Machine learning guided study of composition-coefficient of thermal expansion relationship in oxide glasses using a sparse dataset

Abstract

The coefficient of thermal expansion (CTE) is one of the important macroscopic properties of oxide glasses for diversified applications. The optimization of chemical composition is required to produce silicate glasses with desirable CTE for specific applications. The traditional composition optimization approach relies on a trial-and-error laboratory experimentation process which is highly time-consuming. In order to expedite the optimization process, a mathematical model need to be developed that can predict CTE with given oxide glass composition. For this purpose, an ever-increasing interest of worldwide science community have recently been witnessed in data-driven machine learning models. These models have shown interesting as well as encouraging results for many such applications. The present study focuses on building the random forest machine learning regression model for predicting CTE of silicate glasses. The mole fractions of glass constituents (Al2O3, B2O3, Na2O, MgO, CaO, BaO, and SiO2) are considered model input features for predicting CTE. The figures of merit used to evaluate of the model performance are regression coefficient (R2) and mean absolute percentage error (MAPE). The R2 and MAPE of the model are 0.95 and 8.44 respectively in predicting the CTE of unseen glasses, which demonstrate a reasonably good model performance. The model revealed that the mole fraction of Na2O has distinctively high feature importance compared to the rest of the constituents in predicting CTE. The feature importance of mole fractions of glass constituents other than Na2O lies in a narrow range.