Dynamic Site Heterogeneity in Nanoparticle Catalysis from Hierarchical Machine Learning

Identifying the activity of all individual surface sites on catalytic nanoparticles is challenging but critical for heterogeneous catalysts, especially for understanding site heterogeneity. Here, we developed a hierarchical machine learning approach that links the atomic local environment to both energy and reaction barrier predictions with similar computational costs, opening the feasibility of extensive catalytic reactivity exploration on nanoparticle catalysts. Applying the methodology to propane dehydrogenation at a reaction temperature of 973 K on Pt nanoparticles up to a size of 5 nm revealed a highly dynamic and heterogeneous activity with a wide distribution of propane C–H cleavage barriers (0.5–1.3 eV) and rare highly active or inactive sites. Conventional global optimization methods underestimate the catalytic activity, even by considering a large set of metastable structures. Instead, accurate activity and size effect trend emerge by using ensemble-averaged molecular dynamics trajectories, emphasizing the dynamic nature of site activity and heterogeneity. Notably, classical polyhedral models, like those obtained from a Wulff construction, can predict correct trends for larger nanoparticles but do so through error cancellation between missing dynamics and overestimating probabilities of metastable sites. Furthermore, our simulations reveal a fast barrier fluctuation at the reaction temperature for a given surface Pt atom within 100 ps trajectories with a barrier shift by up to 0.8 eV, indicating that individual atoms can switch between active and inactive states on a picosecond time scale. This work provides a feasible approach to shift the views of catalysts from a combination of static active sites to highly heterogeneous and fast dynamically evolving sites.