accuracy of F1 = 1.00 in wafer presence classification and F1-scores exceeding 0.96 in scratch defect localization, indicating that models trained on synthetic images can achieve strong performance on real wafer images within the tested physical environment. This two-stage evaluation pipeline eliminates the need for costly manual annotation while preserving high detection fidelity. The proposed framework combines synthetic data generation and YOLO-based deep learning to address critical data bottlenecks in industrial defect detection.