Inverse Design of Multilayer Thin-Film Coatings with Neural Networks and Differentiable Transfer-Matrix Optics

Abstract A thin-film coating is a stack of nanometre-thick layers whose reflectance spectrum is set by interference. Computing the spectrum from the thicknesses (forward problem) is easy with the transfer-matrix method (TMM); finding thicknesses that give a desired spectrum (inverse problem) is hard because many different stacks give similar spectra. We build a 200 000-sample dataset of 8-layer TiO 2 /SiO 2 stacks on glass (400–800 nm), train a neural network that maps spectrum to thicknesses, show that the straightforward network fails because of non-uniqueness, fix this with a tandem network whose loss is computed through a differentiable TMM, and finally polish each prediction with an L-BFGS-B optimiser using an analytic gradient. On 300 unseen test spectra the mean spectral error falls from 0.114 (naive network) to 0.059 (tandem) to 0.013 (naive + refinement), i.e. about 9 lower than the naive network, at 0.05 s per target and with 92% of targets matched to within 0.05 mean error.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11712099/v1
Primary Topic
Optical Coatings and Gratings
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Inverse Design of Multilayer Thin-Film Coatings with Neural Networks and Differentiable Transfer-Matrix Optics

Gurman Singh Dhaliwal, Mayank Goswami
1 links
Research Square
Optical Coatings and Gratings
preprint

Inverse Design of Multilayer Thin-Film Coatings with Neural Networks and Differentiable Transfer-Matrix Optics

Gurman Singh Dhaliwal, Mayank Goswami
preprint en

Abstract

Abstract A thin-film coating is a stack of nanometre-thick layers whose reflectance spectrum is set by interference. Computing the spectrum from the thicknesses (forward problem) is easy with the transfer-matrix method (TMM); finding thicknesses that give a desired spectrum (inverse problem) is hard because many different stacks give similar spectra. We build a 200 000-sample dataset of 8-layer TiO 2 /SiO 2 stacks on glass (400–800 nm), train a neural network that maps spectrum to thicknesses, show that the straightforward network fails because of non-uniqueness, fix this with a tandem network whose loss is computed through a differentiable TMM, and finally polish each prediction with an L-BFGS-B optimiser using an analytic gradient. On 300 unseen test spectra the mean spectral error falls from 0.114 (naive network) to 0.059 (tandem) to 0.013 (naive + refinement), i.e. about 9 lower than the naive network, at 0.05 s per target and with 92% of targets matched to within 0.05 mean error.

Research Square
Indian Institute of Technology Roorkee (IN)
Optical Coatings and Gratings
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