Editors' Choice—Connecting Fuel Cell Catalyst Nanostructure and Accessibility Using Quantitative Cryo-STEM Tomography

Pt nanoparticles deposited on two types of carbon black supports were used in this study. Vulcan XC-72 is a relatively compact carbon black with a Brunauer-Emmett-Teller (BET) surface area of ∼250 m²/g. Ketjenblack is a high-surface-area carbon (HSC) with BET surface area of ∼800 m²/g. Pt was deposited on each carbon support using a wet impregnation method at two concentrations: 10 and 50 weight %. The mean Pt crystallite diameters were about 2.5–3 nm determined by XRD using the Scherrer equation at the Pt (220) peak.

The electrochemical surface area (ECSA) of the catalyst powders was measured by hydrogen adsorption (HAD) at room temperature. Thin film catalyst layers were deposited on glassy carbon disk electrode (0.196 cm² area), mounted on an interchangeable rotating disk electrode (RDE) holders (Pine Instruments, USA). A Pt-foil counter electrode and saturated calomel (SCE) reference electrode were used. Detailed procedures on thin film applications and measurements can be found in Reference 2. The Pt ECSA was measured by integrating hydrogen adsorption peaks (assuming 210 μC/cm²-Pt) in cyclic voltammograms in 0.1 M HClO₄ at a 20 mV/s sweep rate.

CO-stripping measurements were performed on membrane electrode assemblies (MEAs) that were prepared using the decal-transfer method to make catalyst-coated membranes, as described elsewhere.² MEAs had an active area of 50 cm². The Pt loadings in the cathode were 0.05 and 0.40 mg_Pt/cm² for 10 and 50 wt% catalysts, respectively. Anode Pt loading was 0.05 mg_Pt/cm². Nafion D2020 ionomer, with an equivalent weight of 950 g/equiv was utilized with an ionomer to carbon weight ratio of 0.95 and 0.6 in the cathode and anode. Reinforced 18 μm thick perfluorosulfonic acid (PFSA) membranes were used. The gas diffusion layers were made from 200 μm thick Teflon-coated carbon fiber diffusion media with 30-μm microporous layers.

ECA measurement by CO stripping in MEAs is similar to that in a conventional liquid electrochemical cell, as previously described.³⁰ After letting the MEA equilibrate at the measurement relative humidity (RH) at 80°C for 30 min, several voltage cycles between 0.05–0.9 V were conducted to clean the catalyst surface. This was followed by 2 minutes of CO purge while holding the electrode at 0.05 V. The cathode channel was then flushed with N₂ for 1 minute, and a cyclic voltammogram starting from 0.05 V to 0.9 V was swept to quantify the CO oxidation charge. ECSA from the CO stripping was calculated assuming 420 μC/cm²-Pt. The procedure was repeated at different RH from 10% to 100% RH. The utilization is defined as the fraction of the CO-stripping ECSA over the ECSA at 100%RH. The low-RH utilization shown in Figure 5c is calculated as the average of the utilization at 10% RH and 20% RH. The error bars in Figure 5c span from the utilization at 10% RH to that at 20% RH.

Zoom In Zoom Out Reset image size

Catalyst powders were dispersed in methanol and dropped onto 100-mesh, hexagonal, copper TEM grids coated with carbon films of 3–4 nm nominal thickness and allowed to dry in air at room temperature. For the Pt-free HSC specimen, 5 nm diameter spherical gold nanoparticles were deposited on the grid prior to the specimen to act as fiducial markers for tomography. To eliminate the possibility of etching the carbon supports, the samples were not plasma cleaned.

Transmission electron microscope (TEM) and Scanning transmission electron microscope (STEM) images were acquired in a FEI Tecnai F20 with a Schottky field emission gun and a 200 kV accelerating voltage. No objective aperture was used for TEM imaging. For STEM imaging, a convergence angle of ∼7–10 mrad was used and the low angle annular dark field (LAADF) signal was collected to enhance the signal for carbon.

