Journal Archive

Johnson Matthey Technol. Rev., 2022, 66, (3), 355
doi: 10.1595/205651322X16508983994949

Combining State of the Art Open Source and Proprietary Machine Learning Technologies to Build a Data Analysis Pipeline for Gasoline Particulate Filters using X-Ray Microscopy, Focused Ion Beam-Scanning Electron Microscopy and Transmission Electron Microscopy


  • Aakash Varambhia*, Angela E. Goode, Ryutaro Sato, Trung Tran
  • Johnson Matthey, Blounts Court, Sonning Common, Reading, RG4 9NH, UK
  • Alissa Stratulat, Markus Boese
  • ZEISS Microscopy, Carl-Zeiss-Str. 22, 73447 Oberkochen, Germany
  • Gareth Hatton, Dogan Ozkaya
  • Johnson Matthey, Blounts Court, Sonning Common, Reading, RG4 9NH, UK
  • *Email: aakashm.varambhia@matthey.com

PEER REVIEWED
Submitted 19th November 2021; Revised 25th March 2022; Accepted 12th April 2022; Online 25th April 2022

 


Article Synopsis

The performance of a particulate filter is determined by properties that span the macro, meso and atomic scales. Traditionally, the primary role of a gasoline particulate filter (GPF) is to reduce solid particles and liquid droplets. At the macro scale, transport of gas through a filter’s channels and interconnecting pores act as main transport arteries for catalytically active sites. At the meso scale, the micropore structure is important for ensuring that enough active sites are accessible for the gas to reach the catalyst nanoparticles. At the atomic scale, the structure of the catalyst material determines the performance and selectivity within the filter. Understanding all length scales requires a correlative approach but this is often quite difficult to achieve due to the number of software packages a scientist has to deal with. We demonstrate how current state-of-the-art approaches in the field can be combined into a streamlined pipeline to characterise particulate filters by digitally reconstructing the sample, analysing it at high throughput, and eventually use the result as an input for gas flow simulations and better product design.

1. Background

Autocatalysts are one of the most effective ways to reduce pollution such as smog from car exhausts (1). The autocatalyst does this by removing vehicle exhaust outputs such as carbon monoxide, nitric oxides and soot to legislated levels (2). The autocatalyst is made up of groups of channels composed of monolithic catalyst supports that are either in a square or honeycomb structure (3).

One type of catalyst is where the washcoat is coated on a porous ceramic wall-flow filter monolith, which has a pattern where every other channel is either an inlet or plug (4). Such a configuration allows the gas to flow through the wall of the monolith, trap the particulate matter, and thereby increase the surface area contact of the gas with the active material. The catalyst coating is usually porous and is made up of alumina and ceria-zirconia supports (5) loaded with platinum group metal nanoparticles.

The quantitative investigation of performance criteria across length scales requires a correlative approach. For such monoliths, the performance of the particulate filter is determined by three aspects across the micrometre to Ångstrom length scales. The interconnecting pores within the wall structure affect the transport of gas to the catalyst active sites (6), this effect can be observed at the macro and meso scales. The morphology and atomic structure of the catalyst nanoparticles affect the selectivity and performance required to reduce pollution (7, 8).

Here, a combination of X-ray microscopy (XRM), focused ion beam–scanning electron microscopy (FIB-SEM) and transmission electron microscopy (TEM) has been used. Analysis in the XRM is carried out by taking multiple two-dimensional (2D) images of a rotating object and reconstructing the images into a digital three-dimensional (3D) structure. Similarly, in the FIB-SEM a smaller area of interest is picked out and a volume is created by precisely milling the sample and recoding the image which is subsequently added to the 3D slice. In the TEM, the samples can be imaged as single images, spectroscopic data cubes and 3D reconstructions by acquiring a tilt series acquisition.

