Unsupervised Estimation of Lower Layer Pattern Regions Based on High Voltage-Scanning Electron Microscope Image Characteristics

As the number of semiconductor layers has been increasing to save power and improve performance, the overlay measurement for pattern misalignment between layers has become more critical. High voltage-scanning electron microscopes (HV-SEMs) measure overlay by simultaneously capturing secondary electron (SE) and backscattered electron (BSE) images. Here, the SE images reflect information about the upper layer patterns, and the BSE images reflect information about both the upper and lower layer patterns. Some conventional overlay measurements estimate upper layer pattern regions (ULPRs) and lower layer pattern regions (LLPRs), and compute their shifts. Supervised segmentation is accurate but requires labor-intensive manual annotation, making unsupervised methods important for semiconductor manufacturing. However, existing unsupervised segmentation methods fail to estimate the LLPRs when the ULPRs and LLPRs have similar brightness in a BSE image. To solve this issue, we propose an unsupervised method that can estimate only the LLPRs by leveraging the HV-SEM characteristics. To estimate only the LLPRs, our method predicts only the ULPRs in the BSE image and subtracts them from the BSE image. The evaluation results showed that our method estimated the LLPRs more accurately than existing unsupervised methods. Furthermore, its overlay measurement accuracy was comparable to a supervised segmentation method.