Machine Learning for Phase Retrieval from 4D-STEM Data

Modern direct electron detectors enable recording the full scattering distribution from an electron beam focused to atomic dimensions at imaging speeds, producing 4-dimensional, phase-space data sets. Current data rates are Gb/s, with Tb/s expected shortly as detector speeds increase, following a Moore’s-law-like scaling as the underlying silicon technology improves [1,2]. Encoded in these large and growing data sets are both the phase and amplitude of the exit wave, from which the expectation values of physical operators can be reconstructed to learn the structure and fields in the probed sample. However, real space images of atomic structure need to be reconstructed from the scattering data. Applying a linear filter such as an annular dark field (ADF) mask quickly generates an interpretable image at the cost of discarding a large fraction of the data. On the other hand, ptychography takes full advantage of the diffraction data to significantly improve signal-to-noise and resolution but is computationally expensive and cannot be applied to data from thick samples.