Fig. 1: Overview of the methods used in this paper.

a Multislice diffraction simulations of many samples with different crystal structures, compositions, orientations, and thicknesses, using various microscope parameters. b Augmentation of the simulated images by applying elliptic distortion, pattern shift, limited signal-to-noise, and background functions. c Deep-learning training. d Experimental geometry for diffraction pattern measurements. e Dataset preprocessing. f Inversion of experimental diffraction images to predict the structure factors using the FCU-Net trained in c.