Stability and Equilibrium Structures of Unknown Ternary Metal Oxides explored by Machine-Learned Potentials

Ternary metal oxides are extensively utilized across various applications and have been comprehensively cataloged in experimental materials databases. Nonetheless, there remains a substantial portion of unexplored combinations of cations with oxide forms in terms of their stability and structures. Discovering new ternary metal oxides solely through experiments is both time-consuming and resource-intensive. Alternatively, theoretical databases offer hypothetical structures based on known prototypes. However, it is possible to miss “hidden” ground state structures that are not represented by prototype structures. This challenge can be addressed through the application of crystal structure prediction (CSP), which employs heuristic methods, such as evolutionary algorithm, to identify the lowest-energy structure for a given composition without prior knowledge. Nevertheless, the effectiveness of this approach is limited due to its reliance on computationally expensive density functional theory (DFT) calculations and an exhaustive search of the structure space, which is often impractical for finding complex equilibrium phases of ternary metal oxides.<br/>Recent years have seen a growing interest in machine-learned potentials (MLPs) an effective alternative to the DFT method. MLPs can rapidly compute energy and atomic forces, achieving results comparable to DFT. Based on the concept that MLPs trained with disordered phases have been successful as surrogate models for DFT in CSP [1], a systematic CSP program called SPINNER (Structure Prediction of Inorganic crystals using Neural Network potentials with Evolutionary and Random searches) was developed, harnessing MLPs within an evolutionary algorithm [2]. In benchmark tests using experimental structure data, SPINNER identified 80% of the structures of 60 different ternary compounds with diverse crystal symmetries and identified 10² to 10³ times faster than DFT-based heuristic methods.<br/>In this study, we extensively employ SPINNER to explore experimentally uncharted chemical spaces [3]. We investigate 181 ternary metal oxide systems, encompassing most cations except those with partially filled 3d or f shells, and search up to 60,000 crystal structures in representative compositions derived from common oxidation states and a machine-learned recommender system to determine the lowest energy crystal structure. Our exploration yields 45 systems containing stable ternary oxides that do not decompose into binary or elemental phases. Interestingly, many of these are noble metal-containing systems that have not yet been well studied and have equilibrium structures that do not belong to the known prototype structure. Comparisons with other theoretical databases highlight the strengths and limitations of informatics-based material searches and point to the synergistic effect between direct and data-mining searches. With a relatively modest computational resource requirement, we contend that heuristic-based structure searches, as demonstrated here, offer a promising approach for future materials discovery endeavors.<br/><br/>[1] C. Hong, <i>et al</i>, <i>Phys. Rev. B</i> <b>102</b>, 224104 (2020).<br/>[2] S. Kang, <i>et al</i>, <i>npj Comput. Mater</i>. <b>8</b>, 108 (2022).<br/>[3] S, Hwang, <i>et al</i>, <i>J. Am. Chem. Soc</i>. <b>145</b>, 19378 (2023).