XRM is well suited to observing monolith properties in 3D, such as coating uniformity and pore network distribution at micrometre scale resolutions (9). FIB-SEM is used to investigate smaller features in 3D, which are often missed by the XRM resolution limit, which is ~0.8 μm in this work (10). To go beyond the device scale, FIB-SEM can be used to extract 50–100 nm thick lamellae from the sample for further TEM imaging (11). This allows the characterisation of nanoparticle atomic structure, as well as higher spatial resolution composition analysis of the sample (12, 13). The TEM also offers further high-level characterisation such as oxidation state mapping as well (14).

When trying to correlate performance to rational catalyst design, several observations, measurements and iterative refinements are necessary. This agile approach requires fast characterisation of materials in a correlative way using a methodology aided by machine learning (ML) and automation (15). Moreover, in industry, this must be done rapidly due to high turnover using microscopes such as the XRM, FIB-SEM and TEM.

Data acquired from the instruments is challenging to quantify due to its complexity and quantity. We showcase a pipeline that can aid with streamlining such analysis. Moreover, expertise in data science and engineering is required to handle the data in a structured manner to build an efficient analysis pipeline. With the emergence of new proprietary tools and open source libraries we demonstrate how both technologies can be leveraged effectively in combination with in-house developed code to deliver comprehensive analysis.

2. Methods

2.1 Sample Preparation

In this paper we concentrate on a research-grade GPF designed to demonstrate our analysis workflow presented here. The filter sample was prepared and chosen as a model sample by ensuring that the washcoat is applied onto the wall and into the wall of the support structure. The monolith support structure is made up of a porous ceramic material and the catalyst coating is made up of alumina with ceria-zirconia supports with palladium nanoparticles.

In order to perform a correlative study, the sample was imaged sequentially using each microscope, starting first with the XRM, then with the FIB-SEM and finally with the TEM. For correlative studies, it is often best to investigate the sample at a large field of view to pick out regions of interest that can then be studied at a higher resolution.

For the XRM, a section of the monolith was cut into an area spanning a few square centimetres from the middle of the sample and embedded in resin. Once both monolith and coating were scanned using the XRM a smaller area was selected from the coating to be analysed by FIB-SEM. It is difficult to determine which area will be representative of the whole sample. A statistical sampling study is required in the long term. This was outside the scope of the present paper as we are demonstrating the 3D analysis workflow, therefore we randomly chose a region of interest to mill.

3D structural information was acquired by sequential milling of the sample using the ion beam and then imaging the polished cross-sections. Following 3D analysis, a TEM lamella was prepared from an immediately adjacent region of the sample. The lamella was mounted onto a TEM omniprobe grid and thinned to electron-transparency for high spatial resolution TEM imaging and spectroscopic elemental mapping.

2.2 Image Segmentation with Machine Learning

Segmentation lays the foundation for all subsequent quantitative analysis, thus proving to be the most critical component of any image analysis workflow (16). Recent ML approaches such as deep learning have proven to be highly successful in solving several tough computer vision tasks (17). This is especially true for instances where conventional approaches such as thresholding have failed before (18). However, deep learning methods require several manually prelabelled datasets to obtain meaningful results. Here in collaboration with ZEISS, Germany, we have deployed a workflow that utilises ML to perform pixelwise image segmentation without needing several images of training data. The ML algorithms used within the software are well known within the computer vision field and discussed later in this section, however our main focus is in creating a pipeline to deploy the workflow in a streamlined manner. ZEISS has provided a GitHub repository (19) with examples for how users can interact with the ZEISS ZEN IntellesisTM segmentation module for their imaging and analysis suites in ZEISS ZEN (Blue Edition) and ZEISS ZEN core software package. Figure 1 shows how this is achieved at a high level for a monolith XRM image.

Fig. 1.

