Rapid Estimation of Catalyst Nanoparticle Morphology and Atomic

Oct 23, 2014 - Katherine E. MacArthur,. †. Vidar T. Fauske,. ‡. Antonius T. J. van Helvoort,. ‡ and Peter D. Nellist. †. †. Department of Ma...
2 downloads 0 Views 4MB Size
Letter pubs.acs.org/NanoLett

Rapid Estimation of Catalyst Nanoparticle Morphology and AtomicCoordination by High-Resolution Z‑Contrast Electron Microscopy Lewys Jones,*,† Katherine E. MacArthur,† Vidar T. Fauske,‡ Antonius T. J. van Helvoort,‡ and Peter D. Nellist† †

Department of Materials, University of Oxford, OX13PH Oxford, United Kingdom Department of Physics, Norwegian University of Science and Technology (NTNU), NO-7491 Trondheim, Norway



S Supporting Information *

ABSTRACT: Heterogeneous nanoparticle catalyst development relies on an understanding of their structure−property relationships, ideally at atomic resolution and in three-dimensions. Current transmission electron microscopy techniques such as discrete tomography can provide this but require multiple images of each nanoparticle and are incompatible with samples that change under electron irradiation or with surveying large numbers of particles to gain significant statistics. Here, we make use of recent advances in quantitative dark-field scanning transmission electron microscopy to count the number atoms in each atomic column of a single image from a platinum nanoparticle. These atom-counts, along with the prior knowledge of the face-centered cubic geometry, are used to create atomistic models. An energy minimization is then used to relax the nanoparticle’s 3D structure. This rapid approach enables high-throughput statistical studies or the analysis of dynamic processes such as facet-restructuring or particle damage. KEYWORDS: catalysis, nanometrology, aberration corrected STEM, quantitative ADF, platinum nanoparticles

M

microscopy (HRTEM), we are limited by the fact that any given micrograph is a two-dimensional (2D) projection of what is really a three-dimensional (3D) object.5 Electron tomography provides one potential solution to this problem by reconstructing a sample’s 3D morphology based on many individual 2D projections (with different viewing angles).6 At atomic resolution, this can be achieved using “discrete tomography”.7 This utilizes the a priori knowledge that atomic-scale samples can be represented as an array of discrete points (atoms) and drastically reduces the number of unique projection orientations required to two as long as they are down low order zone axis.8,9 However, although this does unambiguously solve the 2D data projection problem, axiomatically, all tomographic techniques require multiple exposures to the beam to be recorded from the same region of sample, something which is not compatible with highthroughput analysis or with the imaging of dynamic processes such as electron-beam induced changes, in situ heating, or gas experiments. To attempt to overcome these challenges, here, we present an alternative to STEM tomography, using precisely calibrated atomic-resolution ADF STEM10 to extract as much information as possible from only a single experimental image. These data

etallic nanoparticles are of huge importance for catalyzing a wide variety of reactions, for example, oxygen reduction at the cathode of hydrogen fuel cells.1 To relate the catalytic abilities of these nanoparticles to their structure, it is essential to characterize their surface structures at the atomic scale. Thermodynamic calculations predict singlecrystals of face-centered cubic (FCC) materials such as platinum will form truncated octahedral shapes, with the ratio of their {111} and {100} type surfaces predicted by considering the ratio of their two respective surface energies.2 This ratio can be significantly and controllably modified through the use of surfactants and this has been demonstrated to have a substantial effect on catalytic selectivity.3 Particle morphology also dictates the nearest-neighbor coordinations of exposed surface atoms, broadly affecting the catalytic activity.4 Transmission electron microscopy allows for these catalytic nanomaterials to be imaged directly and in real-space. Among the range of techniques available, perhaps the most readily interpretable is that of annular dark-field scanning transmission electron microscopy (ADF STEM) due to the inherent atomic number dependency, or Z-contrast, of the images. This makes heavy metallic nanoparticles stand out from their lower atomic number supports. Since the development of hardware aberration correction for these microscopes, ADF imaging and simultaneous spectroscopy can now be performed routinely at atomic resolution, giving direct and site-specific information about samples. However, just as for high-resolution transmission electron © XXXX American Chemical Society

Received: July 20, 2014 Revised: October 10, 2014

A

dx.doi.org/10.1021/nl502762m | Nano Lett. XXXX, XXX, XXX−XXX

Nano Letters

Letter

(≈10−7 bar) undergoing structural modification under the electron-beam, something impossible to study by tomography. Pure platinum nanoparticles were taken from a production batch of a commercially manufactured hydrogen-fuel-cell catalyst with particles ranging 2−10 nm in diameter supported on an amorphous 3D carbon support (carbon black). The asreceived sample was dusted onto a standard 3 mm holey carbon grid that had been previously decontaminated by vacuumbaking. Experimental ADF images were recorded using a JEOL ARM200F operated at 200 kV equipped with both probeforming and image-side aberration correctors. The emission current was ≈15 μA immediately after flashing the tip (delivering a probe current of around 35 pA to the sample) but gradually decays. To precisely calibrate the data, the emission used for both detector mapping (12.8 μA in this case), as well as that for each image, should be recorded and the ratio between these incorporated in the analysis. The aperture and camera-length used produced a probe-convergence angle of 27.0 mrad and a potential collection range over the ADF detector spanning 69.2−279.1 mrad. However, unlike previous quantitative ADF studies,13,14,22 and to improve the accuracy of the data normalization, the flux distribution across the detector plane was also considered.23 The three inputs used in this enhanced analysis are shown schematically in Figure 1.

