Accelerating the discovery of direct bandgap doped-spinel photovoltaic materials: A target-driven approach using interpretable machine learning

Abstract

The spinel structure, unifying tetrahedral and octahedral coordination within a single crystal lattice, emerges as a promising alternative to perovskite and traditional semiconductors, particularly in the context of photovoltaic applications. However, the majority of surveyed spinels exhibit an indirect bandgap, imposing substantial limitations on their utility. Consequently, doping is required to tailor indirect bandgap spinels into direct ones. In this study, leveraging machine learning (ML) algorithms, we design a target-driven framework to accelerate the ab initio predictions of unknown doped-spinels using elements from the periodic table. Utilizing this approach, we have introduced a novel machine learning classification model designed for the prediction of the direct-indirect bandgap nature in spinel materials. This pioneering model shows promising potential for achieving rapid and precise predictions, particularly in scenarios involving small dataset. Especially, a full list of potential 3449 (AxA′1-x)B2X4-type doped-spinels and 3809 A(BxB′2-x)X4-type doped-spinels with direct bandgaps is identified from the vast pool of unknown doped-spinel materials. Further, the application of interpretable ML extracts the first ionization energy of the B-site ion as the most important feature for the nature of bandgap, offering insightful design rules. This research provides a novel perspective on unraveling spinel materials, and the proposed design framework proves effective in identifying high-performance materials within a vast chemical space, all while minimizing computational costs.