Electron tomography experiments were performed using STEM imaging as described above with convergence angle of ∼6.9 mrad to provide sub-nanometer resolution over a large depth of field. For automatic segmentation to be feasible, the imaging mode used for tomography must provide strong contrast for weakly-scattering carbon, so that carbon can be reliably distinguished from Pt and void. To meet this requirement, the low angle annular dark field signal was acquired with a camera length of 490 mm, providing a strong overall signal and a smaller intensity difference between carbon and Pt relative to high angle annular dark field (HAADF). An alternative imaging mode that provides a strong signal for carbon is bright-field (BF) transmission electron microscopy. However, phase contrast from amorphous carbon introduces speckle noise in BF TEM imaging, as illustrated in Figure S1, which interferes with interpretation and analysis of the tomogram.

To simultaneously suppress carbon contamination and ice accumulation during data acquisition, samples were maintained at −100C in a Gatan model 914 cryo-tomography holder using liquid nitrogen and a Gatan model 900 cold stage controller. Samples were imaged over a typical tilt range of around ± 75° at 2° increments with a 0.36 nm pixel size and 1024 × 1024 pixels per image. To reduce distortions in STEM images introduced by thermal drift of the sample, at each tilt we acquired a series of fast-scan images typically with ∼1 second per frame and 24–32 seconds total acquisition. After data acquisition each image series was aligned by cross-correlation and summed to produce a high signal-to-noise ratio image at each tilt, averaging out scan noise and limiting drift-induced distortions to around 1–2 pixels. This permits a good reconstruction quality over the full ∼350 nm field of view.

Tomography data preprocessing, reconstruction, segmentation, and analysis were performed in Matlab, using custom code and functions from the Image Processing Toolbox and the Statistics and Machine Learning Toolbox, unless otherwise stated.

To prepare the summed-image tilt series for tomographic reconstruction, hot pixels due to X-ray noise were removed, the average background intensity due to the uniform carbon support film is subtracted from each image, and the images are normalized to maintain a constant average intensity. Images in the tilt series are aligned to a common coordinate system manually using Pt nanoparticles (or the added gold particles in the Pt-free dataset) as fiducial markers, then shifted and rotated to center the axis of rotation. Some datasets also have ∼1% scan non-orthogonality corrected using an image shear operation. Shear and rotation operations used linear interpolation. Images were then binned by two, resulting in a 0.71 nm pixels/voxel edge length.

For the Pt-free HSC, the tomogram was reconstructed using the weighted back-projection (WBP)⁴⁰ algorithm, implemented using the Matlab iradon function. A 1.5 pixel radius Gaussian blur was subsequently applied to reduce noise and finite sampling artifacts. For carbon-supported Pt samples, the presence of Pt makes accurate reconstruction of the carbon more difficult. Because Pt is approximately 30 times brighter than an equal volume of carbon in LAADF STEM, minor artifacts in the Pt reconstruction fall in the intensity range of carbon. These artifacts appear as bright streaks or shadows and interfere significantly in the segmentation of carbon and voids. Tomograms were reconstructed using the simultaneous iterative reconstruction technique (SIRT)⁴¹ with 20 iterations, which provided sufficient artifact suppression to allow recognition of the exterior carbon surface. A discussion of the choice of reconstruction algorithm is presented in the supplemental information (Figure S2).

The carbon-supported Pt tomograms were segmented automatically using a combination of threshold and morphological filtering operations, which provide additional suppression of artifacts. The segmentation procedure is described in detail in the supplemental information and illustrated in Figure S3. Uncertainty in the segmentation procedure was estimated by repeating the Pt segmentation procedure and analysis for reasonable high, low, and medium platinum thresholds.

For the Pt-free HSC specimen, carbon was identified using simple thresholding of the smoothed reconstruction without any morphological filtering. Because the choice of threshold impacts the measured size of pores and pore openings, the analysis was repeated for reasonable low, medium, and high carbon thresholds to estimate uncertainty.

Once the reconstruction was segmented, individual Pt nanoparticles were identified and Pt particle effective diameters were calculated from particle volumes assuming a spherical morphology. Surface areas were calculated for individual Pt particles were calculated using discrete Crofton formula code implemented by Legland, et al.⁴² for larger particles (≳3 nm diameter) and by a spherical-particle approximation for smaller particles (≲2 nm diameter). Pt particles touching the external void were identified as exterior to the carbon support, and those that only touched the filled carbon were identified as interior to the carbon support. This automatic delineation of interior and exterior Pt agreed with visual inspection more than 95% of the time.

