PtyRANNOSAUR: Ptychography with Robust Artificial Neural Networks Optimized for Sub-Angstrom Accuracy and Ultrafast Reconstruction

Abstract

We present PtyRANNOSAUR, a data-driven neural network code that reconstructs atomic resolution electron ptychography data in seconds, 10-100x faster than standard methods. PtyRANNOSAUR uses convolutional autoencoders to map 4D-scanning transmission electron microscopy data to 2D phase images. 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. In addition, the code handles spatial partial coherence, multiple scattering, and scan position errors, which are critical for state-of-the-art electron ptychography reconstructions. By testing PtyRANNOSAUR on experimental and simulated data, we show that the neural networks accurately reconstruct atomic structures of a broad range of materials systems and can achieve high resolutions of $<0.5$ {\AA}, comparable to the best iterative reconstructions of the same data. These advances enable near-live, state-of-the-art electron ptychography reconstructions.