11350 Langmuir 2008, 24, 11350-11360

InVited Feature Article

Particle Size Determination Using TEM: A Discussion of Image
Acquisition and Analysis for the Novice Microscopist
William D. Pyrz and Douglas J. Buttrey*
Center for Catalytic Science and Technology, Department of Chemical Engineering, UniVersity of
Delaware, Newark, Delaware 19716
ReceiVed May 1, 2008. ReVised Manuscript ReceiVed July 15, 2008

As nanoparticle synthesis capabilities advance, there is an increasing need for reliable nanoparticle size distribution
analysis. Transmission electron microscopy (TEM) can be used to directly image nanoparticles at scales approaching
a single atom. However, the advantage gained by being able to “see” these nanoparticles comes with several tradeoffs
that must be addressed and balanced. For effective nanoparticle characterization, the proper selection of imaging type
(bright vs dark field), magnification, and analysis method (manual vs automated) is critical. These decisions control
the measurement resolution, the contrast between the particle and background, the number of particles in each image,
the subsequent analysis efficiency, and the proper determination of the particle-background boundary and affect the
significance of electron beam damage to the sample. In this work, the relationship between the critical decisions
required for TEM analysis of small nanoparticles and the statistical effects of these factors on the resulting size
distribution is presented.

Introduction enhanced performance in catalytic,9,11,12 magnetic,20-22 or

electronic applications.23,24
The stabilization of small particles has become an area of
As synthetic capabilities advance, control over the size of
research that has gained significant momentum in recent years.1-8
small nanoparticles is approaching the level where nanoparticles
Various synthetic approaches have been used for the production
can be constructed from a predetermined number of atoms. As
of small particles; examples include incipient wetness impregna-
the number of atoms in these clusters steadily decreases, the
tion,9-11 dendrimer encapsulation,8,12,13 monolayer protection
challenge shifts toward the development of methods that
of clusters,14,15 as well as electrochemical,16 supercritical solvent,1
accurately and reproducibly characterize population character-
reverse micelle/microemulsion,4,7,17 and sol-gel tech- istics. Many methods currently used to measure particle size
niques.11,18,19 In many cases, the goal is to produce nanoparticles include (scanning) transmission electron microscopy (TEM/
with controlled size, good dispersion, and properties that show STEM),8,25-33 X-ray diffraction (XRD),31,34-36 extended X-ray
absorption fine structure (EXAFS),10,37 probe molecule absorp-
* Corresponding author. Phone: (302) 831-2034. Fax: (302)-831-1048. tion,10,31,38 magnetic measurements,32 optical absorption,33 and
E-mail: dbuttrey@udel.edu.
(1) Aymonier, C.; Loppinet-Serani, A.; Reverün, H.; Garrabos, Y.; Cansell,
F. J. Supercrit. Fluids 2006, 38, 242. (20) Mandal, M.; Kundu, S.; Sau, T. K.; Yusuf, S. M.; Pal, T. Chem. Mater.
(2) Li, Y.; Cai, W.; Duan, G. Chem. Mater. 2008, 20, 615. 2003, 15, 3710.
(3) Niederberger, M.; Garnweitner, G. Chem.sEur. J. 2006, 12, 7282. (21) Park, J. I.; Cheon, J. J. Am. Chem. Soc. 2001, 123, 5743.
(4) Eastoe, J.; Hollamby, M. J.; Hudson, L. AdV. Colloid Interface Sci. 2006, (22) Skomski, R. J. Phys.: Condens. Matter 2003, 15, R841.
5, 128–130. (23) Murray, C. B.; Kagan, C. R.; Bawendi, M. G. Science 1995, 270, 1335.
(5) Wegner, K.; Piseri, P.; Tafreshi, H. V.; Milani, P. J. Phys. D: Appl. Phys. (24) Murray, C. B.; Kagan, C. R.; Bawendi, M. G. Annu. ReV. Mater. Sci.
2006, 39, R439. 2000, 30, 545.
(6) O’Brien, P.; Pickett, N. L. Chem. Mater. 2001, 13, 3843. (25) Bele, P.; Jager, F.; Stimming, U. Microsc. Microanal. 2007, S5.
(7) Chen, C.; Wang, L.; Jiang, G.; Yu, H. ReV. AdV. Mater. Sci. 2006, 11, 1. (26) Fisker, R.; Carstensen, J. M.; Hansen, M. F.; Bodker, F.; Morup, S. J.
(8) Scott, R. W. J.; Wilson, O. M.; Crooks, R. M. J. Phys. Chem. B 2005, 109, Nanopart. Res. 2000, 2, 267.
692. (27) Howie, A.; Marks, L. D.; Pennycook, S. J. Ultramicroscopy 1982, 8, 163.
(9) Komiyama, M. Catal. ReV. Sci. Eng. 1985, 27, 341. (28) Karlsson, L. S.; Deppert, K.; Malm, J. O. J. Nanopart. Res. 2006, 8, 971.
(10) Oudenhuijzen, M. K.; Kooyman, P. J.; Tappel, B.; van Bokhoven, J. A.; (29) Malac, M.; Beleggia, M.; Egerton, R.; Zhu, Y. M. Ultramicroscopy 2007,
Koningsberger, D. C. J. Catal. 2002, 205, 135. 107, 40.
(11) Schwarz, J. A.; Contescu, C.; Contescu, A. Chem. ReV. 1995, 95, 477. (30) Pyrz, W. D.; Park, S.; Vogt, T.; Buttrey, D. J. J. Phys. Chem. C 2007,
(12) Niu, Y. H.; Crooks, R. M. C. R. Chim. 2003, 6, 1049. 111, 10824.
(13) Zhao, M. Q.; Crooks, R. M. AdV. Mater. 1999, 11, 217. (31) Weibel, A.; Bouchet, R.; Boulc’h, F.; Knauth, P. Chem. Mater. 2005, 17,
(14) Hostetler, M. J.; Zhong, C. J.; Yen, B. K. H.; Anderegg, J.; Gross, S. M.; 2378.
Evans, N. D.; Porter, M.; Murray, R. W. J. Am. Chem. Soc. 1998, 120, 9396. (32) Woodward, R. C.; Heeris, J.; Pierre, T. G. S.; Saunders, M.; Gilbert, E. P.;
(15) Templeton, A. C.; Wuelfing, W. P.; Murray, R. W. Acc. Chem. Res. 2000, Rutnakornpituk, M.; Zhang, Q.; Riffle, J. S. J. Appl. Crystallogr. 2007, 40, S495.
33, 27. (33) Duff, D. G.; Edwards, P. P.; Johnson, B. F. G. J. Phys. Chem. 1995, 99,
(16) Menard, L. D.; Gao, S. P.; Xu, H.; Twesten, R. D.; Harper, A. S.; Song, 15934.
Y.; Wang, G.; Douglas, A. D.; Yang, J. C.; Frenkel, A. I.; Nuzzo, R. G.; Murray, (34) Cullity, B. D.; Stock, S. R. Elements of X-Ray Diffraction, 3rd ed.; Prentice
R. W. J. Phys. Chem. B 2006, 110, 12874. Hall: Upper Saddle River, NJ, 2001.
(17) Shchukin, D. G.; Sukhorukov, G. B. AdV. Mater. 2004, 16, 671. (35) Hall, W. H. Proc. Phys. Soc., London, Sect. A 1949, 62, 741.
(18) Cushing, B. L.; Kolesnichenko, V. L.; O’Connor, C. J. Chem. ReV. 2004, (36) Williamson, G. K.; Hall, W. H. Acta Metall. 1953, 1, 22.
104, 3893. (37) Frenkel, A. Z. Kristallogr. 2007, 222, 605.
(19) Guimon, C.; Auroux, A.; Romero, E.; Monzon, A. Appl. Catal., A 2003, (38) Barrett, E. P.; Joyner, L. G.; Halenda, P. P. J. Am. Chem. Soc. 1951, 73,
251, 199. 373.

