Next generation electron microscopes equipped with aberration correctors and sensitive detectors have enabled routine collection of atomic-scale information with sufficient signal-to-noise ratio from a variety of technologically relevant materials. In the case of Scanning Transmission Electron Microscopy (STEM), the ability to focus a large electron flux into a picometer probe has also benefitted the collection of correlative spectral information, primarily in the forms of Electron Energy Loss

more »

... pectroscopy (EELS) and Energy Dispersive X-ray Spectrometry (EDS). Collection of atomic scale EELS and EDS spectra over a grid of pixels, referred to as a spectrum image, is now routinely achievable outside of specialized labs. However, the large inelastic background component in EELS and the limited available detector crosssections in EDS means that each pixel of the spectrum image requires a large integrated dwell time (on the order of milli-seconds to seconds). This results in a relatively poor dose-efficiency, where the material must be subjected to a large incident flux for a comparatively small signal. For atomically resolved spectrum images, there are a limited number of materials with sufficient radiation hardness to withstand such a damaging measurement, as each atomic column may often be subject to multiple integrated acquisitions. Many materials of interest often gain often gain their structure from a complex arrangement of weaker bonds between a hybrid set of metal and organic elements (metal/covalent organic frameworks), which are easily altered due to radiolysis. The long dwell times per pixel also mean total acquisition times for large maps can stretch from hours, to even days. Ensuring a sufficient field of view (FOV) such that the atomic level information is truly representative of the overall material structure is challenging, and often a compromise must be made between sample stability, final resolution, and FOV. Random sub-sampling has been proposed to both reduce the integrated area dose and time-to-collect highresolution images in Scanning Electron Microscopy (SEM) [1]. Observing that not all pixels in an image contribute meaningful information or contrast, the missing pixels can be restored off-line by exploiting statistical correlations between neighbors. Using a random mask assumes important pixels are also randomly distributed, which is generally not true. Thus, many sub-sampling techniques are limited by their 'static' masks and cannot obtain sampling percentages lower than 15-20%. Approaches which can tune the mask of selected pixels 'on-the-fly', referred to as adaptive or dynamic sampling, have recently been shown to achieve state-of-the-art image quality at extremely low sampling rates [2, 3] . Such algorithms follow a general approach as follows: collect an initial set of measurements