(582e) Predicting Thermochemical Properties of Hydrocarbon Adsorbates on Metal Surfaces Using Machine Learning

One of the major challenges in studying the catalysis of surface reactions comes from the complexity of the reaction network. That is, tremendous numbers of surface intermediates/elementary reactions are implicated in many cases, from which the reaction mechanisms along with the rate-limiting steps are hard to identify not only with experiment but also with pure first-principles based calculations (e.g., density functional theory (DFT)) due to the high computational cost. The recent advent of the machine learning (ML) techniques could be applied in computational catalysis to help address this challenge enabling us to predict the thermodynamic properties of an enormous number of species at an affordable cost. In this work, we use a combination of DFT calculations and ML models to explore methods for the prediction of thermochemical properties of hydrocarbon adsorbates on close-packed surfaces.

As a first step, electronic energies and vibrational frequencies for 183 surface derivatives of hydrocarbons on Pt(111) and Ni(111) were calculated with periodic DFT calculations. These results were used as a dataset for subsequent ML models. Representative C₁ to C₆ hydrogenated carbon species were selected to include both linear chains and ringed species so that the following ML models can capture much of the adsorbate-surface interaction. For ML models, a molecular fingerprint representation that can be generated easily from SMILES notation avoids computational effort to translate chemical structures into numerical datasets. Various ML regressions were implemented for property prediction. An enhanced molecular fingerprint was able to predict binding enthalpies with mean absolute error of 6.7 kcal/mol. Predicting hydrocarbon binding for large numbers of surface fragments can help improve our understanding of many industrial reactions such as coke formation at high temperature on metal surfaces.