Materials have various degrees of freedom in their composition and structure, resulting in a large chemical space characterized by these combinations. To accelerate the discovery of promising, unreported materials, methods for identifying favorable compositions and their stable structures efficiently are required. In this study, we propose a crystal structure prediction method that combines low computational cost with high accuracy. This approach integrates symmetry-constrained random structure generation, a universal machine-learning interatomic potential, and first-principles calculations. Reproducibility tests conducted on the crystal structures of eight reported compositions demonstrate the efficacy of our method. To identify promising compositions from the vast chemical space strategically, we developed a machine learning model that estimates the existence probability of unreported compositions using information from reported compositions. Validation involved treating the reported compositions as unreported, achieving an 88% reproducibility rate. As an application of our approach, we explored novel ternary oxides having polar structures by incorporating an additional machine learning model to predict their polarity based on compositional information. Crystal structure prediction was carried out for 18 unreported compositions, each predicted by machine learning to have an existence probability exceeding 0.9 and polar structures. Our exploration yielded three candidates for novel polar oxides that are thermodynamically and dynamically stable. Further, each candidate has an unreported structure, underscoring the potential of our approach to expand chemical space.

Close

PrevNext

View figure in article

• Composition
• Crystal structure
• Oxides
• First-principles calculations
• Machine learning