Elucidating the Fluxionality and Dynamics of Zeolite-Confined Au Nanoclusters Using Machine Learning Potentials

Sub-nanometric clusters (NCs) of transition metal (TM) atoms, typically consisting of fewer than 15 atoms, have exhibited remarkable catalytic activity in various industrial reactions. However, these NCs are thermodynamically unstable and susceptible to deactivation due to sintering effects. Previous experiments have proposed zeolites as effective structural supports to stabilize these NCs. Still, there has been limited exploration of the long-timescale dynamics, including fluxionality and diffusivity, of these zeolite-confined TM NCs (TM@zeolites). Traditionally, investigating dynamics on the timescale of a few nanoseconds (ns) has been challenging using conventional \emph{ab initio} molecular dynamics (AIMD) simulations. This paper uses a self-adaptive workflow that leverages two state-of-the-art machine learning potential (MLP) packages, SchNetPack and Neuroevolution potential (NEP). This workflow was developed via multiple iterations of training, utilizing both AIMD and classical molecular dynamics (MD) based metadynamics simulations under an adaptive sampling framework known as query-by-committee active learning. The result was a highly accurate and robust MLP capable describing the diffusion of Au nanoclusters in zeolites. The MLP was found to be transferable across different temperatures and scalable to different zeolite topologies.