Predicting the Band Gaps of Inorganic Solids by Machine Learning

Abstract

A machine-learning model is developed that can accurately predict the band gap of inorganic solids based only on composition. This method uses support vector classification to first separate metals from nonmetals, followed by quantitatively predicting the band gap of the nonmetals using support vector regression. The superb accuracy of the regression model is obtained by using a training set composed entirely of experimentally measured band gaps and utilizing only compositional descriptors. In fact, because of the unique training set of experimental data, the machine learning predicted band gaps are significantly closer to the experimentally reported values than DFT (PBE-level) calculated band gaps. Not only does this resulting tool provide the ability to accurately predict the band gap for any composition but also the versatility and speed of the prediction based only on composition will make this a great resource to screen inorganic phase space and direct the development of functional inorganic materials.