From the electride concept to an efficient machine learning method

An electride is a material in which electrons detach from atoms and occupy interstitial sites, acting as anions. Many electrides have been discovered or predicted, including those formed under high pressures. However, its definition has become blurred, and their formation mechanisms remain unclear. Several models have been proposed, including the non-nuclear attractor, s–p–d electron transfer, multi-center bonding, and quasi-atom theories. Our recent calculations show that quasi-atom orbitals can form diverse covalent bonding structures and explain the structural evolution of metals and compounds under pressure. Notably, our results suggest that quasi-atoms not only capture the essence of electrides but also provide a promising framework for machine-learning-driven discovery of new materials [1,2]. By integrating the chemical template concept with ML, we developed a 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 [3].