10.1021/la801367j CCC: $40.75  2008 American Chemical Society

Published on Web 08/26/2008
Particle Size Determination Using TEM Langmuir, Vol. 24, No. 20, 2008 11351

Figure 1. Series of typical phase-contrast TEM images (1024 pixels × 1024 pixels) of unsupported Au nanoparticles prepared using dendrimer
encapsulation methods at (a) 240K, (b) 300K, (c) 600K, and (d) 1000K nominal magnification using a JEOL 2010F TEM. The images shown were
obtained under conditions near the ideal Scherzer focus.
small-angle neutron scattering (SANS).32 Each technique has analysis, beam-induced artifacts, and a description of the future
advantages and disadvantages, but only TEM currently allows of TEM and STEM techniques. Several studies prior to this
the direct (real space) visualization of nanoparticles. discuss some of these topics,26,32,39 but few discuss TEM image
In this report, a discussion of the use of TEM as a primary analysis for nanoparticle systems with dimensions of less than
tool for the characterization of nanoparticle size distributions is 2 nm.10,25,40 For a more rigorous treatment of image acquisition
presented. The motivation behind this report is not to provide and particle size analysis for HAADF imaging, see the work by
a new advanced method for particle size analysis but to clearly Treacy et al. where they used HAADF imaging to quantify the
identify the strengths and weaknesses of TEM as a characterization number of atoms in small Pt clusters supported on Al2O3.41
tool. The discussion will address the details of image acquisition
and analysis and will provide a guide for the novice microscopist Experimental Section
to interpret particle size results. Furthermore, several of the
TEM analyses were performed on a JEOL 2010F equipped with
uncertainties associated with the measurement of particle size a Schottky field-emission gun operated at 200 keV with an ultra-
will be addressed to avoid the overinterpretation or improper use high-resolution pole piece providing a point resolution of 1.9 Å.
of the information provided in TEM micrographs. Specific topics
considered are magnification calibration, pixel resolution, (39) Reetz, M. T.; Maase, M.; Schilling, T.; Tesche, B. J. Phys. Chem. B 2000,
automated versus manual measurement, bright-field (BF)/phase- 104, 8779.
(40) Woehrle, G. H.; Hutchison, J. E.; Özkar, S.; Finke, R. G. Turk. J. Chem.
contrast imaging versus high-angle annular dark-field (HAADF) 2006, 30, 1.
imaging, the treatment of the uncertainties associated with image (41) Treacy, M. M.; Rice, S. B. J. Microsc. 1989, 156, 211.
11352 Langmuir, Vol. 24, No. 20, 2008 Pyrz and Buttrey