Pt particle size distributions for interior and exterior particles are shown in Figure S4, and surface-area weighted size distributions are shown in Figure S5. The experimentally measured size distributions were fit to lognormal distributions truncated below 1 nm to account for our limited capability for detecting very small nanoparticles in tomography. Specific surface area from tomography was calculated using the total Pt surface area and volume, assuming the density of bulk Pt. All Pt surface area was included, including surfaces that contact the carbon support – no correction for the "degree of embeddedness"⁴³ was used. The exterior specific surface area was calculated using the surface area of exterior particles only and the volume of all particles, consistent with an assumption of interior particles being present but inaccessible. Further details of the analysis are presented in the supplemental information.

3D electron tomography visualizations were created using the open source tomography platform tomviz (http://www.tomviz.org/). Visualizations of segmented reconstructions were created using binary volumes for each material component exported from Matlab created with medium thresholds. Surfaces are contour renderings at the value 0.5 for each binary component.

Each of the carbon support materials studied here contains some variation in morphology and atomic structure. Carbon aggregates tend to contain clusters of similar primary particles with similar structure extending over hundreds of nanometers.

In Vulcan, primary particles have a broad distribution of sizes which is roughly bimodal,⁴⁴ including smaller 10–20 nm diameter primary particles and large 25–40 nm diameter primary particles. Representative high resolution TEM (HRTEM) images of Vulcan primary particles in each size range (Figure S6) reveal that they have distinct atomic structures as well. The larger primary particles have thick graphitic shells around a relatively amorphous core, while the smaller primary particles are less dense, more amorphous, and do not show dense graphitic shells. Vulcan also includes a few very large 50–90 nm diameter primary particles, although these appear to accumulate very little Pt (Figure S7), and are thus neglected in this investigation.

Primary particles in HSC do not segregate by size as dramatically as in Vulcan, although they show clear differences in morphology. HSC includes two distinct primary particle types, one which appears to be relatively dense and solid, and another that appears to have a relatively low density, particularly at the primary particle centers, making them appear hollow (Figure S8). The relatively solid primary particles appear to be slightly larger, with 20–40 nm diameters, than the hollow primary particles, which have 15–35 nm diameters typically. Representative HRTEM images of the primary particle types (Figure S6) show that the solid particles have a shell of relatively graphitic carbon, although it appears more disordered and less dense than in the large Vulcan particles. The interior of these particles appears to be filled with amorphous carbon. The hollow particles also show a relatively graphitic shell, but have an interior with much lower density. The HRTEM image suggests that the interior may be empty or may contain a few loose, crumpled graphene sheets.

Understanding the structure of the primary particle interior and its porosity is key to explaining the transport behavior and durability of HSC-supported catalysts. Because this structure cannot be inferred from two-dimensional imaging alone, we employed STEM tomography to investigate the 3D structure of HSC carbon supports (Figure 2).

Zoom In Zoom Out Reset image size

The HSC aggregate selected for STEM tomography (Figure 2a) includes both hollow (bottom left) and filled (top right) primary particle clusters. A cross-section from the tomographic reconstruction (Figure 2b) reveals the density (indicated by the intensity) and porosity of the carbon interior. The hollow primary particles show a clear, dense shell, and an interior with effectively zero intensity (same as background) in most regions. Some sparse strips of lower intensity material are visible in parts of the interior, interpretable as either loose graphene sheets or very thin walls between primary particles. Openings in the dense shell of the hollow particles are also visible, with typical diameters around 2–5 nm.

The filled primary particles show a significant amount of material with similar intensity to the shells on their interiors, as well as clear pore regions with zero intensity and some regions with intermediate intensity, interpretable as either low density carbon or small pores similar to or smaller than the ∼1 nm resolution of the tomogram. The larger, round pores observed in filled HSC primary particles have typical diameters around 2–8 nm, and are much smaller than the interior spaces in the hollow primary particles. Close inspection of the pores in filled primary particles show that narrow channels are observed connecting the larger pores to the carbon exterior. One such structure is shown in Figures 2c–2e, highlighted in magenta, with a ∼1 nm channel connecting a larger ∼5 nm diameter void to the primary particle exterior.

