Predicting the Enthalpy of Hydrocarbon Radicals Adsorbed on Pt(111) Using Molecular Fingerprints and Machine Learning

The reliable prediction of properties for the adsorbates, including their enthalpy, has been a long-standing challenge as a first key step in studying surface reactions. It is especially difficult when large adsorbates are involved as the interactions between the adsorbates and surface atoms are complex. Here, we developed machine learning (ML) models for the prediction of the formation enthalpy of various C-2 to C6 hydrocarbon adsorbates on the Pt(111) surface based on 384 density functional theory calculations. Focusing on larger and more intricate adsorbates, two-thirds of the total species were C6 species. Four molecular descriptors that represent the valency and bonding of individual carbons within the adsorbates were generated without intensive computation. They were subsequently used as the features of the ML models with three linear and four nonlinear algorithms. The models were developed with 30 different samplings of train/test sets, and their results were statistically analyzed to ensure the performance of the models. Nonlinear models, especially kernel ridge regression and extreme gradient boosting, outperformed linear models with lower absolute errors. The top two accurate models, based on these algorithms, also displayed remarkable robustness in predicting various species. Employing ensemble average voting with these two models, we achieved the lowest mean absolute error of 0.94 kcal/molC. Finally, ML was used to estimate the formation enthalpy of 3115 hydrocarbon adsorbates on Pt(111), highlighting the promise of these methods to study more complicated reaction networks.