Utilizing deep learning for swift analysis of high-throughput spectroscopic ellipsometry data on anodized oxides of valve metals

Abstract

Spectroscopic ellipsometry is a powerful high-throughput method for mapping the optical properties of combinatorial anodic oxides on alloys. However, the traditional ellipsometry data fitting using non-linear regression highly depends on correct assumptions and is tedious. The determination of the transition concentration of parent alloys that influences its properties to change, based on existing experimental data, without requiring further experimental measurements is also crucial for alloy engineering. Herein, anodic oxides grown on Nb-Ta and Nb-Ti combinatorial thin film binary libraries using a co-sputtering process are prepared (Nb concentration range is 10 ∼ 90 at.%, the oxidation voltage range is 1 ∼ 10 V). A deep learning method is developed to predict the refractive index (n) and extinction coefficient (k) of the oxide film from the ellipsometry data (606 groups). Four algorithms Convolutional Neural Networks (CNN), Convolutional Sequence-to-Sequence (ConvSeq2Seq), Temporal Convolutional Network (TCN), Gated Recurrent Unit Sequence-to-Sequence (GRUSeq2Seq) are compared. CNN achieves superior performance, yielding root mean square error values (RMSE) as low as 0.094 for predicting n and 0.037 for k. Additionally, a predictive model to reveal band-gap trends as a function of parent alloy composition and oxidation voltage at unmeasured locations along the libraries is developed. This approach helps identify the transition concentration that lead to optical property changes, thereby avoiding high experimental costs and potential experimental errors.