Segmentation of the tomogram to distinguish carbon and identify distinct pore regions yields additional insights into the structure of HSC. Figure 2f shows the segmented tomogram rendered in 3D, and Figure 2g shows a cross-section through the segmented tomogram (corresponding to the cross-section in (b)). The very narrow channels shown in (c-e) are not separated from carbon in the segmentation, so pores regions accessible through only these channels appear "closed", and are shown in blue in (f,g). Pore regions with larger openings are shown in red in (f,g), with the shade of red indicating the distance from the nearest exterior opening. The two different types of pores are highly segregated on the carbon aggregate, with the large-opening pores found almost exclusively in the cluster of hollow primary particles, and the smaller, channel-connected pores found almost exclusively in the filled primary particles. Walls between the hollow primary particles are typically incomplete or absent entirely, and as a result these large pores form a single interconnected network throughout the cluster of hollow particles. Openings in the exterior walls are frequent enough that the distance to the nearest opening inside the pore network (g) is typically similar to the primary particle radius.

The prevalence of each type of primary particle was estimated from an ensemble of ADF STEM images by manual classification of primary particle clusters (as in Figure S8). Analysis of the integrated intensity of each type in the image ensemble and porosity of each type in the tomogram (detailed in the supplemental information) suggests that hollow carbon particles compose 43 ± 7% of the total volume and 29 ± 7% of the total mass in HSC.

The pore size distribution for the smaller type, channel-accessed pores can be determined from the segmented tomogram by approximated these pores as spherical to calculate the effective diameter of each pore. The corresponding distribution (Figure S9a) of diameters d is well fit by an exponential distribution ∼e^−d/μ with μ = 1.7 ± 0.3 nm, provided that the distribution is truncated to the minimum detectable pore size. Determining a size distribution for the network of larger pores in hollow primary particles is less straightforward, but can be achieved by assigning a local diameter to each semi-confined pore region (see supplemental for details). Figure S9b shows the volume-weighted pore size distribution for closed and open pores, accounting for the prevalence of each primary particle type. Open pores in hollow primary particles contribute vastly more to the total pore volume than closed pores. Open pores are larger, with most volume in pores with 10–20 nm diameters, while closed pores have most volume at diameters less than 10 nm. The impact of constrictive bottlenecks between pores and at pore openings can be assessed by tracking the limiting bottleneck diameter to access each pore region. The fraction of total pore volume that can be accessed through openings of a given size is shown in Figure S9b. This measurement overestimates the typical opening size somewhat because the size of some openings is exaggerated by blurring due to missing wedge artifacts in electron tomography. Despite this, it is evident that the vast majority of interior pore volume can only be accessed through bottlenecks significantly smaller than the pores.

Measurements of the primary pore diameter distribution from gas desorption isotherms have previously been reported, showing large pore volumes below ∼6 nm.⁴⁵ Because desorption isotherms measure the diameter of the pore opening rather than the interior volume,⁴⁶ we suspect that these measurements are attributable to the small 2–5 nm openings into the hollow primary particles observed in tomography.

Overall, it is clear that the morphology of HSC carbon supports can be highly heterogeneous, with either hollow primary particles with large interior pores or mostly filled primary particles with smaller pores accessible through very small channels. Both pore structures can contain Pt nanoparticles in HSC-supported catalysts. Discussion of reactant transport must consider that Pt in either pore type can only be accessed through relatively small bottleneck constrictions. This restrictive pore geometry is also likely to limit the migration and coalescence of Pt nanoparticles on HSC supports, and will thus have implications for the catalyst durability as well.

