Determination of Nanocluster Size Distributions From HAADF-STEM images

Abstract

Metal nanoclusters on high surface area oxide supports are important catalysts for many applications. To a first approximation, the properties of each system depend on the size of the nanoclusters. It is therefore of great importance to determine the cluster size accurately. One of the methods to investigate individual nanoclusters is high-angle annular dark-field imaging (HAADF) in scanning transmission electron microscopy (STEM). Although a spherical aberration corrected STEM has a good spatial resolution, its intense probe can force nanoclusters move around the surface of the support to make the size measurement difficult. On other hand, a conventional STEM with a less intense probe can avoid the cluster movement at the expense of resolution as well as signal intensity. In this abstract, we discuss our first attempts to quantify the size distributions from STEM images and describe a method that produces reasonable results even for the smallest clusters when the image has a low signal-to-noise (S/N) ratio. Decaosmium carbido carbonyl (Os 10 C(CO) 24 ) and tritantalum (Ta 3 ) clusters were prepared on MgO and silica supports, respectively, by surface-mediated organometallic synthesis [1-2]. These clusters were examined using JEOL JEM-2500SE (S)TEM operated at 200 kV with the probe size of 0.2 nm and the collection angle of 35-90 mrad. Images were taken at a magnification of 4,000,000x and 6,000,000x for Os 10 C(CO) 24 and Ta 3 clusters, respectively. Fig. 1(a) shows a raw HAADF-STEM image for Os 10 C(CO) 24 clusters on a MgO support. The pixel size in the image is 0.063 nm. It is very difficult to measure the cluster size from the raw image because of the noisy intensity profile as shown in Fig. 1(b). Although noise in the image represents the primary difficulty, the size distribution of the intensity profile will also be modified by any blurring effect in the final image -initial probe size, beam broadening, cluster orientation effects. We can smooth the noise in the image by broadening the image using a Gaussian blur, in which each pixel is convoluted with a Gaussian (normal) distribution function with a standard deviation (σ value). When the σ value is small, the noise is not completely smoothed as shown in the profiles (Figs. 1(c) and (d)). When the σ value is large enough, however, the noise is smoothed out so that the cluster size can be measured with the profiles as shown in Figs. 1(e)-(g). Fig. 2 shows a size distribution histogram for twenty clusters in a blurred image with a σ value of 0.25 nm. The averaged sizes of Os 10 C(CO) 24 clusters in the Gaussian blurred images are plotted in Fig. 3(a) as a function of σ value. The error-weighted linear fit intercepts the vertical axis at the value of 0.61 0.11 nm. This value corresponds to the cluster size without Gaussian blur and is slightly larger than the Os-Os-Os atomic distance in Os ± 10 C(CO) 24 clusters (0.57 nm) that obtained with

Electron Microscopy of Solid Catalysts—Transforming from a Challenge to a Toolbox

“…In addition to the noise in the image, the obtained intensity profile across the clusters can be further modified by any blurring effect in the STEM image−initial probe size, beam broadening, and cluster orientation. 339 A Gaussian blur function can be used to smooth the noise in the images: each pixel is convoluted with a Gaussian normal distribution function with a standard deviation so that a clearly recognizable contour of the particle can be obtained. 340 Care must be taken not to oversmooth the profiles and thus enlarge the size of the clusters artificially.…”

“…For the small clusters imaged in STEM with low signal–noise ratio (for instance those in Figures and ), a direct measurement of the cluster size from the raw image could be inaccurate. In addition to the noise in the image, the obtained intensity profile across the clusters can be further modified by any blurring effect in the STEM image–initial probe size, beam broadening, and cluster orientation . A Gaussian blur function can be used to smooth the noise in the images: each pixel is convoluted with a Gaussian normal distribution function with a standard deviation so that a clearly recognizable contour of the particle can be obtained .…”