Learning the diffusion of nanoparticles in liquid phase TEM via physics-informed generative AI

Abstract

Abstract The motion of nanoparticles in complex environments can provide us with a detailed understanding of interactions occurring at the molecular level. Liquid phase transmission electron microscopy (LPTEM) enables us to probe and capture the dynamic motion of nanoparticles directly in their native liquid environment, offering real time insights into nanoscale motion and interaction. However, linking motion to interactions to decode the underlying mechanisms of motion and interpret interactive forces at play is challenging, particularly when closed-form Langevin-based equations are not available to model the motion. Herein, we present LEONARDO, a deep generative model that leverages a physics-informed loss function and an attention-based transformer architecture to learn the stochastic motion of nanoparticles in LPTEM. We demonstrate that LEONARDO successfully captures statistical properties suggestive of the heterogeneity and viscoelasticity of the liquid cell environment surrounding the nanoparticles.