Accurate and efficient machine learning models for predicting hydrogen evolution reaction catalysts based on structural and electronic feature engineering in alloys.

Abstract

Predictive materials design of high-performance alloy electrocatalysts is a grand challenge in hydrogen production via water electrolysis. The vast combinatorial space of element substitutions in alloy electrocatalysts offers a wealth of candidate materials, but presents a significant challenge in terms of experimental and computational exploration of all possible options. Recent scientific and technological developments in machine learning (ML) have offered a new opportunity to accelerate such electrocatalyst materials design. Herein, by incorporating both the electronic and structural properties of alloys, we are able to construct accurate and efficient ML models and predict high-performance alloy catalysts for the hydrogen evolution reaction (HER). We have identified the light gradient boosting (LGB) algorithm as the best-performed method, with an excellent coefficient of determination (R2) value reaching 0.921 and the corresponding root-mean-square error (RMSE) being 0.224 eV. The average marginal contributions of alloy features towards ΔGH* values are estimated to determine the importance of various features during the prediction processes. Our results indicate that both the electronic properties of constituent elements and the structural adsorption site features play the most critical roles in the ΔGH* prediction. Furthermore, 84 potential alloys with |ΔGH*| values less than 0.1 eV are successfully screened out of 2290 candidates selected from the Material Project (MP) database. It is reasonable to expect that the ML models with structural and electronic feature engineering developed in this work would provide new insights in future electrocatalyst developments for the HER and other heterogeneous reactions.