• 목록

Identifying abnormal wafers and analyzing their causes in the semiconductor wafer manufacturing process are essential for improving production quality. Traditional methods typically check and confirm abnormalities through metrology or yield analysis after multiple processes are completed, often delaying issue resolution and causing significant financial losses. To address this limitation, modern quality control systems leverage sensor data from manufacturing equipment to detect abnormalities in real-time and enable immediate responses. However, despite advancements in automation, the task of identifying abnormal wafers based on visualized charts still heavily relies on manual labor, underscoring the need for further automation. This study introduces a quality control framework for automating abnormal wafer identification using object detection methods. A total of 1,725 images were collected, categorized, and labeled them into three types depending on the abnormal trends seen on the images. The three types are Average, Deviation, and Drift. These labeled images were used to fine-tune various object detection models, including Faster Region-based Convolutional Neural Network (Faster R-CNN), You Only Look Once (YOLO), and DEtection TRansformer (DETR)-based models. Among these, the Detection Transformers with Assignment (DETA) model, using a Swin Transformer (Swin-T) backbone, achieved the highest performance with an mAP (0.5–0.95) of 0.456. This study demonstrates that object detection models can be effectively applied in industrial fields such as semiconductor manufacturing, establishing a basis for fully automated quality control systems. Furthermore, it presents the need for developing new evaluation metrics aligned with industrial requirements and designing learning techniques suitable for small, domain-specific datasets, thereby bridging the gap between academia and industry.