Micrographs 1024 pixels × 1024 pixels in size were acquired using accurately assessing the particle size distribution, as will be
a Gatan GIF-200 imaging filter with a model 794 multiscan-CCD discussed in subsequent sections.
camera. Annular dark-field imaging was performed using a JEOL Before the discussion of the third imaging method, it would
EM 24140 annular detector with a 0.5- or 0.7-nm-diameter nanoprobe be instructive to provide a few remarks concerning the basics
at camera lengths of 8-20 cm. Electron energy-loss spectroscopy
of imaging acquisition for both BF and phase-contrast images.
(EELS) was conducted in STEM mode using a 0.5 nm analytical
probe. The energy resolution of the zero loss peak (full width at One major goal of electron microscopy is to be able to acquire
half-maximum) was ∼1.5 eV at a pixel size of 0.3 eV. Spectra were images that are directly interpretable and provide new important
acquired using a dwell time of 2 s and were averaged over 8 frames. information about the materials under study. However, because
HR-STEM was completed on a JEOL 2100F equipped with a of unavoidable imperfections in the manufacture of electro-
CEOS Cs corrector on the illumination system. Geometrical magnetic lenses, most conventional TEMs suffer from a variety
aberrations were measured and controlled to provide less than a π/4 of aberrations that diminish the obtainable resolution.42 A few
phase shift of the incoming electron wave over the probe-defining of the major ones are spherical aberration, chromatic aberration,
aperture of 14.5 mrad. HAADF-STEM images were acquired on a and astigmatism. The most severe is spherical aberration (Cs),
Fischione model 3000 HAADF detector with a camera length such which causes image delocalization, or an inability to define the
that the inner cutoff angle of the detector was 65.6 mrad. The scanning specific location of a feature, where the image of a point is
acquisition was synchronized to the ac electrical power to minimize represented as a smeared disk.42 This is due to the inhomogeneous
60 Hz noise, and a pixel dwell time of 32 µs was chosen. forces that the lens (objective for conventional TEM and
Samples were prepared for TEM by one of two methods: (1) condenser for STEM) transfers to off-axis electron beams, and
dipping a lacey-carbon-coated copper grid (Electron Microscopy the severity of the smearing depends on both the magnitude of
Sciences) into the sample powder or (2) adding a drop of an aqueous
the spherical aberration coefficient (Cs) and the objective lens
nanoparticle suspension to a copper grid coated with an ultrathin
layer of carbon (Electron Microscopy Sciences). In some cases, strength.42 The smearing/delocalization limits the resolution of
sonication was used to disperse the nanoparticles further. the TEM and the ability to interpret an image properly. To
minimize the effects of delocalization, it is critical to image as
close to the Scherzer defocus as possible. The Scherzer defocus
Results and Discussion defined in eq 1 is the optimal objective lens condition for a given
Image Acquisition. For the analysis of small particles, the microscope and limits the effect of delocalization:42
most common methods for imaging in the TEM are bright-field
(BF), dark-field (DF), and phase-contrast imaging. At high ∆fSch ) -1.2(Csλ)1⁄ 2 (1)
magnifications, phase-contrast imaging is commonly referred to ∆fSch is the defocus value, Cs is the coefficient of spherical
as high-resolution electron microscopy (HREM). BF images are aberration, and λ is the wavelength of the incident electron beam.
generated using the direct electron beam, which consists of The proper focus setting is unique to each microscope as a result
forward scattered electrons, electrons that are scattered inelas- of the Cs term and varies with the accelerating voltage. An example
tically and follow a path parallel to the incident electron beam, of how to approximate the Scherzer condition visually compared
and electrons that do not pass through the specimen.42 BF images to both overfocused and underfocused imaging conditions using
are taken by inserting an objective aperture that is small enough fast Fourier transform (FFT) patterns (objective aperture removed)
to allow only the direct beam to pass through. In a BF image, is shown in Figure 2 for an amorphous carbon film. Equation 2
the contrast increases between the substrate and the sample as defines the maximum acceptance angle (Rmax) for the selection
(1) the sample thickness increases, (2) the density of a material of the optimal objective aperture to be used while operating the
increases at constant thickness, or (3) the material is crystalline TEM at Scherzer conditions:43
and is oriented such that Bragg diffraction is favored.42 In the
case of very small particles (<2 nm), all three scattering
mechanisms approach a minimum, and the contrast between the
substrate (typically an amorphous carbon film) and the nano-
Rmax ) ( )
6λ
Cs
1⁄ 2
(2)

particles is limited by the reduction in the number of scattering The image contrast implications of the Scherzer focus condition
centers for the nanoparticle relative to the number of scattering with respect to the contrast transfer function is beyond the scope
centers in the support. of this article, and additional information concerning this topic
and all aspects of electron microscopy can be found in refs 42-46.
Phase-contrast or HREM imaging is similar to BF; however,
A final issue that must be addressed is astigmatism within both
in this case a large-objective aperture or the removal of the
the objective and condenser lens systems. For objective lens
objective aperture is required such that both the direct beam and
astigmatism, the FFT pattern of an underfocused image of
Bragg-diffracted beams are used to generate the image.42 Because
amorphous material can be used to both identify and correct the
multiple beams are used in these images, variations in contrast
astigmatism as shown in Figure 3. The corrected case is shown
are observed when these beams are in phase with one another.42
in the left panel of Figure 3, and examples of severe objective
Increasing the size of the aperture allows more Bragg-diffracted
lens astigmatism are shown in the right panels at various defocus
beams to be included in the image. By incorporating more
values.
diffracted beams, the sharpness improves but the overall contrast
The third technique is dark-field imaging, wherein the scattered
decreases.42 A series of phase-contrast images typical of Au
electrons are restricted to large angles (>15 mrad).42,47 Dark-
nanoparticles supported on a carbon film taken at various
magnifications is shown in Figure 1. The observation of lattice (43) Kirkland, E. J. AdVanced Computing in Electron Microscopy; Plenum
fringes in Figure 1c,d confirms that diffracted beams were included Press: New York, 1998.
with the direct beam in the acquisition of these images. The (44) Reimer, L. Transmission Electron Microscopy: Physics of Image Formation
and Microanalysis; Springer-Verlag: Berlin, 1984.
limited contrast between the nanoparticles and background (45) Watt, I. M. The Principles and Practice of Electron Microscopy; Cambridge
(carbon film in Figure 1) is one of the principal difficulties in University Press: Cambridge, U.K., 1997.
(46) Spence, J. C. H. High-Resolution Electron Microscopy; Oxford University
Press: Oxford, U.K., 2003.
(42) Williams, D. B.; Carter, C. B. Transmission Electron Microscopy: A (47) Fitting, L.; Thiel, S.; Schmehl, A.; Mannhart, J.; Muller, D. A.
Textbook for Materials Science; Plenum Press: New York, 1996. Ultramicroscopy 2006, 106, 1053.
Particle Size Determination Using TEM Langmuir, Vol. 24, No. 20, 2008 11353

