Improved Representations of Heterogeneous Carbon Reforming Catalysis Using Machine Learning.

Predicting adsorption energies of reaction intermediates is critical for determining catalytic reaction mechanisms. Here, we present three combined representations for predicting adsorption energies of carbon reforming species on transition-metal surfaces. Among the three combined representations, the Elemental Properties and Spectral London Axilrod-Teller-Muto (EP&SLATM) representation, which uses separate EP and SLATM representations for the surface and adsorbates, yields the lowest mean absolute error (MAE) of ∼0.18 eV with respect to density functional theory (DFT) adsorption formation energies for 68 adsorbates on four low-index metal facets (Cu(111), Pt(111), Pd(111), Ru(0001)). All three combined representations also have lower MAEs compared with linear scaling relations. Notably, two of the combined representations achieve their results using empirical/experimental molecular structures only (i.e., without recourse to structural optimization based on first-principles methods such as DFT). The combined representations enable improved efficiency for predicting heterogeneous catalytic mechanisms using machine learning approaches, largely bypassing expensive electronic structure calculations. Further, we show that the combined representations enable "cross-surface" training with regression and tree-based machine learning methods. That is, to predict adsorption formation energies on a particular catalyst metal, these methods only need a small amount of training samples (20%) on that metal.