Deep learning model for 3D profiling of high-aspect-ratio features using high-voltage CD-SEM

Abstract

High-aspect-ratio (HAR) channel holes were developed for competitive cost-per-bit 3D-NAND memory. High-throughput and in-line monitoring solutions for 3D profiling of the HAR features are the key to improving yields. We previously proposed an exponential model to identify the cross-sectional profile of the HAR features using backscattered electron (BSE) images of a high-voltage critical dimension scanning electron microscopy (CD-SEM). However, the 3D profiling accuracy was insufficient when the depth of the HAR features was far greater than the focus depth of the electron beam. To address this issue, we developed a deep learning (DL) model, which takes account of the aperture angle and the aberration of the electron beam, to predict the 3D profile from BSE images. The predicted cross-sections of the HAR holes with different bowing geometries were compared with field-emission SEM (FE-SEM) measurements. The results show that the DL model provides higher sensitivity than the exponential model does.