Compressed 4D-STEM—From Conception to Implementation

Scanning transmission electron microscopy has routinely shown its capability to identify complex structural, mechanical, and electronic properties of crystalline materials. However, there are many more materials where the resolution of the final image is limited by its response to the incident electron beam, and ultimately reliable images and analysis is determined by how beam-sensitive the sample is. Recent developments in compressive sensing for scanning transmission electron microscopy (CS-STEM) has opened new avenues for low dose electron microscopy, as well as increased temporal resolution through subsampling for fixed detector imaging methods such as annular dark field (ADF) and bright field (BF) imaging.<br/><br/>4D-STEM is becoming more common in the field, where at each probe location a convergent beam electron diffraction (CBED) pattern is acquired using a pixelated detector to allow for a higher angular range of information from electron scattering. The CBEDs can then be used to acquire a variety of image types such as (high angle) ADF (HA/ADF), BF, ABF, Centre of Mass, Phase Contrast, Strain Mapping, Holography, and Ptychography. However, due to the nature of the data acquired, the time for acquisition is much slower with respect to fixed detectors, and the data sets are on the order of many gigabytes as opposed to megabytes in standard imaging methods. This means that beam exposure increases and other artefacts such as drift and distortion are heighted through the increased acquisition time.<br/><br/>In order to overcome this, we propose integrating compressive sensing with 4D-STEM to developed a new method which aims to both lower the electron dose and increase speed through real and reciprocal space subsampling. We demonstrate practical implementation of subsampling the pixelated detector and subsampling the spatial probe locations to generate a 4D-STEM dataset that is not only significantly smaller in memory size, but is acquired in significantly less time than the full acquisition. The data is then reconstructed through an image inpainting algorithm by first recovering the CBED patterns and then in turn using the CBED patterns to recover a variety of subsampled image types. We present the methods and implementation to our STEM, and the potential for ptychographic reconstruction from subsampled data.