Figure 2. Series of bright-field images and their respective FFT patterns (shown under each image) showing an amorphous carbon film taken with
a variety of focus conditions. Example images of the carbon film are shown underfocused (top left), near the ideal Scherzer focus (-42.5 nm for
the JEOL 2010F), at minimum contrast (bottom left), and severely overfocused (bottom right). Note that the objective aperture was removed.

field images can be obtained by using an objective aperture to the probe convergence semiangle),47 to a first approximation,
select a specific Bragg-diffracted electron beam, or an annular the image contrast scales with the square of the atomic number
detector can be used to detect all of the diffracted electron beams (Z2).42,48 This technique is known as high-angle annular dark-
within a preselected angular range while excluding the direct field (HAADF) imaging or Z-contrast imaging. In the case of
electron beam and beams scattered through angles below the supported nanoparticles, this is a major advantage because a
preselected cutoff. As the atomic number (Z) increases, the majority of the particles are supported on relatively low Z oxides
probability of scattering increases strongly for a constant electron (e.g., Al2O3, TiO2, and SiO2) or various forms of carbon. An
dose. This provides a situation where contrast can be significantly example of the enhanced contrast provided by HAADF imaging
enhanced between two different elements if they are reasonably is shown in Figure 4a for Pt nanoparticles supported on carbon
well separated in the periodic table. For thin samples and imaging (Vulcan XC-72), as compared to a phase-contrast image of the
conditions that accept only high scattering angles (an angle such same sample at similar magnification shown in Figure 4b. For
that diffraction effects are negligible, corresponding to ∼3 times a constant sample thickness, the Pt/C scattering ratio increases
11354 Langmuir, Vol. 24, No. 20, 2008 Pyrz and Buttrey

Figure 3. Series of images of an amorphous carbon film and their respective FFT patterns (shown under each image) showing the effects of corrected
objective lens astigmatism (left image) and severe objective lens astigmatism (images to the right) as a function of focus. The image to the far right
is near the Scherzer focus conditions and can be compared with the image in Figure 2 (top panel, right image) of the same amorphous carbon film
at the Scherzer focus with the objective lens astigmatism corrected. Note that the objective aperture was removed.

Figure 4. (a) HAADF image (512 pixels × 512 pixels) of Pt nanoparticles supported on carbon (Vulcan XC-72) prepared by incipient wetness
impregnation. (b) BF image (1024 pixels × 1024 pixels) of the same sample showing the challenges associated with locating and measuring supported
nanoparticles as a result of the varying contrast of the support material.

from ∼30 using a Z4/3 dependence for BF imaging conditions lighter elements.42 Therefore, at higher accelerating voltages,
to ∼169 assuming a Z2 dependence for HAADF imaging.41 enhanced contrast can be obtained between platinum and a carbon
Following the selection of an imaging technique, the next support because the scattering from the carbon will be diminished
considerations are the choice of accelerating voltage and to a larger extent than for the platinum. However, as you continue
magnification. The choice of accelerating voltage provides an to increase the accelerating voltage, the probability of beam
additional parameter enhancing the contrast between the nano- damage also increases; most microscopes are operated at
particles and the support, particularly for heavy elements (Z >
intermediate accelerating voltages (80-300 keV). On most
30).42 As the accelerating voltage is increased, the probability
of scattering decreases, and the effect is more pronounced for microscopes, the magnification displayed is the nominal mag-
nification; it is imperative that all magnifications are properly
(48) Howie, A. J. Microsc. (Oxford, U.K.) 1979, 117, 11. calibrated for each operating mode using standards. Fortunately,
Particle Size Determination Using TEM Langmuir, Vol. 24, No. 20, 2008 11355