Image segmentation pipeline using machine learning and custom pore masking to segment an XRM image of a monolith channel

The segmentation technique in Figure 1 is similar to the Weka segmentation module in ImageJ (20) which is available as freeware. Similar capabilities are also offered by the ilastik software for free as well (21). An advantage IntellesisTM offers is increased ease-of-use through presets for all parameters which is ideal in an industrial setting. A core strength of segmenting images in the manner presented in Figure 1 is that the segmentation model only needs to be trained on one image before it can be used to predict the pixel labels in other images taken under similar experimental conditions (22). In contrast, more sophisticated deep learning models require a large number of images that might be difficult to gather and record.

However, a disadvantage of such methods is that they are very difficult to benchmark and compare against one another given the lack of a ground truth measurement for most samples and is out of the scope of this publication. In these situations, the segmentation result needs to be judged by an expert as the ML model is trained on a selection of pixels within an image chosen by the microscopist. The microscopist ‘paints’ over various pixels within the image which are associated to different classes. In this manner, the experience of the microscopist is leveraged during the labelling process.

The images then undergo a step to extract features that are of interest. Various filter kernels are used to extract features from the image. Filter modes can be chosen between a ‘classical’ mode utilising the central processing unit (CPU) or a ‘deep feature extraction’ mode utilising the graphics processing unit (GPU). In the classical mode, a range of filters that can extract features such as vertical, horizontal and boundary information are used. In total there are 33 filters that are used to augment the information obtained per pixel before training. In deep feature extraction mode, sample features that describe the shape and textures in the image are extracted from up to the first three layers of a VGG-19 neural network (23).

After filtering, the painted pixels are concatenated into a feature vector that is used as an input to a classifier such as random forest (24) or XGBoost (25) algorithms. The choice of classifier can vary depending on the image and performance desired. For segmentation, random forest classifiers have been popular in recent years due to their versatility. Further post processing is undertaken using conditional random fields to improve the segmentation results by assigning confidence levels to pixels in the image for a given label (26). Once the image has been segmented by IntellesisTM, the label mask can be post processed by a combination of Gaussian filtering and binary erosion to create a mask from which the pores labels can be obtained by thresholding.

Post training, the model can be exported into a file format containing information about the model (training images and labels) and processing steps. With automation features within ZEN, we applied our trained model to segment the subsequent image slices for 3D tomographic reconstruction, Figure 2. Whilst model training is undertaken within IntellesisTM, data preprocessing such as normalisation, dataset rotation and cropping is performed using in-house Python code.

Fig. 2.

Image segmentation processing pipeline for 3D tomography data for XRM and FIB image stacks

After initial dataset alignment through rotation and cropping, the images are normalised with respect to the centre image using the histogram matching technique ensuring that all pixels share a similar intensity distribution range. Normalisation is a situational preprocessing step to be used in instances where the intensity within the image stack varies slice by slice (27). Normalisation ensures that the segmentation model can accurately classify pixels with the desired outcome without systemic errors in classification.

An image is selected as a training image usually from the middle of the stack as the initial input for the classifier. Starting with ZEN 3.2 (Blue Edition) and ZEN core 3.1 it is possible to import externally trained deep learning models created as Python packages (28) into the IntellesisTM software. Recently, ZEISS has also provided a repository with examples of how to perform automated analysis on GitHub (19). Once the classifier is trained it can be used to segment the rest of the images within the stack.

Segmented labels can be used to create 3D reconstructions of the dataset for further processing. With such label datasets, it is then possible to undertake meaningful measurements of a sample such as percentage composition measurements, pore and particle size measurements, tortuosity and use of voxelised data as inputs for gas flow simulations.

2.3 XRM Data Acquisition and Processing

The XRM tomography dataset was acquired on a ZEISS XradiaTM 520 Versa microscope operated at 80 kV acceleration voltage with a pixel size of 4.74 μm with 20 s dwell time. Data acquisition in the XRM is carried out by taking multiple 2D transmission images of an object from different orientations. Once the acquisition was complete, the data was reconstructed using inbuilt automatic tomography reconstruction algorithms within the ZEISS Scout‐and-Scan software. It is possible to create image slices from any direction of the sample, Figure 3.

