Fig. 1: Overview of LEONARDO workflow.
From: Learning the diffusion of nanoparticles in liquid phase TEM via physics-informed generative AI
a Schematic overview of our workflow for extracting single particle trajectories from LPTEM movies. By imaging the stochastic motion of single gold nanorods in water as they move and interact with the window of the liquid cell microfluidic chamber of LPTEM, we collect a large dataset of single-particle trajectories from LPTEM experiments. b Schematic of the LEONARDO model, a transformer-VAE with self-attention mechanisms in the encoder and decoder, mapping input trajectories to a low-dimensional latent space and reconstructing them with the aid of a physics-informed loss function. c Latent space representation of unseen trajectories encoded by the trained model, showing distinct clustering and overlaps indicative of different diffusion behaviors. d Demonstration of the generative power of LEONARDO in simulating new synthetic LPTEM trajectories by sampling from different regions of the latent space.