dosing the entire sample with the electron beam. The continuous
motion of the beam limits the total electron dose for a given
sample area, providing gentler imaging conditions for samples
exhibiting instabilities under the electron beam. This was recently
demonstrated using less intense beam conditions than in our
previous studies of MoVNbTeO where Te completely sublimed
from the structure during HREM studies57 but was readily
observed, though perhaps partially removed, using HAADF
STEM.58 Another advantage of STEM is that the placement of
the nanoprobe can be controlled, allowing for single-particle
spectroscopic analysis of composition or valance states using
energy-dispersive X-ray spectroscopy (EDS) or EELS, respec-
tively. However, special care must be taken during spectroscopic
studies on small particles in which the STEM probe is held
stationary during spectral acquisition. Long-time exposure by a
stationary STEM probe can lead to significant modifications in
particle size, composition, and/or morphology. An example is
Figure 5. Pixel size as a function of the nominal magnification for both provided in Figures 7 and 8, where beam damage results over
the BF and HAADF imaging modes for the JEOL 2010F. time during an EELS study of the Si L-edge from SiO2
nanoparticles. In Figure 8, the Si L-edge was sampled at six
most microscopy facilities have calibrated instruments, but different locations in a single particle. It is clear from the sequence
periodic verifications are recommended. The relationship between of images in Figure 7 that significant damage is caused by the
the magnification and pixel edge length is shown in Figure 5 for stationary STEM probe at each acquisition location. Furthermore,
the JEOL-2010F using both BF and HAADF operating modes the chemical nature of the nanoparticles was altered, as evidenced
with a 1024 pixel × 1024 pixel micrograph. For particle size by changes in the post-edge features indicating the reduction of
analysis, a natural inclination might be to select the highest SiO2 to amorphous Si. A second example of damage induced by
possible magnification; however, this will ultimately result in a stationary STEM probe is shown in Figure 9 for a 5 nm Pt-Re
insufficient sampling. nanoparticle. In this case, a stationary probe was used to perform
Another problem with imaging at very high magnification is five consecutive EDS analyses. After each acquisition, an image
the enhanced likelihood of particle-selection bias by the was obtained, and there is a clear change in both the size and
microscopist. With fewer particles in the field of view, attention morphology of the nanoparticle. A report by Egerton et al.
could be preferentially drawn to one particle type over another provides a detailed discussion outlining the different mechanisms
(i.e., selecting particles exhibiting lattice fringes or focusing only contributing to beam damage in thin specimens.56 For samples
on the smallest nanoparticles at the expense of larger ones). It that are very sensitive to the beam and require low-dose imaging
is also possible that an entire population of particles (varying in conditions, image plate (IP) scanners provide an alternative data
size, composition, or morphology) is inadvertently excluded collection technique. The IP scanner has a large linear dynamic
because of the small sampling sizes, the latter being a risk with range and is capable of detecting single electrons, but its use is
all TEM studies. By imaging at both low and high magnifications somewhat limited because of degraded resolution under high
or by using SEM and/or light microscopy, more particles will electron doses.59,60
be present, which should decrease the likelihood of bias and/or In general, the imaging conditions must be chosen such that
unsampled particle populations. a compromise is reached that provides a sufficient number of
Another drawback with decreasing particle size, particularly particles in each image, but they should be selected with a
for metals, is that nanoparticles tend to become less stable under resolution providing an acceptable number of image pixels such
the electron beam. Instabilities may arise from particle motion, that the uncertainties associated with the measurements are
structural fluctuations, atom sputtering, coalescence, and particle minimized. Additional considerations associated with particle
destruction/decomposition.30,49-56 The sequence of phase- stability and electron beam damage are necessary to ensure reliable
contrast images in Figure 6 shows an instability arising during results for particle size analysis and/or particle composition.
a study in which enough energy was provided for two particles Image Measurement, Analysis, and Interpretation. Once
to mobilize, meet, and then coalesce into a single larger particle. an imaging strategy has been chosen and images have been
To test for sample stability, a sequence of images can be acquired acquired, the methods for image processing and particle size
and compared for physical, structural, and positional changes measurement must be selected. Before discussing the various
over time. The high electron doses typically associated with types of image processing techniques and available algorithms,
high-resolution imaging exacerbate damage and/or restructuring. we begin by considering the discrepancy between the measured
STEM imaging provides an alternative to the high-dose electron and true size for a single particle as represented in eq 3
beam required for high-resolution BF imaging in TEM mode.
xi ) x + ε (3)
The advantage of STEM is that a rastering nanoprobe (∼0.1-1.0
nm diameter) is used to analyze samples rather than continuously where xi is the measured size, x is the true (but unknown) size,

(49) Ajayan, P. M.; Marks, L. D. Phys. ReV. Lett. 1988, 60, 585. (57) DeSanto, P.; Buttrey, D. J.; Grasselli, R. K.; Lugmair, C. G.; Volpe, A. F.;
(50) Ajayan, P. M.; Marks, L. D. Phys. ReV. Lett. 1989, 63, 279. Toby, B. H.; Vogt, T. Z. Kristallogr. 2004, 219, 152.
(51) Ajayan, P. M.; Marks, L. D. Phase Trans. 1990, 229, 24–6. (58) Pyrz, W. D.; Blom, D. A.; Vogt, T.; Buttrey, D. J. Angew. Chem., Int.
(52) Buffat, P.; Borel, J. P. Phys. ReV. A 1976, 13, 2287. Ed. 2008, 47, 2788.
(53) Lu, B.; Wang, C. S.; Zhang, Y. H. Appl. Phys. Lett. 1997, 70, 717. (59) Bele, P.; Ochs, R.; Angert, I.; Schröder, R. R. Microsc. Res. Tech. 2000,
(54) Ru, Q. Appl. Phys. Lett. 1997, 71, 1792. 49, 281.
(55) Williams, P. Appl. Phys. Lett. 1987, 50, 1760. (60) Gregory, C. D.; Howe, J.; Courtois, C.; Chambers, D. M.; Woolsey,
(56) Egerton, R. F.; Wang, F.; Crozier, P. A. Microsc. Microanal. 2006, 12, N. C.; Förster, E.; Uschmann, I.; Renner, O. Central Laser Facility Annual Report,
65. 2003/2004.
11356 Langmuir, Vol. 24, No. 20, 2008 Pyrz and Buttrey

Figure 6. Sequence of phase-contrast images that captures Au nanoparticle mobility and coalescence induced by the energy provided from the incident
beam. The Au nanoparticles were prepared using dendrimer encapsulation methods. The total time that the sample was exposed to the electron beam
from the acquisition of the left image to the acquisition of the right image was ∼1 min.

Figure 8. EELS spectra of the Si L-edge for six different locations from
a SiO2 nanoparticle (Snowtex ZL Colloidal Silica, Nissan Chemical) in
Figure 7. The numbers represent the locations in Figure 7, and the top
spectrum is from a reference Si powder. The asterisk represents spectra
that were smoothed using a seven-point moving average.