are then used to count the number of atoms in each column visible in the image. The unknown z positions of the atoms are then determined using a modified molecular dynamics relaxation. This fast, automated, approach allows for higher data throughput to either analyze many images of the same particle for dynamic studies or many particles for statistical surveys. Tomographic methods require not only significant beamdose for their imaging but also still further damaging beam-dose for microscope alignment, sample tilting, and field of view repositioning. Perhaps surprisingly then, even though tomographic techniques expose beam-sensitive samples such as nanoparticles to large electron doses,11 it is a requirement of their reconstruction that the sample must be assumed not to have changed during the series. To overcome the practical challenges of achieving atomic-resolution with supported nanoparticles, microscopists often utilize very large particles,8 embed fragile samples in a more robust matrix that may act as a heat-sink,7,9 or support them on a more easily controlled and aligned substrate.12 Unfortunately, each of these approaches results in a potentially incomplete or unrealistic representation of particle morphology. Catalyst nanoparticles for technological applications then present a “triple hit” of challenges; their small size and surface atom mobility render them beam-sensitive, the irregular nature of the sample support (often porous amorphous carbon-black) presents challenges of mechanical stability during imaging, and finally (as shown later), the contribution to image contrast from these irregular carbon supports is less simple to account for than with the free-standing,8 encapsulated,9 or oxide-supported cases.12 The principle of ADF STEM quantification, first proposed by Singhal et al. in the late 90s,13 is to express image contrast on an absolute scale through accurate detector calibration. Quantitative ADF STEM has had a resurgence recently with detectorsensitivity mapping and image-normalization now allowing experimental data to either be compared directly with STEM image simulation14,15 or decomposed using statistical techniques.9,16,17 However, comparing image peak intensities or contrast ratios can be very sensitive to several experimental unknowns including; defocus variations,18 astigmatism (or other residual aberrations),19 source size/coherence,20 and scan-noise.21 Alternatively, statistical methods can be used to decompose the image data into Gaussian components, with the appropriate number of components being determined by selecting a minima in an integrated classification likelihood (ICL) plot.9 These techniques do not rely on the absolute values of the atomic-column intensity data but rather on their relative distribution, and as a result, it can be less affected by small particle mist-tilt or by magnification miscalibration for example. However, care is required in computing and interpreting these plots to avoid an over- or underestimate in the number of components (and, hence, in the sample mass).12 Instead, in this work, we make use of an alternative approach, that of the integrated scattering cross section. This method, though requiring accurate ADF detector and magnification calibration, has been found to be robust to many of these harder to determine parameters.18,19 This cross-sectional approach allows for easier creation of simulated reference data as well as far simpler comparison with the experimental results. Here, we demonstrate our approach by observing the evolution of the shape of a single nanoparticle in vacuum

Figure 1. Illustration of the three input data used for the quantitative ADF analysis: (a) the experimental image, (b) the detector plane fluxdistribution, and (c) the measured detector efficiency.

Analysis of the flux distribution showed that while the ADF detector was sensitive to an outer angle of 279.1 mrad, it was only illuminated out to 137.5 mrad (see Supporting Information for detailed analysis). Along with giving the appropriate detector efficiency to normalize the experimental data, this also dictated the correct angles for the reference simulations, that is, 69.2−137.5 mrad. Following detector calibration, the experimental images were converted to an absolute scale where the image intensity is now B

dx.doi.org/10.1021/nl502762m | Nano Lett. XXXX, XXX, XXX−XXX

Nano Letters

Letter

Figure 2. Selected frames from the (a) start, (b) middle, and (c) end of the time-series data. Frame times and indicated emission currents are as shown. Line annotations indicate the terminating facets where visible. As the series progresses further, parts that indicate the amorphous-type sections of the perimeter are labeled “A”. Field of view is ≈10 nm.

expressed as a fraction of the incident beam current.14 To mitigate effects from the observed sensitivity asymmetries of ADF detectors,22,24,25 a “weighted-average detector-sensitivity” was used to normalize the images. The weighting merely reflects the flux distribution in the detection plane and is described in detail elsewhere.23 The image magnification was verified through calibrating the Fourier spots of the known crystal planes. Following automated peak-finding to locate the x−y coordinates of all atomic columns (method as in 21), a mask was generated that defines the outer bounds of the nanoparticle. The background intensity (from the carbon black) outside this mask was then extrapolated to estimate the contribution inside the mask and subtracted.26 This additional step is critical for supported samples, as failing to account for this additional scattering to the ADF detector leads to an overestimation of the particle thickness.12,26 The image is segmented into Voronoi cells20,27 before finally the integration to yield the scattering cross sections, with units of Mbarn (1 Mbarn = 10−22 m2 or 0.01 Å2), for all individual atomic columns.18 The “Absolute Integrator” software used in this work to perform this is available free of charge for academic/ noncommercial use from www.lewysjones.com. These Voroinoi cells prove a computationally efficient and robust means of segmenting the experimental image, as for thin samples (