Reference-less defect inspection method based on unsupervised deep learning for semiconductors

In advanced semiconductor manufacturing, the importance of high-sensitivity inspections based on scanning electron microscopy (SEM) has increased for yield improvement. The sensitivity of SEM-based inspections generally improves as the Signal-to-noise ratio (SNR) increases. However, there is a trade-off between the SNR and imaging throughput of SEM images. To improve inspection throughput, improving the robustness of inspection methods to noise is an important issue. We propose the reference-less defect inspection method based on deep learning to improve robustness to noise. The proposed method detects defects by comparing the inspection image with an average estimated reference image improved SNR by averaging estimated reference images which multiple-reference estimation model estimates from an inspection image. The results of evaluation utilizing low-SNR SEM image datasets of semiconductor device samples shows that the proposed method enables high-accuracy inspection even for low-SNR images.