Atomic Structure-free Representation of Active Motifs for Expedited Catalyst Discovery

Abstract

For CO* and H* binding energy prediction, we develop new representation of catalyst surface which split surface into three types of site, first nearest neighbor of adsorbates and second nearest neighbor in same layer and sublayer. From this representation and machine learning regression model, we achieve reasonable accuracy (0.120 eV for CO* and 0.105 eV for H*) with quick training (~200 sec using CPU). Because our representation does not require density functional calculation and atomic structure modelling, it can predict binding energies of possible active motifs without time-consuming steps.