(521ds) Predicting the Energies of Adsorbate-Saddle Point Interactions Using Machine Learning

Heterogeneous catalysis underpins many important chemical technologies. It is typically very expensive or generally infeasible to run experiments at all conditions of interest so it is useful to be able to generate microkinetic models to describe chemistry without the need for specific experiments. However, these microkinetic models require estimation of many important thermodynamic and kinetic parameters. In particular, at high coverages adsorbates and saddle points have significant energetic interactions with adjacent adsorbates that affect both thermochemistry and kinetics. While computational methods exist for accounting for these coverage dependent effects they tend to be highly expensive and specific to the adsorbates involved. We present a tool for predicting the energetics of arbitrary inter-adsorbate interactions. We generated a dataset of optimized geometries of saddle points for a diverse set of reactions interacting with a variety of co-adsorbates. Mapping these structures into 2D we are able to train a subgraph isomorphic decision tree (SIDT) to predict interactions between arbitrary adjacent adsorbates. We demonstrate the accuracy of the generated SIDT and discuss the limitations of the approach.