Answers are grounded in this paper and cite their sources.
Herein, we propose a deep learning-based method for spectroscopic ellipsometry, termed deep learning ellipsometry (DLE), designed for practical and high-throughput optical metrology. Our DLE framework integrates four key innovations: (i) ultra-large scale training with up to 100 million datasets, (ii) a novel hybrid approach that combines linear-scale modeling for the extinction coefficient ( k ) with logarithmic-scale modeling for the absorption coefficient ( α ), (iii) a noise-robust learning strategy to handle experimental spectral noise, and (iv) an optimized deep convolutional neural network (CNN) architecture using zero padding to maintain accuracy even with complex data. Our results show that the number of training datasets and a deep CNN algorithm are critical for accuracy, particularly for deducing optical constants of crystalline phase materials, exhibiting complex optical transitions, whereas, to address the challenge of experimental noise, a transfer learning scheme is essential, enabling fine-tuning with real-world spectra and enhancing robustness. We have validated our DLE approach using 32 experimental thin films on Si substrates, achieving precise optical constants and structural parameter estimations without prior knowledge of the samples. Specifically, low analysis errors of 0.032 for k and 8.6% for the thin film layer thickness have been confirmed. For the first time, we have successfully deduced highly accurate logarithmic α spectra in a wide energy range of 0.75~6.0 eV without requiring transmission/reflectance spectra, enabling high precision bandgap determination with a statistical R-squared value of 0.997 and an average mean absolute error of only 31 meV. The total analysis time of DLE for optical constants, film structures and bandgap determination is less than 2 s, demonstrating the ultrafast analysis speed of DLE approach, compared with manual analyses (~5 min). To further improve cases with minor spectral discrepancies, we introduce a Kramers-Kronig consistent B-spline post-analysis using DLE outputs as initial guesses. This study establishes DLE as a powerful, accurate, and general-purpose AI method for ellipsometry, paving the way for ultrafast, automated optical analysis in both research and industrial applications.
No takes yet. Share an insight, caveat, or question.
Oka et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: