Physics Main Calendar

Abstract: We present PtyRANNOSAUR, a data-driven neural network model that rapidly reconstructs atomic resolution electron ptychography data in less than 10 seconds on an A100 GPU. PtyRANNOSAUR uses a convolutional autoencoder to map 4D-scanning transmission electron microscopy (STEM) data to 2D phase images for samples up to 25 nm. Each model is trained on a large database of crystal structures and is tailored for a range of experimental parameters, such as accelerating voltage, convergence angle, defocus, and sample thickness. This approach yields high quality reconstructions without requiring any fine-tuning of hyperparameters. By testing PtyRANNOSAUR on experimental and simulated data under a range of conditions, we show that the neural networks are capable of accurately reconstructing a broad range of atomic structures, including single crystals, twisted van der Waals materials, nanoparticles, and multilayer thin films. Notably, the model operates 10-100x faster than iterative methods while producing images of comparable and sometimes higher quality. These advances enable near-live ptychography reconstructions powered by machine learning.