Advancing 4D STEM Using Computational Techniques

This is the authors' abstract. We don't add key points for this paper.

Four-dimensional scanning transmission electron microscopy (4D STEM) is a state-of-the-art method for collecting transmitted scattering events in a STEM. Unlike integrating radial detectors, the full diffraction pattern is imaged in the far-field for each probe position, allowing momentum resolution of the transmitted electrons. Using these diffraction patterns, a variety of different computational methods can be employed to analyze the specimen, such as applying a virtual detector, calculating the center of mass, or retrieving the phase shift induced by the object [1]. The increased uptake of 4D STEM over recent years is no doubt attributed to (i) the advances in detector technologies, where single electron strikes can be measured with little to no noise [2], and (ii) the advances in 4D STEM analysis algorithms such as ptychography, differential phase contrast and parallax imaging [3]. Furthermore, the high frame rates of these direct electron detectors (in some cases exceeding 100kHz, but more typically on the order of 10kHz) has reduced the practical limitations of data-acquisition such as sample drift or beam damage. Despite this, 4D STEM remains at least an order of magnitude slower than conventional integrated detector imaging, and as such this typically limits the field of view to mitigate for drift and other instabilities, as well as potentially leading to beam induced effects such as damage and/or contamination [4]. Recent developments in probe sub-sampling for 4D STEM [5, 6] have allowed for reduced sampling requirements at the sample plane whilst acquiring 4D STEM data. Not only does this increase the speed of 4D STEM acquisition, it also allows the user to more accurately align their STEM with respect to the detector plane. Typical alignment involves using an integrated signal (such as Z-contrast) to correct for mis-tilt or low-order aberrations, but we shall utilise computed 4D STEM images to improve alignment. By using sub-sampling, feedback from the microscope can be achieved more rapidly, and thus the resulting acquisition will be both faster and better aligned. Fig. 1. demonstrates the improved image clarity when using a centre of mass magnitude image over a Z-contrast image, allowing for more precise alignment of specimen tilt. In this talk, we will present the live alignment, acquisition, and processing of 4D STEM data using sub-sampled measurements to demonstrate how 4D STEM experiments can be accelerated in practice. Live 4D STEM alignment can be done at the speeds of fixed radial detectors without sacrificing signal by leveraging the sparsity of the underlying images. By using a more sensitive modality such as centre of mass, the alignment can be fine-tuned to optimise the output of the final acquisition. Examples of Z-contrast (top row) and centre of mass magnitude (bottom row) simulated images for different specimen tilts with improved correction from left to right. Sub-sampled (top) and recovered (bottom) virtual image frames taken from an experimental acquisition of sub-sampled 4D STEM data as viewed in real time. The first image (a) is aligned, and then the magnification increased to induce damage (b) before demagnifying (c). The data was acquired using a Quantum Detectors Merlin EM detector on a JEOL ARM200F at CNR-IMM, Catania [5].