A Machine Learning Approach for the Prediction of Formability and Thermodynamic Stability of Single and Double Perovskite Oxides

Perovskite oxides continue to attract
huge interest due to their
fascinating and wide-ranging properties for diverse applications.
The tunability of these properties may be further enhanced by increasing
their compositional complexity via double perovskite-ordered configurations
containing multiple cations. In this work, we focus on an exhaustive
chemical space of single and double oxide perovskites and optimally
explore this space to identify novel compositions that are likely
to form stable compounds. Critically, we examine the relationship
between formability, the practical ability to synthesize a compound,
and stability, the thermodynamic preference to form the structure.
Our formability and stability training data sets were enumerated from
the available experimental literature and in-house density functional
theory computations and contained 1505 and 3469 examples, respectively,
representing state-of-the-art in the current open literature in perovskite
and double perovskite compounds. Subsequently, cross-validated and
highly accurate machine learning classification models are built using
these training data sets and employed to screen for novel stable oxide
perovskites. The study identifies (1) atomic features relevant to
prediction of formability and stability in perovskite and double perovskite
compounds, (2) the importance of including energy contributions due
to local structural relaxations going beyond the high symmetry perovskite
phase, and (3) 437,828 double perovskite compounds that are likely
to be stable and 891,188 compounds that are likely to be formable.
From the intersection of this large chemical space of formable and
stable oxide perovskites, 414 compositions are identified as the most
promising candidates for future experimental synthesis of novel oxide
perovskites. The developed models may be generalized and have implications
beyond perovskite discovery if applied to other families of compounds.