conditions, and the measured results can represent more than

just the intrinsic variability in particle size. Two significant factors
that contribute to uncertainty in both the mean and the standard
Figure 7. HAADF image of a SiO2 particle (Snowtex ZL Colloidal deviation are the pixel resolution and the bias associated with
Silica, Nissan Chemical) (a) before EELS analysis and (b) after collecting
spectra from the first three locations. (c) HAADF image of the left edge the determination of the particle edge at coarse magnification.
of the SiO2 particle in panel a showing acquisition locations 4-6. (d) The pixel size as a function of magnification for a 1024 pixel
HAADF image of the SiO2 particles after all six spectra acquisitions. × 1024 pixel image is shown in Figure 5, and the fractional
contribution to the nanoparticle size from one added pixel as a
function of magnification for conventional TEM is shown in
and ε is the associated error. The contributions to ε include the Figure 10. Consider the measurements of two populations, one
pixel resolution, the determination of the particle-substrate with a mean size of 5.00 nm and the other with a mean of 1.00
boundary, the magnification calibration, and image distortions nm, each taken in BF with a nominal magnification of 200K. It
due to stage drift, improper focus conditions, delocalization due is the authors’ opinion that careful measurement of a particle
to spherical aberration, and various sources of astigmatism diameter, whether manually determined or from an automated
(objective lens, condenser lens, coma, etc). Contributions to counting routine, should be accurate to within 2 pixels, one for
systematic errors such as the magnification error and lens each edge. For the JEOL-2010F with a 1024 pixel × 1024 pixel
distortions shift the mean but, with a few exceptions such as image at 200K magnification, the image pixel length is 0.12 nm.
barrel distortions,42 generally do not influence the shape of the Therefore, the measured diameter of a 5.00 nm particle would
distribution. We will assume that measurements are made on a be composed of ∼42 pixels, and the diameter of a 1.00 nm
TEM that has been carefully calibrated and properly focused and particle would be ∼9 pixels. As depicted in Figure 10, extraneous
that all lens distortions and stage drift have been minimized. pixels added or subtracted from the particle boundary due to
The most common way to express the size distribution is discrepancies associated with the determination of the particle-
typically through a Gaussian or log-normal function, characterized background boundary will have a minimal effect on the
by the mean (µ) and the standard deviation (σx). For small particles measurement of the 5.00 nm particle; each pixel represents ∼2.4%
(∼2 nm or less), the value of the mean and standard deviation of the total diameter. However, for the 1.00 nm particle, the
are very sensitive to the choices in both imaging and analysis contribution of each pixel is significant, representing ∼12% of
Particle Size Determination Using TEM Langmuir, Vol. 24, No. 20, 2008 11357

Figure 9. HAADF image of a 5 nm nanoparticle from a Pt-Re catalyst Figure 11. (a) Bright-field image showing larger Bi-Ni nanoparticles
prepared using incipient wetness impregnation that was analyzed using generated from electron-beam-induced fragmentation of micrometer-
a stationary STEM probe for elemental analysis. The images show a sized, hydrothermally prepared Ni-Bi alloy particles exhibiting very
single particle from (a) the original image, (b) after a 100 live second good contrast variation compared to the background. (b) BF image
EDS acquisition, (c) after 10 min of exposure to the rastering STEM following thresholding using the intensity histogram of the raw image.
beam, (d) following the second 100 live second EDS acquisition, (e) (c) HAADF images showing good contrast between the particles and
after 5 min of exposure to the rastering beam, (f) after the third 100 live the support. (d) HAADF image following thresholding using only the
second acquisition, (g) after the fourth 100 live second acquisition, and intensity histogram of the raw image. Note that the images showing the
(h) after the fifth 100 live second acquisition. threshold mask were captured prior to the use of particle-selection
algorithms.

Image processing approaches can be classified as manual,

software-assisted, or fully automated. Manual analyses can be
accomplished with direct measurements on a calibrated micro-
graph, such as by using a negative with a ruler and a magnifying
glass; it may also be software-assisted, using image analysis
packages such as Digital Micrograph or ImageJ,40 which offer
an array of simple line-measurement tools. The advantage of the
manual and assisted methods is that each particle is scrutinized
and measured, the signal-to-noise requirement is less stringent,
and analysis is straightforward. This is particularly useful for
supported nanoparticles because the support often varies in
thickness and leads to a nonuniform background. It is critical,
however, that a consistent measurement method is employed
Figure 10. For nanoparticles of various sizes, the fraction of the (i.e., measuring horizontal or vertical diameters, shortest diameter,
nanoparticle diameter that is represented by a single pixel within a 1024
× 1024 BF image as a function of nominal magnification. longest diameter, etc.) and that the technique is properly disclosed
when reporting results. Significant drawbacks to manual mea-
surements include user biases, the tedious nature of repeated
the total diameter. In the absence of careful measurement when measurements, and limited sampling statistics because only a
trying to measure images with a 0.12 nm pixel resolution for a few hundred particles can be measured within a reasonable time
1.00 nm nanoparticle population, an uncertainty of a single pixel period. Manual measurement is particularly burdensome if
at each end of a 1.00 nm particle may therefore be ∼24% of the comparison between several samples is required.
true diameter. Clearly, the use of a coarse pixel size can lead to
serious interpretation errors when comparing or trying to Alternatively, for analyses requiring improved counting
differentiate between two nanoparticle populations that are statistics, automated software packages such as Digital Micro-
expected to be reasonably close in size (i.e., 1.50 nm versus 1.75 graph and ImageJ40 offer a variety of thresholding algorithms
nm). Furthermore, changing the data analyst responsible for designed to locate and measure particles. In addition to the particle
measuring the size of the particles could drastically alter the final diameter, many other measures of size (e.g., perimeter and area),
conclusions of a particle size study if biases on boundary selection morphology (e.g., roundness, minimum and maximum radii),
differ. Clearly, an important objective should always be to work and size distribution descriptors (e.g., mean, median, and mode)
with image pixel sizes that are sufficiently small so as not to can be calculated for a particle population. The complexity of
compromise the size assessment and to perform measurements the algorithm is quite variable, from simple thresholding based
with criteria that are independent of the individual analyst. on the unprocessed image intensity histogram to more sophis-
11358 Langmuir, Vol. 24, No. 20, 2008 Pyrz and Buttrey

Figure 12. Bright-field image of small Au dendrimer-encapsulated nanoparticles that do not exhibit good signal-to-noise variation compared to the
background. (b) Simple thresholding algorithms fail for the measurement of small nanoparticles (<5nm) in BF images. In this case, attempts to
threshold the nanoparticles fully led to the inclusion of several extraneous background pixels (which is clearly evident in regions near the scale bar).
Note that the image showing the threshold mask was captured prior to the use of particle-selection algorithms.

