Advanced Compressive Sensing and Dynamic Sampling for 4D-STEM Imaging of Interfaces

Abstract

Interfaces in energy materials and devices often involve beam-sensitive materials such as fast ionic, soft, or liquid phases. 4D scanning transmission electron microscopy (4D-STEM) offers insights into local lattice, strain charge, and field distributions, but faces challenges in analyzing beam-sensitive interfaces at high spatial resolutions. Here, a 4D-STEM compressive sensing algorithm is introduced that significantly reduces data acquisition time and electron dose. This method autonomously allocates probe positions on interfaces and reconstructs missing information from datasets acquired via dynamic sampling. This algorithm allows for the integration of various scanning schemes and electron probe conditions to optimize data integrity. Its data reconstruction employs a neural network and an autoencoder to correlate diffraction pattern features with measured properties, significantly reducing training costs. The accuracy of the reconstructed 4D-STEM datasets is verified using a combination of explicitly and implicitly trained parameters from atomic resolution datasets. This method is broadly applicable for 4D-STEM imaging of any local features of interest and will be available on GitHub upon publication.