Atomic Defect Identification with Sparse Sampling and Deep Learning

Scanning Transmission Electron Microscopy (STEM) is a high-resolution characterization technique that can resolve atomic lattices. Recent advances in high-brightness sources enable in-situ STEM experiments with acquisition rates reaching 20 frames per second. However, high-doses of electron radiation damage the atomic lattice and limit frame-rates. Thus the real-time visualization of lattice transformations with sub-angstrom resolution requires innovative tools to track defect evolution while limiting electron radiolysis. Here we present a trained deep learning model that automatically tracks lattice defects in graphene while scanning around 60% of the total area for an in-situ experiment. Our approach extracts relevant physical information from STEM movies without requiring any knowledge of the lattice symmetry. Atomic defect identification using sparse sampling and deep learning was demonstrated on a multi-image dataset of defects in graphene as part of the 2021 Smoky Mountains Data Challenge.