to that of the nearest neighbors and rescales the pixel on the basis
of a normalization factor calculated from the predetermined
scaling coefficients included in the matrix.61 As an example,
consider a 3 × 3 median filter that rescales the pixel intensity
by comparison with the nearest horizontal, vertical, and diagonal
neighbors and then replaces the central pixel intensity with the
median value of surrounding pixels.61 This operation is applied
to every pixel in the image with the objective of removing stray
pixels with unusually high or low intensity values, at the expense
of slightly decreasing sharpness. Reference 61 provides examples
and explanations for several of the common kernels and filters
in which some examples are maximum, minimum, median, mode,
Laplacian, Sobel, midpoint, sharpen, and smooth. These matrices
are available in a variety of sizes and shapes, and most of them
allow the user to modify the existing coefficients to create custom
smoothing algorithms.
Once appropriate smoothing and sharpening processes have
been applied, thresholding algorithms are used to select pixels
that have intensities that are representative of the desired
nanoparticles. The most basic thresholding technique allows the
selection of an upper bound and a lower bound of intensity values
that define a feature of interest. All pixels that have intensity
Figure 13. High-resolution HAADF-STEM image showing Ni-Bi values within these bounds are selected. Defects or stray pixels
nanoparticles generated from the electron-beam-induced fragmentation can be corrected by additional processing techniques that can fill
of micrometer-sized Ni-Bi alloy particles prepared hydrothermally that
are atomically resolved and surrounded by individual atoms, dimers,
holes, remove extraneous background pixels with erosion tools,
and small atom clusters. The image was taken on an aberration-corrected or add additional pixels to nanoparticle edges using dilation tools.
JEOL 2100F TEM. Full descriptions of these processes and tools are found in ref
61. After thresholding, a counting algorithm is employed, the
ticated multistep routines.25,26,39,61 The multistep routines typi- objects are counted, and the particle size distributions are
cally employ several preprocessing steps to smooth uneven generated, along with any other desired size or morphology
background, enhance contrast between the particle and the evaluation.
substrate through a variety of kernels and filters, and finally to Among the dangers associated with significant processing are
locate and count particles.25,61 A typical background smoothing loss of information and addition of artifacts that introduce errors
routine applies a matrix kernel that compares the pixel intensity into the analysis. Despite being automated, these routines also
(61) Russ, J. C. The Image Processing Handbook, 3rd ed.; CRC Press: Boca suffer from user bias because the user or algorithm author sets
Raton, FL, 1999. the criteria for discriminating the particle edge from the
Particle Size Determination Using TEM Langmuir, Vol. 24, No. 20, 2008 11359

