Graphs and Kernelized Learning Applied to Interactions of Hydrogen with Doped Gold Nanoparticle Electrocatalysts

Abstract

Understanding hydrogen adsorption on metal nanoparticles is a key prerequisite for designing efficient electrocatalysts for water splitting and the hydrogen evolution reaction. However, this seemingly simple elementary reaction step is affected by several factors arising from the chemical environment at the catalyst, and deciphering the most important contributions to optimal interactions requires numerically heavy electronic structure calculations. Here, we combine graph-based representations of the local atomic environment of hydrogen in copper- and palladium-doped 25-atom gold nanoparticles with several kernel-based machine learning (ML) methods to predict the interaction energy between hydrogen and the nanoparticle catalyst. We demonstrate that simple distance-based kernel models are able to predict the interaction energy within 0.1 eV when trained by reference data from state-of-the-art density functional theory calculations. Analyzing the model performance with respect to attributes of the hydrogen node highlights the locality of hydrogen adsorption. This implies the viability of combining graphs with kernel-based ML models for studying hydrogen chemisorption in complex environment data efficiently.