Fig. 3.

XRM scan showing four channels from a monolith. The slice view directions along x-, y- and z-directions provide an overview of the washcoat coating along the channel

An image stack with 994 images was then segmented along the z-direction using the proposed image analysis pipeline in Figure 2. At the XRM length-scale we are able to observe the monolith’s properties at the macro scale such as phase volume fractions, pore connectivity, size and tortuosity which are the key properties that need to be measured at this scale.

Phase volumes for the monolith sample were calculated by counting the number of voxels within the segmented image and calculating the percentage volume fraction. We note that there are more accurate approaches for calculating the volume by triangulating the voxel dataset into a mesh (29). However, this does not affect the volume drastically compared to measuring surface area of the material, which is often overestimated by the voxel counting approach. Phase distributions along a selected slice direction throughout the sample were computed by examining percentage changes to segmented label stacks. This is useful for investigating anisotropy of the washcoat distribution along the monolith channel.

Pore connectivity in a sample is also a parameter that is often critical in determining gas flow and bulk transport properties (30). A well-connected monolith wall pore network ensures that most active sites are reached as gas travels through the channel wall. To measure connectivity, a nearest neighbour detection approach within Python’s scikit-image library (31) was used where a label detection kernel is scanned within the image to assign a label to each pore object within the segmented image (32). Similar analysis can also be performed using AvizoTM software but the main advantage of using open source libraries in Python is that most of the analysis is incorporated into a single pipeline without any data conversion, file read-write and automation bottlenecks.

Provided that the pores are segmented it is also possible to perform a distance-transform based morphological sphere fitting to obtain pore sizes. We have used the definition that a voxel is part of a pore with a given diameter for a sphere that can completely fit or be packed within the pore space (33). Figure 4 shows what the fitting procedure looks like visually for a synthetically generated pore structure. The superposition of all ‘sphere’ maps with a given diameter can be included in the pore space as equal to the pore volume they occupy relative to the total volume.

Fig. 4.

Pore diameter measurements using a morphological distance transform algorithm to ‘fit’ spheres to pores within a binary segmented image

The fitting approach can be further expanded to emulate mercury porosimetry undertaken in laboratory tests. This was not investigated within this paper and will be addressed in a future study. Alternatively, gas flow through the wall of a monolith channel can be simulated to understand gas pathways in more detail. The simulation was performed using the segmented stereolithography data meshed at refinement level 5 in snappyHexMesh processed by OpenFOAM® software.

2.4 FIB-SEM Data Acquisition and Processing

FIB-SEM tomography was performed using a ZEISS Crossbeam 550 microscope. ZEISS Atlas 5 software was used to align XRM and FIB-SEM datasets and locate the region of interest, which in this case was focussing on the washcoat (Figure 5). A tomogram was then collected with a pixel size of 3 nm along the x- and y-directions, and a 9 nm pixel size along the z-direction.

Fig. 5.

FIB region of interest chosen from the XRM dataset scan using the ZEISS Atlas 5 software. A series of images from the region of interest were stitched and a curtaining removal filter was applied before segmentation. The workflow was applied to all the images in the stack

One of the milling artefacts often present in inhomogeneous samples is ‘curtaining’ (34) as seen in Figure 5. This effect can be noticed as faint vertical lines on the polished cross-section due to variations in sputtering rates across the sample. The curtaining effect severely hampers any segmentation routine due to uneven intensity changes in the image. A Fourier filtering routine followed with a median filter was applied to all the images to correct for the curtaining effect using a script coded in Python using a mythology proposed by Schwartz et al. (35).

It is worth mentioning that an additional preprocessing step that can also be undertaken is an intensity anisotropy correction. Often, collected images exhibit an intensity gradient due to shadowing of electrons originating from the bottom of the sample face and the FIB-SEM detector geometry. For our images we did not observe a major intensity gradient and did not apply this correction, however a publication by Tallion et al. (29) provides further detail on how such a correction can be performed.

