Adaptive mixture-of-experts denoising for deep learning of optical ellipsometry spectra

Abstract

Spectral noise limits the reliability of deep-learning ellipsometry (DLE), particularly under short measurement times required for high-throughput materials characterization. Here, we propose an adaptive denoising framework that integrates a Mixture-of-Experts (MoE) module into a U-Net architecture for spectroscopic ellipsometry. The MoE dynamically assigns expert weights according to input-dependent noise characteristics, enabling data-driven specialization across diverse noise regimes. To further enhance robustness, we introduce an iterative denoising strategy coupled with a pretrained DLE-based selection mechanism that enforces reliable evaluation of candidate outputs. The model is trained exclusively on simulated spectra with artificial noise, yet it is shown to be effective on experimental data from 27 semiconductor thin-film materials measured using two distinct ellipsometry systems. The proposed MoE-based method reduces the global DLE prediction error by 34.5%–65.0%, depending on the measurement time, achieving the maximum 65.0% reduction under the shortest measurement time (1 s) where noise is most severe. By adopting the MoE denoising architecture, bandgap estimation accuracy improves significantly, with the mean absolute error reduced by 44.3% using the iterative MoE approach. In addition to improving prediction accuracy, the proposed method is 43 times faster than the corresponding transfer learning approach and eliminates the need for substrate-dependent retraining. These results demonstrate that MoE-driven adaptive specialization provides a computationally efficient and highly reliable strategy for robust spectral denoising in deep-learning-based ellipsometric analysis.