Exposing New Atomic‐scale Information about Materials by Improving the Quality and Quantifiability of Aberration‐corrected STEM Data

Abstract

Aberration‐corrected scanning transmission electron microscopy (STEM) is providing previously unattainable views of materials at the atomic scale. The quality of STEM data is now often limited by environmental and experimental factors instead of instrument factors (e.g. electron optics). Some of these environmental limitations can be overcome by collecting and processing STEM data using new data science techniques. These techniques expose new atomic‐scale materials information by improving our ability to measure atomic column positions, 3D structure, single point defects, and atomic‐scale composition. The precision in locating atomic column positions in STEM images is fundamentally limited by the image signal to noise ratio (SNR), but typically practical limits including sample and microscope instabilities that produce distortions in STEM images are encountered before reaching the SNR limit. We have developed a non‐rigid registration (NRR) technique that corrects image distortion of all length scales and enables averaging to enhance the SNR.[1–2] Sub‐pm precision images of single crystal materials have been achieved by NRR and averaging high angle annular dark field (HAADF) STEM images. NRR has allowed measurements of pm‐scale bond length variations of Pt nanocatalyst atoms that may help explain their catalytic activity. Figure 1 shows NRR and averaged HAADF STEM data of a Pt nanocatalyst on an alumina support from a side‐view that exhibits ˜2 pm precision. Displacement measurements of each atomic column position reveal moderate Pt surface bond length contraction and strong but localized strain of Pt atoms near nanoparticle‐support interface. Some of the interface strain is transferred up the twin boundary. Determining the three dimensional atomic structure of materials from two dimensional S/TEM images is a major hurdle. The standardless atom counting technique is one promising route to measure local sample thickness by quantitatively comparing experimental and simulated HAADF STEM images and can be used to deduce 3D structure.[3] Unlike previous examples, NRR and averaging STEM images have allowed standardless atom counting with the uncertainty no longer dominated by Poisson noise [1]. This should allow the unique determination of the number of atoms in atomic columns, although this has not yet been demonstrated due to other sample limitations. Point defects are critical to the properties of a wide range of materials, but imaging single defects is challenging. Quantitative STEM has allowed imaging single substitutional and interstitial dopant impurity atoms[4], but experimentally imaging single vacancies has remained elusive. We have used HAADF STEM frozen phonon multislice simulations to predict the detectability of La vacancies in LaMnO 3 by the reduced atomic column intensity and atomic column distortions around the vacancy. NRR and averaging HAADF STEM images of LaMnO 3 improves the SNR and the image precision sufficiently to potentially detect single La vacancies. Experimental images contain candidate single La vacancies that have local atomic column distortions and intensity variations which match simulated predictions. Atomic‐resolution composition maps can be created using STEM energy dispersive x‐ray spectroscopy (EDS) spectrum imaging (SI). However, long total dwell times that may introduce spatial distortions are required because of low x‐ray production and collection efficiency. The most common approach to minimize distortions is to sum multiple SIs using online drift‐correction software that discards the individual SIs and HAADF images. The quality of EDS SIs of a Nd 2/3 TiO 3 sample was improved by saving the simultaneously acquired raw HAADF and EDS SI series, and applying post acquisition NRR and averaging. The resulting elemental maps show less spatial distortions and more atomic localization of x‐rays. In addition, a novel non‐local principle component analysis further enhances the quality of EDS SIs compared to conventional denoising methods.[5]