background. Accurate particle edge determination requires been recent reports indicating that this has been further improved
significant pixel intensity differences at the particle-substrate to ∼0.5 Å in 300 kV instruments.72 Furthermore, the correction
boundary, as shown in Figure 11 for representative BF and of spherical aberration for both STEM and HREM significantly
HAADF images of reasonably large nanoparticles on a carbon minimizes the delocalization issues that plague uncorrected
film before and after a simple pixel-range thresholding routine instruments. Figure 13 shows an image of Bi-Ni nanoparticles
was applied. For both the BF and HAADF image, the automated taken with an aberration-corrected STEM in HAADF mode using
thresholding routine provides a mask that is in good agreement a JEOL 2100F. The image shows both atomically resolved
with the original images. nanoparticles and an array of single atoms, dimers, and small
For the smallest nanoparticles (<5 nm) supported on a thin clusters consisting of only a few atoms. The atomically dispersed
carbon film, simple thresholding routines that use only the particles may go unobserved with instruments lacking aberration
intensity histogram are inadequate, as demonstrated in Figure correction. This new generation of corrected instruments will
12. A close examination of Figure 12b reveals the inaccuracies permit the determination of precise cluster sizes and distributions
resulting from exclusive use of the intensity histogram for the and may allow a more reliable estimation of the number of atoms
differentiation of particles from the support because a large present. More recently, aberration-corrected TEM imaging and
number of extraneous pixels are included in the thresholded image reconstruction of a through-focal series of images have
mask. It is also clear that the selected nanoparticles were not allowed the extraction of 3-D information. In this technique,
accurately masked because several particles have voids and incremental images of an object are taken using known focus
regions of missing pixels, which is partially due to the speckled steps that represent different “slices” of the object in the vertical
contrast of the underlying amorphous carbon-film support. The direction. Stacking these images then provides a depth profile
analysis of supported nanoparticles is further complicated by the of the object, and comparison of images within the stack allows
variable thickness of the support, which leads to significant the estimation of objects in the direction parallel to the electron
variations in image contrast, as seen in Figure 4a. beam. This technique was used identify single Hf atoms within
Improper application of these counting routines to noisy images the bulk of the SiO2 layer of a HfO-SiO2-Si interface.73 A
or images with nonuniform backgrounds often leads to significant second technique that is being used to gather information in the
over- or underestimation of particle size as a result of adding or third dimension is tomography. In tomography, an object is
removing questionable pixels at the particle edge or inadvertent imaged at incremental tilt angles about a single axis (up to 70°
inclusionofsupportparticlesinthedistribution.Theparticle-substrate in each direction), and the stack of images is similarly
boundary, determined through thresholding, can be verified reconstructed into a cube of information that provides a 3-D
manually for each particle, but this becomes time-consuming view of the object of interest.74-76 Tomography can be used for
and the automated process no longer saves time relative to manual a wide variety of materials, particularly in biological samples
measurement. However, in a recent report by Bele et al., a self- and also in the study of catalysis and porous materials such as
written algorithm using the local adaptive threshold technique zeolites.74-76 Aberration correction, through-focus image re-
is used to analyze an image by segmenting it into a group of construction, and tomography will become powerful techniques
small subimages that are individually processed.25 The advantage for the analysis of controlled nanoparticles with the goal of
of this algorithm is that an image with a background having producing tailored clusters with a specific number of atoms.
variable contrast (i.e., supported nanoparticles) can be analyzed
because the edge-selection criteria are based on pixels local to Conclusions
the particle of interest, rather than using a global signal-to-noise As nanoparticle sizes steadily decrease toward clusters of only
ratio that is applied to all pixels within the entire image.25 They a small number of atoms, the challenges associated with the
have demonstrated effective particle size counting for Pt implementation of quantitatively reliable characterization tech-
nanoparticles supported on carbon with a mean diameter of 1.5 niques increase. For small nanoparticles (∼2 nm or less), TEM
nm.25 Drawbacks to these new advanced algorithms are the provides a unique tool that allows direct visualization down to
increased computing time necessary to process each image and a single atom. Proper quantitative analysis of particle size requires
the lack of commercially available packages. a compromise among the selections of imaging type (TEM vs
Future of TEM Particle Size Analysis. The future of STEM), magnification, and analysis method (manual vs auto-
nanoparticle size characterization with TEM is quite promising. mated) with the goals of optimizing the measurement resolution,
Aberration-corrected microscopes are becoming more common the contrast between the particle and background, the number
outside of the national laboratories, and the applications of these of particles in each image, the analysis efficiency, the proper
instruments are appearing with increasing frequency.29,58,62-71 determination of the particle-background boundary, and limiting
For both TEM and STEM operation at 200 kV, the aberration- the electron beam damage. Images obtained at coarse magnifica-
corrected point resolution is ∼1.0 Å in dark field. There have tions may clearly depict a nanoparticle population but severely
limit reliable quantification of the particle size distribution.
(62) Batson, P. E.; Dellby, N.; Krivanek, O. L. Nature 2002, 418, 617. However, imaging at very high magnification often leads to
(63) Chang, L. Y.; Kirkland, A. I.; Titchmarsh, J. M. Ultramicroscopy 2006, sampling an inadequate number of particles, and the concomitant
106, 301.
(64) Dellby, N.; Krivanek, O. L.; Nellist, P. D.; Batson, P. E.; Lupini, A. R. increase in electron flux may contribute to instabilities and beam
J. Electron Microsc. 2001, 50, 177. damage. These limitations can easily lead to errors or bias, which
(65) Haider, M.; Rose, H.; Uhlemann, S.; Kabius, B.; Urban, K. J. Electron
Microsc. 1998, 47, 395. can be up to 30-50% of the average size for a population of
(66) Haider, M.; Rose, H.; Uhlemann, S.; Schwan, E.; Kabius, B.; Urban, K.
Ultramicroscopy 1998, 75, 53. (72) Dahmen, U. Microsc. Microanal. 2007 2007, 23.
(67) Hirata, A.; Hirotsu, Y.; Nieh, T. G.; Ohkubo, T.; Tanaka, N. Ultrami- (73) van Benthem, K.; Lupini, A. R.; Oxley, M. P.; Findlay, S. D.; Allen, L. J.;
croscopy 2007, 107, 116. Pennycook, S. J. Ultramicroscopy 2006, 106, 1062.
(68) Liu, Z.; Suenaga, K.; Iijima, S. J. Am. Chem. Soc. 2007, 129, 6666. (74) Baumeister, W.; Grimm, R.; Walz, J. Trends Cell Biol. 1999, 9, 81.
(69) Rose, H. Optik 1990, 85, 19. (75) Baumeister, W. Curr. Opin. Struct. Biol. 2002, 12, 679.
(70) Voyles, P. M.; Muller, D. A.; Grazul, J. L.; Citrin, P. H.; Gossmann, (76) Ziese, U.; Koster, A. J. Appl. Catal., A 2004, 260, 71.
H. J. L. Nature 2002, 416, 826. (77) Cowley, J. M. Diffraction Physics; North Holland: New York, 1981.
(71) Yoshida, K.; Kawai, T.; Nambara, T.; Tanemura, S.; Saitoh, K.; Tanaka, (78) Egerton, R. F. Electron Energy-Loss Spectroscopy in the Electron
N. Nanotechnology 2006, 17, 3944. Microscope; Plenum Press: New York, 1996.
11360 Langmuir, Vol. 24, No. 20, 2008 Pyrz and Buttrey

1.00-1.50 nm particles. The size distribution would then have Acknowledgment. The authors thank C. Ni with the Keck
little statistical significance because the results would not Microscopy facility for TEM access and assistance and
appropriately capture the details of the particle population. The acknowledge support from the U.S. Department of Energy
result would be a convolution of user bias with the real particle (grant no. DE-FG02-03ER15468). The authors gratefully
size variations that is not easily differentiated. After a careful acknowledge T. Vogt and D. A. Blom at the University of
consideration of all variables, the user can be confident that South Carolina for access to and assistance with the aberration-
properly measured particle size distributions have statistical corrected JEOL 2100F, S. Park and T. Vogt for the preparation
significance and accurately reflect the population. References of the Bi-Ni samples, D. A. Simonetti and J. A. Dumesic for
42-46 and 77-80 are provided for the reader interested in
providing the Pt-Re catalysts samples, M. Weir for Au
additional information on diffraction, EELS, EDS, or HREM
nanoparticle samples, and R. M. Crooks, Sir J. M. Thomas,
imaging.
D. A. Blom, and B. A. Ogunnaike for insightful discus-
(79) Thomas, J. M.; Williams, B. G.; Sparrow, T. G. Acc. Chem. Res. 1985, sions.
18, 324.
(80) Thomas, J. M.; Midgley, P. A. Chem. Commun. 2004, 11, 1253. LA801367J