Abstract
The computational design of suitable multiferroic double perovskite oxides requires finding materials that exhibit sizable polarization, magnetization, and coupling between them. Oxides with the chemical formula of AA′BB′O6 with building blocks of ABO3 single perovskite oxides in centrosymmetric Pnma symmetry are strong candidates that have been reported to satisfy such criteria. The system lowers to noncentrosymmetric, polar P21 symmetry if A/A′ layered and B/B′ rocksalt cation orderings are imposed. A detailed compositional search over a variety of chemical spaces followed by evaluating their polarization may lead to the identification of more of these compounds with ferroelectric ordering. The standard density functional theory practices to estimate polarization within the Berry phase formalism require the systems to be perfectly insulating. The number of compounds that can be evaluated using this method is therefore limited. In this work, we introduce a predictive learning strategy based on importance sampling to build a series of machine learning models using results from first-principles simulations to predict polarization and the corresponding switching barrier. The geometry-driven features related to charge states and cationic radii play key roles in predicting the switching barrier with complementary contributions from the key structural mode-based order parameters. These modes become important to draw reasonable predictions of polarization components from machine learning models. Our predictive models identify candidates with high polarizations and low switching barriers from a pool of double perovskite oxides, suitable for future investigation for their potential applications in spintronic devices.
