INVESTIGATING BAND GAP DIRECTNESS USING MACHINE LEARNING

Bandgap directness is the basis of most optoelectronic device applications, e.g., direct band gaps materials are expected to have high efficient light emission. However, no unified theory exists to explain why a given material has direct or indirect band gaps. Using all semiconductors’ band structures from Materials Project, a total of 18,372 materials, we have used classification Machine Learning methods for general prediction of band gap directness and, more importantly, for extraction of interpretable knowledge of direct-indirect transitions. Here we used a visualization tool for the rules in the decision trees method of machine learning, referred to as Explainable Matrix (ExMatrix), with the main purpose of deepening the current understanding of such transitions. We applied this methodology for prototypical groups of semiconductors such as zincblende, rocksalt, wurtzite and perovskites, leading to insights regarding the influence of chemical composition, structural parameters and individual atomic properties on band structures. Such results might lead to new strategies for engineering band gap directness, thus widening the applicability of indirect band gap semiconductors.