Center-Environment Feature Model for Machine Learning Study of Spinel Oxides Based on First-Principles Computations

Abstract

Spinel oxides have attracted extensive attention due to their unique physical, chemical, optical, and electronic properties with their applications in lithium batteries, photocatalysts, and ferroelectricity. Owing to a large number of possible cation substitutions, many novel potential oxides are yet to be discovered with interesting properties. Inspired by the data-driven materials design approach, in this work, we developed machine learning (ML) models based on the first-principles computational data to investigate the energy and structure properties of normal cubic spinel oxides. The density functional theory (DFT) calculations were first carried out for 5329 spinel oxides with cubic AB2O4 structures where A and B sites were substituted with 73 elements, respectively. We predicted 451 new spinel oxides more stable than all the studied known experimental structures worth for further experimental investigation. We found the “good” A/B elements stabilizing spinel oxides include II–IV group elements and rare-earth elements in the periodic table. Furthermore, we proposed a new "Center-Environment" (CE) model to construct features containing both the composition and structure information as inputs to machine learning algorithms. Based on the DFT data, we developed ML models using a support vector regression algorithm to predict accurately and efficiently the formation energies, lattice parameters, and band gaps of spinel oxides. The composition design principles proposed in this work prompt the experimental discovery of new spinel oxides and the CE feature model can be generally applied in the data-driven materials design by ML methods.