(8a) Understanding Barrierless Adsorption in Electro-Catalytic Systems Using Accelerated Transition Path Sampling

The adsorption and desorption processes that often initiate or terminate catalytic reaction sequences are challenging to account for in kinetic models, since they are often barrierless processes where conventional transition-state theory (TST) is not applicable. In particular, the adsorption or desorption of carbon monoxide (CO) on metal surfaces is a very important process as it is a rate limiting step for many catalytic reactions. The free energy profiles of these processes are even more difficult to calculate when there is solvent present, because the potential energy surface (PES) will be rugged and fluctuating. One approach to studying the kinetics and thermodynamics of a process such as this is to employ a rare event sampling method such as transition path sampling (TPS) which requires molecular dynamics simulations to generate dynamical trajectories of all the atoms in the system. However, molecular dynamics using density functional theory (DFT) is highly computationally expensive and not practical. Data-driven machine learning force fields exploit machine learning methods to quickly predict forces and energies in atomistic simulations. We use classical force-fields as well as machine-learned potentials to carry out TPS to calculate free energy profiles of adsorption/desorption of CO on Pt(111). We use the Atomistic Machine-learning Package (Amp) to train a high dimensional neural network potential on the reference data calculated using DFT GGA (BEEF-vdW) with finite-difference mode. First, we perform TPS calculations for desorption of CO at solid-liquid interface to identify all of the relevant reaction coordinates. Then we use meta-dynamics simulations to calculate free-energy profile across multivariable as reaction coordinates and compare results with umbrella sampling calculations done along one reaction coordinate. We also compare the CO desorption at solid-liquid interface with solid-gas interface to quantify the effect of solvents on the dynamics of the desorption process. This approach can be easily generalized to other machine-learning potentials and opens up a path to study kinetics of systems with non-smooth PES.