The differences in primary particle structure and porosity between HSC and Vulcan result in dramatic differences in the distribution of platinum on each support. STEM tomography was performed for each of the Pt/HSC and Pt/Vulcan specimens to allow visualization and quantitative measurement of the platinum distribution. Figure 3 shows 3D visualizations of segmented tomograms for each of the samples, with carbon-interior Pt particles shown in blue and carbon-exterior Pt particles shown in red, alongside 2D cross-sections from the tomograms. It is qualitatively apparent that in HSC interior particles are dominant, as the 3D visualizations appear mostly blue, while in Vulcan exterior particles are dominant, as the 3D visualizations appear mostly red. Note that Vulcan does have interior Pt particles, which are clearly visible in the tomogram cross-sections, although they are relatively small, having diameters of only 1–2 nm. As we will discuss in more detail below, the majority of Pt surface area resides in the carbon interior in all HSC specimens and on the carbon exterior in all Vulcan specimens.

Zoom In Zoom Out Reset image size

For Vulcan, the variation in primary particle size and morphology was investigated by acquiring separate tomograms of small and large carbon primary particles for each platinum loading (Figure S10). To calculate quantitative values that are comparable to bulk properties, we consider the relative weighting of the large and small Vulcan primary particles. Weighting factors for the larger (22.5–45 nm) and smaller (0–22.5 nm) Vulcan primary particles was calculated using the size distribution reported by Ferraro et al.⁴⁴ and the concentration of platinum measured in the tomograms for each loading and Vulcan type.

While the larger primary particles have approximately 7 times more volume than small primary particles, we find that, due to their higher surface area and porosity, the smaller primary particles support approximately 2.7 times more Pt per carbon volume in 10%Pt/Vulcan and approximately 20% more in 50%Pt/Vulcan. This implies that small primary particles support approximately 28% of the Pt in 10%Pt/Vulcan, and 15% of the Pt in 50%Pt/Vulcan. Uncertainty in the weighting due to the choice of particle type cutoff diameter and platinum content was included in calculations. The two dominant primary particle types in HSC, hollow and filled, were discussed previously. All reconstructions shown here for HSC represent the filled primary particle type. Hollow HSC particles in the platinized samples were found to be more radiation sensitive, and reconstructions of sufficient quality for automated statistical analysis could not be obtained by the same methods. This may be caused by the Pt particles catalyzing beam-induced carbon oxidation with small amounts of water adsorbed on the sample surface. All platinized samples showed some degree of damage during tomography, although the hollow HSC structure showed more motion, possibly because the lower density of carbon provides less structural stability.

Quantitative analysis of the segmented 3D reconstructions allows us to determine the trends in catalyst distribution as the support morphology and Pt content are varied (Figure 4). For HSC specimens we find that the majority of Pt surface area (60–70%) is embedded in the carbon interior (consistent with independent TEM tomography measurements).^25,47 For Vulcan specimens, the majority of Pt surface area resides on the carbon exterior, with the higher-porosity small primary particles having a relatively larger fraction on the support interior. For all specimens, increasing the Pt content decreases the fraction of catalyst area on the support interior. This effect, which is large for Vulcan and small for HSC, can have two main contributing factors: the number of additional nanoparticles that nucleate on the carbon interior or exterior, and how the particle size changes with loading. For HSC, as Pt content increases additional Pt particles have roughly equal probability of forming on the interior and exterior, causing no significant change in the fraction of Pt particles on the interior. In Vulcan, additional particles preferentially nucleate on the carbon exterior because of the limited interior space available, causing the interior particle fraction to drop. The observed Pt distributions are consistent with reasonable expectations for Pt nanoparticles synthesized directly on the carbon support, as was done in this study.

Zoom In Zoom Out Reset image size

Figure 4b shows a comparison of mean diameters derived from distribution fits. The size distributions of Pt particles on the interior and exterior in all specimens are well fit by lognormal distributions (Figure S4), truncated to account for the minimum detectable nanoparticle size. (Fit parameters are summarized in Figure S11 and Table S1.) For all carbon support morphologies, exterior particles have a larger mean size than interior ones. The mean diameter of interior particles becomes significantly larger with increasing support porosity, whereas the mean diameter of exterior particles changes little with support morphology. Increasing the Pt content tends to increase the mean particle size on all supports, but the mean size for exterior particles increases significantly more. Furthermore, the mean size for interior particles increases more for more porous supports. This indicates that interior pores confine the growth of large nanoparticles, especially in low porosity supports, where interior spaces can quickly become saturated. As Pt content increases from 10 wt% to 50 wt%, the combination of preferential nucleation of additional nanoparticles on the exterior of Vulcan and the faster growth in exterior particle sizes relative to HSC results in a large decrease in the interior surface area fraction for Vulcan and a small one for HSC, as shown in Figure 4a.

