Breaking the Fixed-Substrate Limitation of Deep Learning Ellipsometry with Pseudo-Substrate Spectra

A machine learning method, deep learning ellipsometry (DLE), enables direct and rapid estimation of the optical constants and structures of thin films from spectroscopic ellipsometry (SE) spectra. However, conventional neural-network-based DLE models are restricted to the specific substrate used during training, and applying them to a different substrate requires ~100 h of retraining. Here, we propose a framework that overcomes this fixed-substrate limitation by introducing a two-input feature-wise linear modulation DLE (FiLM-DLE). In this model, the SE spectra of a thin-film structure are input together with those of the substrate measured before film deposition, which we term the pseudo-substrate spectra. Any substrate, including a multilayer stack, is thus represented solely by its measured SE response, analogous to the pseudo-dielectric function, so that no prior knowledge of the substrate structure is required. Trained on a virtual dataset of 165 million samples representing metal, glass, and SiO2/Si substrates with different SiO2 thicknesses, a single FiLM-DLE model, trained once, extracts thin-film optical constants and structures with high accuracy and shows strong applicability to substrates not included in the training datasets. Experimental validation on 47 thin-film samples across diverse substrates yielded sample-averaged mean absolute errors, relative to conventional manual analysis, of 0.047 for refractive index, 0.030 for extinction coefficient, and 1.52 nm for thickness, with each analysis completed in under 2 s. FiLM-DLE thus eliminates the need for a separate DLE model for each substrate and enables rapid, high-accuracy optical characterization of thin films on diverse substrates.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-16
DOI
https://doi.org/10.48550/arxiv.2611.20694
Primary Topic
Optical Polarization and Ellipsometry
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Breaking the Fixed-Substrate Limitation of Deep Learning Ellipsometry with Pseudo-Substrate Spectra

James N. Hilfiker, Hitoshi Tampo, Yuki Ikarashi, Takuji Maekawa et al.
arXiv (Cornell University)
Optical Polarization and Ellipsometry
preprint

Breaking the Fixed-Substrate Limitation of Deep Learning Ellipsometry with Pseudo-Substrate Spectra

James N. Hilfiker, Hitoshi Tampo, Yuki Ikarashi, Takuji Maekawa, Naoki Kurokawa, Hiroyuki Fujiwara
preprint en

Abstract

A machine learning method, deep learning ellipsometry (DLE), enables direct and rapid estimation of the optical constants and structures of thin films from spectroscopic ellipsometry (SE) spectra. However, conventional neural-network-based DLE models are restricted to the specific substrate used during training, and applying them to a different substrate requires ~100 h of retraining. Here, we propose a framework that overcomes this fixed-substrate limitation by introducing a two-input feature-wise linear modulation DLE (FiLM-DLE). In this model, the SE spectra of a thin-film structure are input together with those of the substrate measured before film deposition, which we term the pseudo-substrate spectra. Any substrate, including a multilayer stack, is thus represented solely by its measured SE response, analogous to the pseudo-dielectric function, so that no prior knowledge of the substrate structure is required. Trained on a virtual dataset of 165 million samples representing metal, glass, and SiO2/Si substrates with different SiO2 thicknesses, a single FiLM-DLE model, trained once, extracts thin-film optical constants and structures with high accuracy and shows strong applicability to substrates not included in the training datasets. Experimental validation on 47 thin-film samples across diverse substrates yielded sample-averaged mean absolute errors, relative to conventional manual analysis, of 0.047 for refractive index, 0.030 for extinction coefficient, and 1.52 nm for thickness, with each analysis completed in under 2 s. FiLM-DLE thus eliminates the need for a separate DLE model for each substrate and enables rapid, high-accuracy optical characterization of thin films on diverse substrates.

arXiv (Cornell University)
Optical Polarization and Ellipsometry
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon