Prediction of glass refractive index based on a multi-algorithm stacked integration model

Abstract

Machine learning (ML) has been recognized as a viable method to accelerate the design of materials with targeted properties. A number of composition–property relationships for modeling oxide glasses with machine learning have been recently reported, including density, modulus of elasticity, glass transition temperature, and refractive index. As we know, most current models for predicting the refractive index of glass compositions usually use a single machine learning algorithm. But a single machine learning model may not fully capture all complex relationships in the data, resulting in model bias. Herein, we use a multi-algorithm stacked integration approach—namely stacking different algorithms—to rectify model bias caused by a single algorithm. We provide insight into a new strategy of applying stacked integration models in oxide glass prediction. SPearman’s coefficients are used to discover relationships between features and target variables, combined with SHAP additive interpretation and analysis. In this paper, a stacked integration model is proposed to predict the refractive index of oxides, and RMSE is reduced by 3.6%–38% compared to a single model. The potential of stacked integration models to predict refractive index is proven through our work.