Sampling real-time atomic dynamics in metal nanoparticles by combining experiments, simulations, and machine learning

The atomic dynamics of metal nanoparticles (NPs), prominent already at low temperatures, is crucial for their properties but also challenging to elucidate. Recent advances in experimental approaches may provide atomically resolved snapshots of the structure of NPs in relevant regimes, but limitations in experimental data acquisition hinder the reconstruction of the atomic dynamics present within them. Molecular simulations -- typically starting from ideal/perfect NP structures -- allow tracking the motion of atoms over time, but suffer from limited sampling and provide results that, being dependent on the initial (putative) structure, are often only indicative. Here, combining state-of-the-art experimental and computational approaches, we demonstrate how it is possible to tackle the inherent limitations of both methods and resolve the atomistic dynamics present in metal NPs under realistic conditions. Annular dark-field scanning transmission electron microscopy (ADF-STEM) enables the acquisition of a time series of ten high-resolution images of an Au NP. Each image is taken at intervals of 0.6 seconds, providing data on a second timescale during the experimental sampling. These are used to reconstruct atomistic 3D structures of the real NP that are then used as starting configurations for ten independent molecular dynamics (MD) simulations. Unsupervised machine learning analysis of the data extracted from the MD trajectories using advanced structural and dynamical descriptors allows tracking and resolving the real-time atomic dynamics present within the NP under relevant conditions. This provides new perspectives into the realistic atomic dynamics within such NPs. We expect that such integrated experimental/computational approaches will become fundamental in various fields where the dynamics of NPs plays a key role, from catalysis to, e.g., nanoelectronics and biomedicine.