Modern Scanning Transmission Electron Microscopes (STEM) provide sub-Angstrom beam sizes and high spatial coherence. The fundamental challenge to a quantitative analysis of scattering in the STEM is a difficult inverse problem with ubiquitous dynamical scattering. In this talk, we present examples where a deep learning approach to this inverse problem was successfully applied and discuss its advantages and current limitations.
First, we show that deep convolutional neural networks (DCNN), trained on multislice simulations, learn to accurately predict the 3-D oxygen octahedral rotations in complex oxides from annular bright-field images, with sub-degree accuracy and unit-cell resolutions. Second, we show that DCNN can successfully “invert” coherent convergent-beam electron diffraction (CBED) data to uncover the local 3-D atomic structure for a limited set of material classes, with sub-Angstrom resolutions.
This research used resources of the Oak Ridge Leadership Computing Facility (supported by the Office of Science of the Department of Energy).