(521dr) Elucidating the Fluxionality and Dynamics of Zeolite-Confined Au Nanoclusters Using Machine Learning Potentials

Sub-nanometric clusters of transition metal (TM) atoms (usually < 13 atoms) have demonstrated remarkable catalytic activities for several reactions, including low temperature CO oxidation, propylene epoxidation, and oxidation of aldehydes. However, as these nanoclusters (NCs) are not thermodynamically stable, significant catalyst deactivation due to sintering is often observed over time. One approach for overcoming these limitations relies on stabilizing the NCs within the pores of a zeolite. Although several experimental studies have demonstrated the efficacy of this approach, the atomistic processes associated with catalyst deactivation remain under-explored. Specifically, while density functional theory (DFT) is routinely used to study surface reactions, modeling the fluxionality of zeolite-confined transition metals (denoted as TM@zeolite) remains challenging. Here, using Au@LTA as a prototypical example, we show that graph neural network-based machine learning potentials (GNN-MLPs) can be used to investigate the size, shape, and dynamics of Au NCs. The PaiNN-based model is developed using ab-initio molecular dynamics (AIMD) data that span various NC sizes (1 – 10 atoms) and ten zeolite topologies. Furthermore, as traditional nudged elastic band methods do not capture the fluxionality of the NCs, we have used enhanced sampling techniques (e.g., metadynamics and the weighted ensemble method) to calculate the free energy barriers of diffusion within the zeolite pores. This work highlights the growing relevance of MLPs in describing phenomena that remain currently intractable using DFT. We anticipate that our findings will facilitate the rational design of TM@zeolite catalysts for industrially relevant reactions.