Data-driven physics-informed descriptors of cation ordering in multicomponent perovskite oxides

The structural tunability and compositional diversity of multicomponent perovskite oxides have enabled their various applications. The cation ordering in these oxides, ranging from disordered to ordered, profoundly controls their properties but concurrently complicates the design and discovery of these materials. While first-principles simulations and machine learning can typically predict properties associated with a particular ordering, inferring which ordering—if any—will dominate experimentally remains challenging. In this work, we establish data-driven, physics-informed descriptors of experimental ordering in multicomponent perovskite oxides to offer systematic benchmarks between machine learning, theory, and experiments. While state-of-the-art machine learning interatomic potentials only partially capture experimental ordering, descriptors obtained from first-principles simulations correctly rank up to 93% of compositions in an experimental dataset of 190 perovskite oxides between cation ordered and disordered. These descriptors can accelerate the high-throughput virtual screening of multicomponent oxides by predicting dominant ordering before experiments to avoid costly, exhaustive simulations of cation arrangements.