Skip to content

Deep Learning-Based ROI Alignment for Automated SEM Workflows

Oct 2026 · International Symposium for Testing and Failure Analysis · 0 citations

Abstract

Over the past several years, scanning electron microscopy (SEM) based automated acquisition workflows have been widely used in semiconductor analysis to rapidly inspect fine patterns and defects in advanced memory and non-memory devices. In particular, during repetitive high-magnification imaging across multilayer structures, pattern alignment that stably maintains the region of interest (ROI) is a critical factor governing metrology reliability and throughput. However, most automated SEM workflows rely on rule-based methods such as intensity-profile peak finders or template matching, whose fixed configuration cannot accommodate the full diversity of imaging conditions. This imposes two coupled costs that grow as conditions diversify: the lower success rate forces repeated reacquisition that consumes instrument time, while restoring accuracy requires repeated manual re-tuning that consumes operator time. In this study, a deep learning ROI recognition framework based on a one-stage object detection model and a semantic segmentation model was evaluated within the Thermo Fisher iFast automation workflow for a NAND SEM workflow, and was compared with a conventional rule-based edge-finder baseline (intensity-profile peak detection) under the same Test-set and operational-specification conditions. The evaluation tasks consisted of (i) word-line (WL) nitride-layer-band alignment and (ii) channel-hole (CH) centering using adjacent word-line-cut patterns as recognition references. Approximately 200 SEM images were collected per task, and evaluation was performed on a fixed Test set of 30 images while varying training-set size (N_train) from 10 to 136 (step 10). On the 30-image Test set, both deep learning models substantially improved operational metrics over the edge-finder baseline (WL: success rate 74.8% → 98.1-99.4%, F1 72.5% → 98.7-99.7%; CH: success rate/F1 100%, FP/FN = 0). Together with single-image end-to-end inference latency below 20 ms for both models, these results confirm that both models provide alignment reliability and real-time feasibility sufficient for deployment in automation workflows.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.