Data Science-Driven Bulk Electronic Structures Descriptors in Cubic Perovskite Oxides for Oxygen Evolution Reaction Activity

The discovery of new materials has been accelerated using machine learning in electrocatalysts in which the main performance criterion is the work function of oxides. It is possible to predict the work function and reveal quantitative information regarding its dependence on the structure by machine learning models. However, the experiments and calculations required to obtain the work function are costly and challenging. As a result, many descriptors based on geometry and bulk electronic structure have been offered for predicting the work function of new compositional candidates, especially for the perovskite oxide family. Among many alternatives to replace noble metals in electrocatalytic applications, perovskite oxides have been extensively studied, and promising candidates have been suggested from this class of materials for better performance. However, due to their complex electronic and crystal structure, a generally applicable descriptor to cover all possible compositional candidates is still lacking. The compositional dependence of the relationship between the most frequently used bulk electronic structure descriptor, O2p band center, and work function limits its use, verifying the need for a new descriptor. Consequently, the lack of such a descriptor is the barrier to the accelerated discovery of new simple and complex compositions by machine learning. In this study, we have investigated the potential of new bulk electronic structure descriptors for characterizing the work function of the BO₂-terminated surfaces in cubic perovskite oxides using a dataset from the scientific literature, including 583 different simple perovskite compositions with 26 different A-site and 23 different B-site elements. We have first identified the bonding-related reasons that cause the composition dependence of the relationship between the O2p band center and work function by univariant exploratory data analysis. The bonding-related features that cause the compositional dependency are electronegativity, ionicity, covalency, ionization energy, valence orbital energy, orbital radii, band-filling, bond hybridization, and charge distribution. We have generated several new descriptors by combining these features with the O2p band center in varying mathematical forms. A new descriptor that is a function of O2p band center, valence orbital energy, d-orbital radii, bond hybridization, and charge distribution has been found to have a Pearson correlation coefficient of 0.82 with the work function that is 30% higher than the previously suggested and widely used bulk electronic structure descriptor O2p band center. The strategy suggested here is exemplified for the cubic perovskite oxides with a definite surface termination; however, it is general enough for application to other crystals and surface terminations as well as other materials and application areas showing the strength of combining data science with domain knowledge.