Physics‐Enhanced Learning for Automated Determination of Material Optical Constants

Abstract Accurate determination of optical constants is a prerequisite for various fundamental research and optoelectronic applications of material. Spectroscopic ellipsometry (SE) is almost the standard tool for determining optical constants and thin‐film thicknesses, but inverse ellipsometry data analysis traditionally requires prior knowledge about sample materials and involves tedious and time‐consuming human‐expert intervention, which is widely recognized as a major challenge. Existing deep‐learning‐based methods allow fast analysis but suffer from poor generalization and lack of physically plausible results. Here, a Smart Ellipsometry Network (SENet) is reported that adopts physics‐enhanced learning‐based inference for the automated determination of optical constants without prior knowledge or expert intervention. Trained with artificially generated materials that have no connection or resemblance to real‐world samples, SENet achieves excellent generalization to realistic thin‐film samples in various unseen scenarios, such as nonideal structures, varied measurement conditions, and different substrates. Notably, SENet ensures that predicted optical constants always fulfill the Kramers–Kronig (KK) relations. Moreover, SENet enables highly efficient implementation and yields reliable results in ≈30 s using an ordinary laptop without graphics processing unit (GPU) acceleration. Overall, the developed method opens a new avenue for automated, accurate, and fast determination of material optical constants in a single‐click manner.