Machine-learned potential makes long simulations of CO hydrogenation on rhodium affordable

The work

Technical · from the record

Gaussian Approximation Potentials were trained on DFT energies and forces for the hydrogenation of CO to CHO on Rh(111) and used in place of DFT to run the biased molecular dynamics required for umbrella integration. Training was iterative: biased dynamics with the current potential produced new configurations, which were selected by farthest point sampling, labelled with single-point DFT and added to the training set, with energies and forces reported converged after 17 iterations. Production umbrella integration used 100 ps trajectories per window at 523 K, amounting to 16 ns of dynamics overall, with separate potentials trained on revPBE+vdWsurf and BEEF-vdW reference data. The free energy barrier for CHO decomposition nearly vanishes at the revPBE+vdWsurf level and a barrier of 0.13 eV remains at the BEEF-vdW level, lower than the harmonic approximation estimates, while unbiased trajectories seeded near the barrier gave transmission coefficients of 0.87 and 0.81.

How AI was used

A Gaussian Approximation Potential, combining a two-body term with a SOAP many-body representation, was fitted to DFT energies and forces for the Rh(111) surface with CHO and CO+H adsorbates, serving as a surrogate for the DFT potential energy surface so that extensive biased sampling became tractable. The initial training set comprised CI-NEB path images, light-element dimer curves and optimised and rattled pristine surface configurations. Training then proceeded in an automated loop: the current potential ran biased molecular dynamics in randomly chosen umbrella windows, diverse new structures were extracted by farthest point sampling on kernel distances, evaluated with single-point DFT, used first as a validation set and then appended to the training set for the next fit; from iteration twelve onward the potential also ran NEB calculations so its minimum energy path could be compared with the DFT one. A one-dimensional collective variable was constructed as a linear combination of the C-H distance and the CO tilt angle along the DFT NEB path, and harmonic biasing potentials were placed along it. The converged potentials drove the production biased dynamics with a Langevin thermostat; free energy gradients from the window averages were integrated with Gaussian process regression to give the free energy surface and its uncertainty, from which transition state theory rate constants were computed alongside harmonic and Pitzer-Gwinn estimates built on DFT vibrational frequencies. The same potentials ran unbiased trajectories seeded in the initial state and near the barrier to estimate half-lives and transmission coefficients. A second potential was fitted to the same final configurations recalculated with BEEF-vdW, without further training iterations.