Smart ellipsometry with physics-informed deep learning

Abstract

Abstract Ellipsometry is a century-old optical measurement technique extensively used to determine thicknesses and optical constants of thin films. Nevertheless, as ellipsometry is essentially an indirect characterization method, ellipsometry data analysis, which traditionally involves tedious and time-consuming human-expert intervention and requires adequate knowledge and experience to model the optical properties of film materials, has long been universally considered to be challenging. We proposed a smart ellipsometry network (SENet), with which optical constants satisfying the Kramers–Kronig relations can be achieved automatically without any prior knowledge of material optical properties. The proposed method was verified experimentally using a variety of thin films, with the materials covering metals, dielectrics and semiconductors in an ultraviolet–visible–near-infrared broad spectral range. Furthermore, our method demonstrated remarkable generalization capabilities in the analysis of thin films at different angles of incidence, in different spectral ranges, and prepared on different substrates from those in the generation of the training dataset. The developed method paves the way for ellipsometry as a general characterization technique that allows for the automatic, accurate, and fast determination of material optical properties.