To understand the impacts of catalyst morphology on the accessibility of the Pt, we measured the electrochemically active surface area (ECSA) of the catalyst specimens using two techniques: hydrogen adsorption/desorption (HAD) in an acidic aqueous electrolyte and CO-stripping in fuel cell MEAs. Figure 5a shows a comparison of the specific surface area (SSA) measurements made by STEM tomography (closed symbols) and the electrochemical (open symbols) techniques. We find that the ECSA measured by HAD in acidic aqueous electrolyte (black circles) is broadly consistent with the total physical surface area observed in tomography (blue squares), showing decreasing SSA with increasing Pt loading on either support, and slightly higher SSA for HSC than for Vulcan at each loading. Note that the SSA calculated including only the surface area of exterior particles is insufficient to account for the ECSA. This indicates that all of the catalyst is accessible to protons in liquid within our uncertainty, including Pt surface that is in contact with the carbon or embedded inside the support. The HSC-supported Pt analyzed here is embedded in the relatively solid HSC structure. Because the vast majority of embedded catalyst particles appear to be proton-accessible, we infer that proton conduction can occur in the narrow ∼1 nm channels shown in Figures 2c–2e when they are saturated with aqueous electrolyte.

The ECSA measured in MEAs by CO-stripping at 100% relative humidity (RH) follows the same trend with Pt content observed in HAD and tomography. However, the overall ECSA is lower in CO-stripping, compared to either HAD or tomography, especially for HSC. This may be due to dissolution of very small Pt nanoparticles in the MEA as CO-stripping was performed after one day of fuel cell operation while HAD and tomography were done on fresh samples. In addition, ECSA measurement at elevated temperature (80 vs 25°C) have been shown to give slightly lower values.⁴⁸

Shinozaki et al. showed that CO stripping could be useful in expressing the accessibility of Pt/C catalysts.³⁰ The electrochemical oxidation of CO requires H₂O and proton accessibility, through either ionomer or water condensed on the catalyst. As the RH decreases, only CO adsorbed on Pt that is in close proximity to ionomer will be oxidized.⁴⁹ Shinozaki et al. demonstrated that catalysts with a large fraction of interior Pt such as Pt/HSC showed a large decrease in CO stripping charge at low RH, while catalyst with a small fraction of interior Pt such as Pt/Vulcan showed negligible decrease in the CO stripping charge. This is consistent with the assumption that ionomer cannot infiltrate small pore openings to interact with Pt inside the carbon primary particles.

Using this method, we determined the Pt ECSA and utilization at different RH for the four catalysts (Figure 5b). The decrease in the CO stripping area with RH depends on the existence of pathways through ionomer or capillary-condensed water to Pt, which are influenced by the distribution of Pt particles with respect to the carbon, the carbon pore size, and the pore hydrophilicity. To quantify the Pt particles inside the carbon pores that are inaccessible to ionomer, CO stripping measurement should include measurements at low RH where capillary condensation is minimal. However, measurement at 0% RH is not possible because the CO oxidation reaction requires water. Quantification of the interior particle fraction is straightforward for most catalysts, where the CO stripping area reaches a plateau at low RH. However, for some catalysts, such as 50%Pt/HSC, the CO stripping area continues to change at low RH, contributing significant uncertainty to the measurement. As shown in Figure 5c, we find that the platinum utilization at low RH (10–20%), normalized to the ECSA at 100% RH, is consistent with the fraction of total surface area on the carbon support exterior measured by tomography. The overall agreement suggests that it is reasonable to interpret the utilization lost at low RH, measured by RH-dependent CO-stripping, as an indication of the fraction of catalyst surface that is embedded in primary pores of the support, and that these pores are not infiltrated by ionomer.