An Effective Machine Learning Frame for Materials Discovery Structured by a Chemical Concept

Despite extensive studies on binary compounds with high non-metal compositions, there remains a large, unexplored chemical space, particularly regarding non-integer non-metal-to-metal ratios. By integrating the chemical template concept with machine learning algorithms, we developed a specialized structure discovery workflow that significantly enhances the efficiency of predicting stable compounds. Our method led to the identification of 13 new structural prototypes and 31 stable metal superhydrides, representing a 23% increase in discoveries. Metal superhydrides, known for their high hydrogen content and polyhedral hydrogen cages, are promising candidates for high-temperature superconductivity. The method enables us to discover many structures containing over 50 atoms per primitive cell. Additionally, 19 of the newly identified superhydrides exhibit Tc > 100 K, highlighting the potential for higher Tc materials within the 3D hydrogen clathrate structures.