Machine learning analysis of broadband optical reflectivity of semiconductor thin film

Broadband reflection spectroscopy is a non-invasive and non-contact tool widely used to measure optical dielectric constants and thickness of thin films. However, a lot of time and effort are consumed to analyze data before the results can be attained. Here we construct an artificial neural network (ANN) using scattering matrix formalism and U-net architecture, and apply it to analyze infrared reflection of SiO₂ thin film grown on Si substrate. The ANN returns multiple outputs—frequency-dependent optical refractive index (n), absorption coefficient(κ), and thickness of the film (d)—with high precision with 0.6 nm thickness difference. Furthermore, the ANN can fit large number of reflection data taken at numerous positions (500) of the thin film in short time less than 150 ms, and creates fine-scale thickness map with 0.6 nm thickness resolution. This work demonstrates that U-net-based ANN is a powerful method of reflectivity analysis and can be applied to other thin-film materials.