Compressed sensing for beam sensitive materials imaging in S canning T ransmission E lectron M icroscopy

Abstract

Transmission electron microscopy (TEM) is a very powerful technique to investigate materials down to their atomic components. Its versatility allows quantifying samples from their shape to the nature of their constituents and surroundings. However, the strong interaction of the electron beam with matter potentially induces damages in samples under investigation, especially for those composed of soft matter such as zeolites, metal organic frameworks (MOFs) and most life science samples. Current workarounds involve the reduction of the beam intensity, such as in the so‐called low dose imaging, or increase of the detector performances as provided e.g. by direct electron detectors. Recently, improvement in signal processing lead to the development of compressed sensing [1, 2], a numerical algorithm based on the assumption that real life images are sparse in some particular well‐chosen basis. Images can then be expressed with much less components than what required by the Nyquist sampling theorem. By extension, not all pixels in a given image are necessary and only a random selection of them is sufficient to retrieve the original image with fidelity. By definition, compressed sensing is a very dose efficient technique as only parts of the sample need to be exposed to the electron beam to reconstruct a faithful image. So far, only theoretical implementations of compressed sensing were investigated in the TEM community, focused mostly on the case of tomographic reconstructions [3]. Very recently, we demonstrated the first physical implementation of compressed sensing in a Scanning TEM (STEM) based on the use of a solenoid as a fast beam blanker [4]. The solenoid is placed in the condenser plane of a STEM in a specially designed condenser aperture holder with feedthrough electrical contacts. By synchronizing the STEM signal with the current source driving the solenoid, we successfully acquired compressed images by shifting the beam away from the region of interest on blanked pixels. The case of both medium scale imaging (Fig.1) and high resolution imaging (Fig. 2) was investigated and reconstructed using the SPGL1 algorithm [5] with a signal compression of 80%. The present setup, although still quite far from low dose imaging conditions, has lead to successful imaging of a cross grating and a SrTiO 3 sample. Latest results revealed that a checkerboard pattern, the most demanding pattern for the fast beam blanker, could be acquired with a dwell time of 40 µs and a beam deflection of 225 nm, as shown in Fig. 3. By further improving the experimental setup, we expect reducing the acquisition dwell time to values within the µs range. Such improvements will offer possibilities for realizing improved images and tomography experiments of electron beam sensitive materials.

Development of a Practicable Digital Pulse Read-Out for Dark-Field STEM

“…The future outlook for this approach may include a shift to faster clock signal readouts, as well as a move to the new generation of solid-state diode detectors (Si-PMTs) (Buzhan et al, 2003). We would hope to make further use of other timing signals to more seamlessly automate the line and frame trimming, and in future, our approach could be coupled with novel beam-blanking schemes to eliminate entirely the wasted dose lost during flyback and between frames (Béché et al, 2016). We could also move towards in-hardware pulsecounting, however this would also lead to increased costs.…”

Fast reconstruction of atomic-scale STEM-EELS images from sparse sampling

“…These results raised a lot of interest toward inverse problems which estimate the image based on partial spatial acquisitions which is referred to as inpainting. It remains an active research area for STEM [8,9] and SEM [4], among others.…”

Reconstruction of Partially Sampled Multiband Images—Application to STEM-EELS Imaging

“…Standard inpainting methods for 2D images are now well-documented and can be applied to two-dimensional HAADF images, for instance by minimizing the Sobolev norm or the total variation or the image (see for instance [16], [17]). Alternative regularizations include dictionary learning, proposed to solve the sensitive-material problem in HAADF acquisition [4], [18]- [20], leading to satisfactory results. However, in the case of EELS multi-band images (hereafter referred to as spectrum-images), the spatial-spectral structure of the data needs to be taken into account within the reconstruction process.…”