Combining Density Functional Theory and Machine Learning for Compositional Optimization of Alloyed Perovskite Electrocatalysts

Multicomponent metal oxides, such as perovskite oxides, hold promise for use as sustainable alternatives to Ir-, Ru-, and Pt-based electrocatalysts at scale. Perovskites can accommodate a wide variety of elements in their A- and B-sites, enabling tuning of their structural and electronic properties through compositional alloying. These properties, which are obtainable from density functional theory (DFT) calculations, can be used as low-dimensional descriptors that correlate with experimental stability and activity in, for example, the oxygen evolution reaction (OER). Established descriptors of stability include energy above convex hull and energy above Pourbaix hull, while those for catalytic activity include the difference between oxygen 2p-band center and the Fermi level, the difference between the metal 3d- and oxygen 2p-band centers, as well as oxygen vacancy formation energy. We are therefore presented with a combinatorial problem of determining which A- and B-site compositions optimize such descriptors. The compositional search space for A_xA’_1-xB_yB’_1-yO₃perovskites with up to two different elements in their A- and B-sites is O(10⁶), making it intractable to calculate descriptors exhaustively using DFT. Here, we thus combine high-throughput DFT calculations with crystal graph convolutional neural networks (GCNs) to screen for stability and activity descriptors.<br/><br/>Previous high-throughput screening efforts have largely focused on ABO₃ perovskites, without alloying of the A- and B-sites, or narrow compositional ranges of perovskites with A- and/or B-site alloying. To predict descriptors of highly-alloyed perovskites over a wide cation space, we have used our high-throughput virtual screening platform to generate a dataset of over 10,000 perovskites. Notably, the presented dataset includes over 4,000 (AA’)(BB’)O₃ perovskites with varied A- and B-site alloying ratios and with over 2,000 unique cationic combinations. This significantly increases the amount of available highly-alloyed perovskite DFT data, with previous alloying studies including only a few hundred unique combinations of cations. Moreover, we have generated a dataset of over 2,000 (AA’)(BB’)O₃ perovskites with oxygen vacancies; to our knowledge, the only available high-throughput dataset studying vacancy formation in alloyed perovskites. Using our datasets in combination with data available in the literature, we have trained GCNs to predict the aforementioned descriptors from unrelaxed cubic perovskite structures and used these models to predict descriptors for O(10⁶) perovskites containing over 20 different A- and B-site cations, respectively. Leveraging these predictions, we describe trends in the structural and electronic properties of alloyed perovskites across a vast compositional space and work to establish improved structure-property relationships between computed descriptors and experimental catalytic performance. The presented work provides the community with a benchmark dataset for further study of alloyed perovskites and analyses of compositional trends, laying the groundwork for improved design of perovskite electrocatalysts.