Directly predicting limiting potentials from easily obtainable physical properties of graphene-supported single-atom electrocatalysts by machine learning

Abstract

The oxygen reduction reaction (ORR), oxygen evolution reaction (OER), and hydrogen evolution reaction (HER) are three critical reactions for energy-related applications, such as water electrolyzers and metal–air batteries. Graphene-supported single-atom catalysts (SACs) have been widely explored; however, either experiments or density functional theory (DFT) computations cannot screen catalysts at high speed. Herein, based on DFT computations of 104 graphene-supported SACs (M@C₃, M@C₄, M@pyridine-N₄, and M@pyrrole-N₄), we built machine learning (ML) models to describe the underlying pattern of easily obtainable physical properties and limiting potentials (mean square errors = 0.027/0.021/0.035 V for the ORR/OER/HER, respectively) and employed these models to predict the catalytic performance of 260 other graphene-supported SACs containing metal-N_xC_y active sites (M@N_xC_y). We recomputed the top catalysts recommended by ML towards the ORR/OER/HER by DFT, which confirmed the reliability of our ML model, and identified two OER catalysts (Ir@pyridine-N₃C₁ and Ir@pyridine-N₂C₂) outperforming noble metal oxides, RuO₂ and IrO₂. The ML models quantitatively unveiled the significance of various descriptors and quickly narrowed down the candidate list of graphene-supported single-atom catalysts. This approach can be easily used to screen and design other SACs and significantly accelerate the catalyst design for many other important reactions.