Extrapolating Energetics on Clusters and Single-Crystal Surfaces to Nanoparticles by Machine-Learning Scheme

A Bayesian linear regression scheme using a local structural similarity kernel as a descriptor is used to predict the energetics of atoms and molecules on nanoparticles. Examination of the binding energies of N, O, and NO with RhAu alloy single-crystal surfaces and particles indicates that regression models predict the binding energies on nanoparticles having diameters greater than 1.5 nm within an error range of 100–150 meV when the DFT data on single-crystal surfaces are used for training. By contrast, when the DFT data on small clusters are used for training, the regression models produce an error range of 200–400 meV. Kinetic analyses using the predicted energetics of the direct decomposition of NO on RhAu nanoparticles indicate that catalytic activity increases with a decrease in the particle diameter to 2.0 nm, whereas the activity drops when the diameter decreases to 1.5 nm. Detailed examinations of the free energy diagrams and the structures of active sites indicate that the drop in catalytic activity derives from the disappearance of active alloyed corner sites on the small nanoparticles as a result of Au segregation at the corners of narrow facets.