AI-assisted anomaly detection from SEM images in semiconductor fabrication

Scanning Electron Microscopy (SEM) allows for the detection of defects in semiconductor processes. However, the introduction of new semiconductor processes inevitably leads to the emergence of new types of defects. These defects demand time-consuming detection process. Deep learning-based classification models have gained attention for defect inspection, but these models are trained on existing defect types and therefore cannot classify novel defect types. Conventional anomaly detection models are designed to identify abnormal cases, but they are generally trained only on normal data with relatively simple patterns. To overcome these challenges, we propose a multi-task hybrid transformer that leverages multi-task learning (MTL) and an Efficient Self-Attention Module (ESAM) to detect novel anomalies from existing defect types in SEM images. By performing classification on existing defect types, network learns to form compact and well-separated clusters across diverse existing defects, thereby improving its ability to distinguish novel patterns. ESAM enhances feature representation by capturing both local and global contexts while maintaining computational efficiency. Evaluated on SEM defect datasets, our proposed model outperformed six anomaly detection models, achieving the highest AUC and real-time inference speed. These results demonstrate the practical potential of our approach as a reliable and efficient solution for robust defect detection in semiconductor manufacturing environments.