Solar energy offers a promising means of addressing energy supply and storage problems, but this potential is not fully realized due to a lack of suitable semiconducting materials. The discovery of new materials with desirable properties has historically been conducted either using an experimental or a first-principles density functional theory based study. These approaches are extremely time-intensive, and therefore, cannot be applied effectively to study a large number of systems. In such situations, machine learning can be used to make predictions about properties of new compounds from known data, providing a more efficient route to materials discovery. Here, this approach is used to predict the bandgap of a series of oxysulfide perovskites (of the form of ABOXS3−X, X = 0,1,2,3), in general, and sulfur-rich ABOS2, in particular. Atomic properties of constituent elements in the perovskite structures via 1.048 millions possible subsets of features are employed to train the models. Further, feature selection, kernel ridge regression, and k-nearest neighbors classification methods are applied to downselect the promising ABOS2 based oxysulfide perovskites for water-splitting. The accuracy of each model is determined using standard statistical metrics. Finally, seven stable but yet unsynthesized sulfur-rich oxysulfide perovskites (BiInOS2, BiGaOS2, SbInOS2, SbGaOS2, SbAlOS2, SnZrOS2, and MgSnOS2) that show potential for water-splitting applications are proposed.
Dr. Priya Johari