AI-Driven Super-Resolution Enhancement of SEM Images for Semiconductor Failure Analysis
AI-Driven Super-Resolution Enhancement of SEM Images for Semiconductor Failure Analysis
Wednesday, October 7, 2026: 3:00 PM
Summary:
We present a deep learning super-resolution framework that enhances low-resolution SEM images of SRAM arrays to high-resolution quality for semiconductor failure analysis. The model trains on paired backscattered electron image datasets captured at multiple resolutions (768x547 to 6144x4376 pixels) across identical regions, covering horizontal field widths from 1.6 to 11.6 um. A dual-branch encoder-decoder architecture incorporates an autocorrelation branch that exploits SRAM periodicity as spatial priors, constraining the reconstruction solution space beyond what is achievable on arbitrary image content. Reinforcement learning-guided optimization balances reconstruction fidelity against perceptual quality, preserving physically meaningful features—contact integrity, via profiles, interconnect morphology—rather than generating visually plausible but analytically misleading artifacts. Evaluation uses both standard metrics (SSIM, PSNR) and task-specific measures including downstream anomaly detection accuracy. Results show 4x super-resolved images achieve structural similarity comparable to native high-resolution acquisitions while reducing acquisition time by an order of magnitude, enabling rapid low-resolution survey scans with computational enhancement of regions of interest.
We present a deep learning super-resolution framework that enhances low-resolution SEM images of SRAM arrays to high-resolution quality for semiconductor failure analysis. The model trains on paired backscattered electron image datasets captured at multiple resolutions (768x547 to 6144x4376 pixels) across identical regions, covering horizontal field widths from 1.6 to 11.6 um. A dual-branch encoder-decoder architecture incorporates an autocorrelation branch that exploits SRAM periodicity as spatial priors, constraining the reconstruction solution space beyond what is achievable on arbitrary image content. Reinforcement learning-guided optimization balances reconstruction fidelity against perceptual quality, preserving physically meaningful features—contact integrity, via profiles, interconnect morphology—rather than generating visually plausible but analytically misleading artifacts. Evaluation uses both standard metrics (SSIM, PSNR) and task-specific measures including downstream anomaly detection accuracy. Results show 4x super-resolved images achieve structural similarity comparable to native high-resolution acquisitions while reducing acquisition time by an order of magnitude, enabling rapid low-resolution survey scans with computational enhancement of regions of interest.