Amorphous Catalysis: Machine Learning Driven High-Throughput Screening of Superior Active Site for Hydrogen Evolution Reaction

The prediction of chemisorption energy to facilitate the high-throughput screening of active catalysts has been long pursued but remains challenging. In particular, amorphous materials usually exhibit superior activity and have drawn ever-increasing attention in heterogeneous catalysis. However, the insight into the basic structure–property relation remains far from sufficient owing to their disordered structure and untracked surface state, let alone the effective prediction of adsorption energy. Here, employing the amorphous Ni2P catalyst as an example and powerful machine learning (ML) models, we propose an effective strategy that enables fast and quantitative prediction of the adsorption energy of hydrogen on amorphous Ni2P surfaces. Specifically, our method decomposes the difficult prediction of adsorption energy on amorphous surfaces into two subproblems: frozen adsorption energy and relaxation energy. By training with a set of ab initio adsorption energies within a wide configuration space, we succeed to predict the adsorption energies with ∼0.1 eV error by adopting the feature only relying on local chemical environment. Our strategy allows us to successfully implement the high-throughput exploration of active sites for hydrogen evolution reaction (HER). This work builds a predictive model of site-specific chemisorption energy, and the related statistical analysis underpins the fundamental understanding of the chemical bond, which could largely facilitate rational design of active amorphous catalysts.