TrustSpec: Simulation-Grounded Physics-Driven Spectral Inversion With Uncertainty Quantification for SiC Epitaxial Thickness Metrology

Abstract

Accurate thickness metrology is essential for silicon carbide (SiC) epitaxial layers, yet classical Fourier-transform infrared (FTIR) fringe analysis relies on artifact-specific preprocessing, degrades below the fringe-resolution limit, and provides no calibrated confidence for individual measurements. This paper presents TrustSpec, a simulation-grounded, physics-driven spectral inversion framework intended as a proof of concept for trustworthy SiC epitaxial thickness metrology. A dispersive two-interface interference model with randomized instrument disturbances is used as a spectral digital twin to generate labeled training spectra, and the same twin is reused at inference time as an analysis-by-synthesis referee. A deep ensemble estimates thickness from a physics-conjugate Fourier-magnitude optical-thickness representation; the ensemble sample standard deviation and a windowed re-synthesis consistency ratio define a two-channel trust plane for release, review, or rejection. In a reproducible simulation study over 0.4–25 μm, TrustSpec achieves a mean absolute error (MAE) of 0.110 μm and remains effective below 1 μm, where a classical FFT pipeline fails. Split-conformal recalibration yields 94.6% empirical coverage for a nominal 95% interval. The combined trust policy releases 90.5% of in-distribution spectra at an MAE of 0.066 μm and rejects all 600 tested simulated 28–38 μm thickness-extrapolation spectra; a 25–28 μm test shows a gradual rejection transition near the training boundary. Channel ablations, referee-window sensitivity, shifted-nuisance stress tests, and model-mismatch experiments delimit the scope of these simulation results and motivate batch residual monitoring, reference-wafer checks, and periodic two-angle acquisition for drift detection. Validation on measured SiC wafers remains necessary before production deployment.