Learning the Diffusion of Nanoparticles in Liquid Phase TEM Using a Physics-Informed Generative AI

Liquid phase transmission electron microscopy (LPTEM) has emerged as a promising technique for single particle tracking at the nanoscale, enabling us to visualize and characterize the motion and interaction with unprecedented spatiotemporal resolution. Here, we have leveraged the generative power of AI models to learn the physics of experimental trajectories obtained from LPTEM movies and generate synthetic single-particle trajectories. To this end, we have developed a generative AI model trained on a large data set of experimental trajectories of gold nanorods moving near the SiN membrane of the LPTEM chamber. We show the model is capable of learning the underlying correlations in sequence-based data and, thus, learning the time-dependent dynamics of the experimental trajectories in its continuous latent space representation based on statistical/physical properties. We demonstrate that our model can separate trajectories into different diffusion classes and recognize the differences in their statistical properties. More importantly, our model can be used as a black-box simulator for generating synthetic single-particle trajectories from LPTEM. The latter is an extremely useful feature in that it generates an unlimited number of trajectories for downstream tasks, e.g., in developing an AI-based workflow for automating in situ electron microscopy experiments.