After preprocessing the stack, the images were segmented using the same pipeline shown in Figure 2. In the segmented image, the ceria-zirconia is shown in pink, the alumina is shown in green, and all other colours are associated with the rest of the monolith. Corresponding energy dispersive X-ray (EDX) mapping was performed to confirm the elemental composition of the sample.

The pores within the washcoat layer are present at the nanoscale, they are therefore not well resolved by the FIB-SEM acquisition parameters used in this work, where a balance had to be struck between the total volume analysed and the voxel size used. Rather than pore tortuosity, an alternative measurement that can provide insight into diffusion through the washcoat layer is to quantify the amount of tortuosity of each phase (36). The degree of tortuosity can provide an indication of how easy it is for a gas or liquid to diffuse through a material (37). It is a property linked to effective diffusivity and both electrical and ionic conductivity through a material (38).

To calculate tortuosity, we have used a definition based purely on geometric considerations using a method proposed by Tallion et al. (29). This approach makes it simpler to calculate the physical tortuousness within the microstructure. An advantage of calculating tortuosity this way means that the calculated values can be easily compared between different phases as they do not rely on a material’s diffusivity or conductivity. We demonstrate how this procedure works using three synthetically generated model samples with increasing compression, Figure 6. For demonstration purposes, the spheres were packed tighter so that the empty space between where the gas would flow is more tortuous.

Fig. 6.

Model samples with 0%, 10% and 20% compression applied to make the empty space between the spheres more tortuous. The 3D hexagon represents the ‘top’ of the sample marking the direction of gas flow

In this approach, tortuosity is calculated as the ratio of the geodesic distance (39) and the Euclidean distance. We have calculated the Geodesic distance using the fast-marching algorithm with the scikit-fmm Python library. The Euclidean distance was calculated as a straight-line measurement parallel to the direction of interest. By taking a ratio of the geodesic and Euclidean distance maps, we can obtain a 3D tortuosity map of the material or the pore phase of interest, Figure 7.

Fig. 7.

A 3D tortuosity map for a model sample overlaid with the position of model spheres to geometrically demonstrate how tortuous a path may be for gas to flow through

By slicing the 3D tortuosity map, an average tortuosity profile along a direction of interest can be obtained, Figure 8. With this demonstration, it can be seen that calculating the tortuosity between model samples can provide an effective measurement that is straightforward to compare between different samples. Moreover, the tortuosity line profiles provide more detailed variational information along the direction of the sample compared to a global average value for a volume which is what is often calculated in literature (40).

Fig. 8.

Tortuosity line profiles for the empty space between the spheres along the direction of gas flow in model samples with varying compression

Tortuosity in the ceria-zirconia phase is usually of interest as it is the pathway gas would take to reach the sites where the palladium nanoparticles are located. A portion of the volume in Figure 5 was cropped to calculate the geometric tortuosity starting from the surface of the washcoat traversing towards the substrate wall and ending with a subsequent line profile calculation.

2.5 TEM Data Acquisition and Processing

Characterisation of the nanoparticles was undertaken using the TEM which can atomically resolve the structure of the particles through imaging and compositional characterisation. Before TEM imaging was undertaken, a lamella from the washcoat region was extracted using the FIB-SEM, Figure 9. High angle annular dark field (HAADF) atomic resolution imaging was performed using a probe-corrected ARM200CF (JEOL, USA) scanning transmission electron microscope (STEM) operating at 200 kV.

Fig. 9.

