A Deep Learning Approach for Thickness Determination of AZO Thin Films
- 1Universidad Nacional Autónoma de México , , ,
Abstract
We developed a database and a machine learning model to predict the thickness of aluminum-doped zinc oxide (AZO) thin films from their optical transmittance spectra. We measured the spectra of 145 thin films with thicknesses between 100 and 1500 nm and fitted each with a 12-parameter function, one parameter being the film thickness. We divided the thickness range into five regions and modeled the distribution of each fitting parameter in each region as a sum of Gaussian functions. Random sampling from these distributions produced parameter sets used to construct synthetic transmittance spectra (X), each paired with its film thickness (Y) to form one database sample (X, Y). We generated 240,000 samples per region, yielding a database of 1.2 million samples for supervised training with thickness as the target. The initial machine learning model, a fully connected neural network, showed poor predictive accuracy. We therefore improved the architecture by incorporating an autoencoder with skip connections to extract the most informative features from the transmittance spectra. This change substantially improved performance, yielding mean absolute errors of 15.8(3) nm on the training set and 15.7(3) nm on the validation set. When applied to the transmittance spectra of the experimental AZO films, the trained model predicted thicknesses with a mean absolute percentage error of 6.91%. Because the model was trained using data derived from AZO measurements, it applies only to AZO thin films. The methodology, however, is general and can be applied to thin films of any composition.
Related
- ALD grown (001) κ-Ga2O3 epitaxial thin films on c-plane sapphire: influence of thickness and thermal annealing
- Engineering shallow and deep level defects in κ-Ga2O3 thin films: comparing metal-organic vapour phase epitaxy to molecular beam epitaxy and the effect of annealing treatments
- Interpretable inverse design of optical multilayer thin films based on global neural adjoint
- Magnetic Resonance Properties of Thin Py Films Suitable for Excitation of Spin Waves
- Scalable Room-Temperature Synthesis of Glass-like MOF Thin Films on Arbitrary Substrates with Versatile Functionality
BibTeX
@article{guerrerososa2026deep,
title = {{A Deep Learning Approach for Thickness Determination of AZO Thin Films}},
author = {Víctor Manuel Guerrero-Sosa and Angélica Carrillo-Verduzco and Juan Carlos Alonso-Huitrón and Alan Gabriel Amaro-Colin and Miguel Esquivel-García and Xim Bokhimi},
journal = {ACS Omega},
year = {2026},
month = oct,
doi = {10.1021/acsomega.6c10558},
url = {https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c10558/5458896/A-Deep-Learning-Approach-for-Thickness},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A Deep Learning Approach for Thickness Determination of AZO Thin Films
AU - Víctor Manuel Guerrero-Sosa
AU - Angélica Carrillo-Verduzco
AU - Juan Carlos Alonso-Huitrón
AU - Alan Gabriel Amaro-Colin
AU - Miguel Esquivel-García
AU - Xim Bokhimi
PY - 2026
DA - 2026/10/27
JO - ACS Omega
DO - 10.1021/acsomega.6c10558
UR - https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c10558/5458896/A-Deep-Learning-Approach-for-Thickness
AB - We developed a database and a machine learning model to predict the thickness of aluminum-doped zinc oxide (AZO) thin films from their optical transmittance spectra. We measured the spectra of 145 thin films with thicknesses between 100 and 1500 nm and fitted each with a 12-parameter function, one parameter being the film thickness. We divided the thickness range into five regions and modeled the distribution of each fitting parameter in each region as a sum of Gaussian functions. Random sampling from these distributions produced parameter sets used to construct synthetic transmittance spectra (X), each paired with its film thickness (Y) to form one database sample (X, Y). We generated 240,000 samples per region, yielding a database of 1.2 million samples for supervised training with thickness as the target. The initial machine learning model, a fully connected neural network, showed poor predictive accuracy. We therefore improved the architecture by incorporating an autoencoder with skip connections to extract the most informative features from the transmittance spectra. This change substantially improved performance, yielding mean absolute errors of 15.8(3) nm on the training set and 15.7(3) nm on the validation set. When applied to the transmittance spectra of the experimental AZO films, the trained model predicted thicknesses with a mean absolute percentage error of 6.91%. Because the model was trained using data derived from AZO measurements, it applies only to AZO thin films. The methodology, however, is general and can be applied to thin films of any composition.
SN - 2470-1343
ER -