Machine learning for deconstructing contributions of atomic characterizations to achieve hybridization-determined electron transfer in a perovskite catalyst

Abstract

Perovskite-type oxide has various environmental and energy-related applications and attracts continuous attention in the field of catalysis. In previous reports, the catalytic activity of perovskites is usually attributed to their higher specific surface area, more lattice oxygen, and abundant valences. In this work, a machine-learning method was developed to predict the catalytic performance and deconstruct contributions of atomic features. As a result, the model evaluation index (R²) was >0.8 and root mean square error (RMSE) was <0.13. Screened-out experimental data were consistent with predicted tendencies, and LaMnO₃ was the best formula for the catalytic oxidation of toluene. Besides, the contribution of the Batsanov equilibrium van der Waals radius of the B-site element (63%) and atomic number of the A-site element were 63% and 19%, respectively. The Pyykkö covalent radius of the A-site element (PCR_A) and d electron density of the B-site element (D_d) occupied 18% of the total contribution. Furthermore, since molecular scale characterization failed to explain catalytic tendency and atomic features, a combined macro impedance and micro hybridization method was developed. In brief, a wider range of total hybridization resulted in smaller impedance, faster electron transfer speed, and enhanced catalytic activity. The main results of this work would facilitate the development of an effective catalyst with a more precise understanding of its catalytic mechanism.