AI-Driven Super-Resolution Enhancement of Scanning Electron Microscopy Images for Semiconductor Failure Analysis

Abstract

Scanning electron microscopy (SEM) is central to semiconductor failure analysis (FA), but high-resolution acquisition is slow, and fast survey scans often lack the detail needed to decide where to look closely. We report two results on enhancing low-detail SEM images of static random-access memory (SRAM) arrays. First, an honest evaluation of learned single-image super-resolution (SR): against bicubic interpolation on real paired backscattered-electron (BSE) data, SR is tied on pixel metrics (structural similarity, SSIM; peak signal-to-noise ratio, PSNR) because those metrics reward smoothing on noisy BSE - yet SR is structurally better, recovering SRAM cell modulation depth to 102-112 % of the native high-resolution (HR) value (versus 94-103 % for bicubic) and tracking the SRAM lattice spectrum closest to HR, while suppressing shot noise. Second, and more fundamental, we cast enhancement as physics-informed reconstruction: from one measured image we infer a simple, piecewise-constant scene of metal lines, dielectric voids, and bright tungsten contacts; we render that scene back through a fitted SEM point-spread function (PSF) and confirm it reproduces the measurement (relative-L2 error 0.04, correlation 0.984); then we render the same scene through an idealized point-probe to obtain a sharp visualization of the inferred geometry. We are explicit about the bound: SEM resolution is not wavelength-limited (the 10 keV electron wavelength is ≈0.12 Å, far below 1 Å), so this is prior-constrained, measurement-verified deconvolution of blur and noise - not new measurement, and not atomic resolution.