Predicting strain effects on adsorption energy based on atomistic structure and density of states

Strain can significantly tune potential energy landscape and the activation energy of chemical reaction, which changes the catalyst property and surface mechanochemistry process. But many structural degrees of freedom results in high dimensional data space, which brings challenge to evaluating the strain regulation performance and unveiling the regulation mechanism. Here, we propose a machine learning based approach to predict adsorption energy of strain tuned oxygen-alloy substrates systems based on two types of information, i.e., atomic structure and electronic density of states (DOS), respectively. Through the former, the relationship between the atomic structure and oxygen adsorption energies is established, but lack of physical explanation on the regulation mechanism. With the aid of the latter, it is demonstrated that the adsorption energies can be also well predicted using DOS features extracted from graph convolution and statistical analysis. The explainable machine learning Methods enables us to identify the physical significance of the moments of DOS and explain the limitations of D-band theory. This finding of this work may deepen our understanding of chemical reaction potential energy surface in response to strain and enlighten us how to construct the structure-spectrum-property relationship for other mechanochemistry problems.