Predicting Surface Strain Effects on Adsorption Energy with Graph Neural Networks

Abstract

Modifying the adsorption energies of reaction intermediates on different material surfaces can significantly improve heterogeneous catalysis by reducing energy barriers for intermediate elementary reaction steps. Surface strain can increase or decrease the adsorption energy depending on the surface composition, adsorbate composition, surface facet, and adsorbate site, breaking traditional scaling relationships which inhibit energy barrier alteration in reactions such as ammonia synthesis. We aim to generate a model that maps the adsorption energy response to a given input strain for a range of adsorbates and catalyst structures. After generating a training dataset of strained copper binary alloy catalyst + adsorbate complexes from the Open Catalyst Project and calculating the adsorption energy with first-principles calculations (dataset made available), we train a graph neural network to learn the relationship between catalyst + adsorbate structure, surface strain, and adsorption energy. The model successfully predicts the nature of the adsorption energy response for 85% of surface strains, outperforming simpler model baselines. Using the ammonia synthesis reaction as an example system, we identify Cu-S alloy catalysts as promising candidates for strain engineering since the majority of surface strain patterns raise the adsorption energy of the *NH intermediate. We find that the strain response of similar adsorbates on the same surface can greatly vary due to the competition between surface relaxation under strain and relaxation of the coordination environment. Our presented machine learning approach can be applied to additional datasets to identify target strain patterns that can reduce energy barriers in heterogeneous catalysis.