(a) FIB based TEM lamella extracted using in situ lift out technique, the sample was imaged at: (b) different magnifications; and (c) various elements were mapped using the EDX detector (green = alumina, magenta = cerium, blue = palladium)

The column of images on the left in Figure 9 illustrates the lamella preparation by FIB-SEM, while the image in the middle shows the thinned region to be imaged with the TEM. Images on the right show an area with an annular dark field (ADF) image with its simultaneously acquired spectroscopic maps. The EDX map in Figure 9 was taken from an interface area where ceria-zirconia meets alumina. It is also important to note that the palladium nanoparticles are detected within the ceria layer. The palladium particles are the active ingredient in this case where the catalysis happens. The structure and size of the nanoparticles is important as it determines the surface area available for catalysts and was imaged using the TEM.

Here we have carried out some high resolution, aberration corrected imaging of the FIB slice, specifically, the ceria-zirconia component. This component is a very important part of this catalyst structure in determining the oxygen available for catalysis to progress. This component acts as an oxygen regulator within the structure (41). It is important that its structure and composition are uniform.

We also attempted to record a series of images with increasing magnification from the cerium region of the sample, however performing simultaneous spectroscopic acquisition proved to be challenging. As this part of the study was to demonstrate a proof of concept that it is possible to obtain atomic resolution imaging with spectroscopy, we performed spectroscopy on a similar ceria sample. In the future we will be exploring this aspect further to bridge this length scale correlatively. After sufficiently zooming into the sample, elemental mapping was performed by electron energy loss spectroscopy (EELS) spectrum imaging (SI) using a GIF Quantum® ER 965 model spectrometer. To produce the elemental maps from the EELS-SIs, power-law background stripping was applied and Hartree Slater cross-sections (42, 43) were employed for fitting zirconium M edge, cerium M edge and oxygen K edge.

3. Results and Discussion

For a correlative through the length-scale study, the sample was first analysed using the XRM, then the FIB-SEM and finally the TEM. Each technique has its advantages and disadvantages. In the simplest case, the field of view that can be observed by each instrument is inversely proportional to the resolution that can be achieved. For instance, at the atomic scale only a few nanoparticles can be observed in detail for a given field of view, whereas at the meso scale, several nanoparticles can be observed but with less resolution.

XRM can provide the macro scale measurements such as washcoat morphology and large pore size distributions. One of the key parameters of interest is the anisotropy and the average composition of washcoat, substrate and pore characterisation of the wall, Figure 10. A uniform washcoat, pore and substrate distribution can be observed along the channel indicating good control over process engineering. Consistent and precise coverage means that the performance can be finely controlled by changing the washcoat coverage methodology. When such a workflow is aided with iterative testing, rapid analysis processing pipelines and simulations, the process engineering can be fine tuned to develop rational GPF design at the device level.

Fig. 10.

Composition characterisation along a monolith channel. The line profiles in the line plot are colour coded with the same colours as the pie chart. The pie chart provides the total volume of the various phases within the 3D scan, whereas the line profiles provide the coverage information of each image slice along the z-direction shown in Figure 3

Coverage information is not sufficient if one wants to understand the flow of gas through a monolith channel. A qualitative approach that can be used is visualising the connectivity in 3D for a material or pore. Volumes of the newly assigned pore labels relative to the total pore volume provide a representation of how well connected the pores are within a monolith wall. Isolated pores are identified by label segments that are not connected to the larger pores, ranked by overall volume fraction. When this technique is applied in 3D, it is possible to visualise the pore connectivity of a monolith wall in a straightforward manner, Figure 11 visualised by AvizoTM software.

Fig. 11.

Pore regions segmented and labelled from a selected region of a GPF monolith. The coloured labels represent the pore regions and their label assignments. The catalyst coating is shown in cyan, and the substrate is shown in pink

For a more quantitative measurement, the labelled pore regions in Figure 11 can be ranked by volume as a function of the total pore volume and be interpreted as a measure of connectivity. Figure 12 shows the top 10 labelled regions of the pore volume as a function of percentage connectivity. There are four dominant connected regions within this channel that make up more than 60% of the pore space.

Fig. 12.

