Application of Inpainting Methods for the Improvement of STEM-Based Time-Resolved Elemental Mapping

In situ electron microscopy has seen rapid growth due to the availability of commercial platforms that enable the analysis of microscopic processes under external stimuli [1]. However, compositional analysis during in situ experiment using complementary spectroscopy methods such as energy dispersive X-ray spectroscopy (EDX) had only seen limited application due to the acquisition times needed for spatially resolved mapping. Furthermore, we cannot simply increase the electron flux to reduce the acquisition times, because we also need to avoid radiation damage. As more sensitive detectors are being developed, the acquisition conditions can also be improved using compressive sensing (CS) methods such as inpainting algorithms that collect only a fraction of the standard images or data frames [2] To explore the limits of chemical analysis both during in situ and ex situ experiments, we performed analysis of Ni(OH)2 nanoplates with low concentrations of incorporated Fe within a liquid cell (Hummingbird Scientific) using a Spectra Ultra 300 kV S/TEM (Thermo Scientific) that is equipped with a state-of-the-art Ultra-X EDX detector (collection solid angle>4.45 Sr). Due to the serial map acquisition, time and electron dose ēD dependence can be adjusted by selecting individual frames form the EDX data cube. From these datasets, we rendered reconstructions using a Beta-Process Factor Analysis via Expectation Maximization algorithm (BPFA-EM) that is part of SenseAI’s Image-Inpainting software (Sense AI Innovations Ltd.) [3]. The advantage of this algorithm is that it relies on the input data itself to compute a solution without the need of training data and so, we can use the sequential EDX frames to optimize the results. [4]. In the work presented here, we will discuss reconstructions obtain from data acquired ex situ after reaction where we compare the strong Ni signal and the weak Fe signal at different sparsity levels and cumulative ēD yields. Fig. 1 shows the high angle annular darkfield (HAADF) image of the 200-300 nm catalysts beside the fully acquired (390 frames) Ni-Kα signal map, which showed clear edges and thickness contrast. The raw EDX spectrum also shows that Fe signal is above background noise. The limits of the inpainting are explored in Fig. 2 where we compare the reconstructions from different numbers of integrated frames (less is sparser). In the case of Ni, a full inpainted frame was recovered at both 40 and 4 frames (10% and 1% of the total signal). In contrast, a border effect is evident in the noisy Fe map, which an intermediate image filtering step performed using the concurrent HAADF image to avoid misrepresentation of signal from the support. At the lowest detection limits (∼0.01% of the maximum Ni signal) artifacts arise in the areas that are poorly probed. Possible optimization approaches and insights into in-situ experiments will be provided to further improving new avenues of research [5]. (A) HAADF image of the Ni(OH)2 catalyst observed ex situ after reaction. (B) The corresponding EDX-Ni map acquired over 390 frames at 16 e־/Ų for a total time of 790 s. (C) EDX spectrum from the first 40 frames of the data cube to highlighting the fringe presence of Fe which is incorporated by the Fe(NO3)3 electrolyte used during the in situ reaction. Scale bar = 200 nm. Unprocessed and processed Ni and Fe elemental maps at different cumulative e¯D. (A, B) Ni-K maps rendered using only 10% and 1% of the total e¯D by means of selecting less frames. (C, D) As before but for the Fe-K signal. (A′-D′) Map reconstructions applying an inpainting algorithm into their corresponding ‘noisier’ EDX maps. At only 4 frames (or 9 s time) the total irradiation is e¯D= 130 e־/Ų. Scale bar = 200 nm.