Deep Learning-Based 2D ROI Alignment for Automated Semiconductor SEM Imaging
Deep Learning-Based 2D ROI Alignment for Automated Semiconductor SEM Imaging
Tuesday, October 6, 2026: 12:50 PM
Summary:
Automated SEM is essential for semiconductor failure analysis, yet rule-based ROI alignment often degrades under diverse imaging perturbations. We integrate a deep-learning ROI recognition framework into a NAND top-view workflow on Thermo Fisher iFast. Word-line (WL) nitride-layer bands (low-magnification y correction) and channel-hole (CH) centering via word-line-cut patterns (high-magnification x correction) are compared against a rule-based profile-peak baseline on a fixed 30-image holdout using detection and segmentation models (~200 training images per task). Both models outperform the baseline (WL success 62% to ≥99%; CH false positives 46 to 0) with end-to-end latency of 11–12 ms (detection) and 17 ms (segmentation). WL operational targets are achieved with ~70–90 labels. Results demonstrate robust, real-time ROI alignment for automated 3D NAND SEM acquisition.
Automated SEM is essential for semiconductor failure analysis, yet rule-based ROI alignment often degrades under diverse imaging perturbations. We integrate a deep-learning ROI recognition framework into a NAND top-view workflow on Thermo Fisher iFast. Word-line (WL) nitride-layer bands (low-magnification y correction) and channel-hole (CH) centering via word-line-cut patterns (high-magnification x correction) are compared against a rule-based profile-peak baseline on a fixed 30-image holdout using detection and segmentation models (~200 training images per task). Both models outperform the baseline (WL success 62% to ≥99%; CH false positives 46 to 0) with end-to-end latency of 11–12 ms (detection) and 17 ms (segmentation). WL operational targets are achieved with ~70–90 labels. Results demonstrate robust, real-time ROI alignment for automated 3D NAND SEM acquisition.