The utility of composition-based machine learning models for band gap prediction

Abstract Given the importance of band gaps in electronic and optoelectronics applications, methods that can reliably predict the band gap of any material are in demand. Here, we show that the rule-based ensemble Cubist model that uses descriptors derived from elemental compositions, provides rapid and accurate estimates of experimental band gaps. The generalizability of the model was tested using two independent test sets, both of which yielded squared correlations ⩾ 0.85. The model was also found to yield lower errors compared with most density functionals specifically crafted for the determination of band gaps. Furthermore, the inclusion of a distance based model domain applicability metric facilitates the assessment of the reliability of the model predictions, thereby improving prospective screening performance.