Top 10 pore labels ranked as percentage connectivity derived from the total pore volume within the monolith

For any XRM measurement, we are only able to measure the macro scale pores and therefore the connectivity within the sample may be higher than reported. To measure the range of pores that we can observe a sphere fitting routine was used indicating a minimum pore diameter of 9.48 μm and a mean pore diameter of 29.06 μm, Figure 13. The primary y-axis of the plot provides the likelihood of a pore existing within the pore volume space given all the measured diameters. The secondary y-axis provides the cumulative percentage a pore is either greater than or smaller than a given pore diameter represented by the black dashed lines. We can observe that ~30% of the pores are greater than the average pore diameter. In this instance we were not able to verify the measured pore sizes through other means such as mercury porosimetry for this sample as we mostly concentrated on demonstrating the 3D analysis workflow. For a more rigorous analysis and experiment design, one should measure the pore sizes through physical measurements and tune the XRM resolution appropriately to image most of the pores within the monolith.

Fig. 13.

Measured pore size distribution from the XRM scan where the mean and median pore size are 29.06 μm and 28.44 μm respectively

Given the resolution limitation of the XRM, we are mostly interested in how gas would flow through the wall of a monolith and identify the major flow pathways that affect the performance of the GPF. Simulated streamlines of the gas velocity vector magnitude U provide good qualitative information of the major pathways that a gas would flow through, Figure 14. However, we are computationally limited by the number of streamlines we can visualise within a 3D structure. In principle it is possible to visualise all streamlines within a structure, but the visualisation depends highly on the mesh refinement and XRM resolution.

Fig. 14.

Gas flow simulation for a selected area of monolith wall (green rectangle) where the colour coded streamlines present the magnitude of the velocity vector U along x-, y- and z-directions in metres per second

Most of the gas flow is determined by the macropores within the monolith where the flow is proportional to the pore size. For example, in the case of a cylinder, the flow in pores is proportional to pore size to the power 4 and can be related to the pressure drop using the Hagen-Poiseuille equation (44). However, in some instances pore blocking and narrowing that is missed by the XRM resolution can affect interpretations of what we measure. In situations like these, we must image the sample using higher resolution X-ray beamline tomography or FIB-SEM.

Studying the gas flow within the microstructure in detail requires the inclusion of micro- and mesopores, which is difficult to achieve using our XRM. To characterise this in detail one would have to resample the 3D XRM dataset and incorporate the micropores from the FIB-SEM. With this limitation in mind, we determined that further work is necessary to link gas flow simulations to pressure drop and performance of a monolith. To incorporate submicron pores, we would need to characterise submicron porosity using the FIB‐SEM, create a statistical twin of the washcoat and texturise it over the XRM dataset. This is currently outside the scope of this publication and will be looked at in the future.

Alternatively, as a result of the segmented FIB‐SEM images it is possible to do a phase analysis of the washcoat region. We can analyse this set further after the reconstruction to reveal the material content for each of the phases within the structure. This reconstruction is shown in Figure 15 together with EDX maps recorded while the sample was being taken through the FIB tomography routine using the ZEISS Atlas 5 software. EDX maps in a FIB-SEM are useful for confirming the elemental composition of various components, but compositional mapping often comes with limited spatial resolution due to large interaction volumes between the electron beam and the SEM sample.

Fig. 15.

3D reconstruction and visualisation of the subset of the alumina phase viewed from two different directions. The images on the right show the segmented overlay and the subsequent EDX map from a slice taken after milling the sample

The best way to distinguish the various components in 3D remains through backscattered images once the intensities of components have been indexed to specific phases by EDX maps. In this sample, it was possible to map the ceria-zirconia and alumina components to specific greyscale ranges, and therefore these phases can be thresholded using the backscattered intensity which has much higher spatial resolution. Following this method, the research grade GPF sample was calculated to have a 76% ceria-zirconia and a 24% alumina composition. Experimentally, the sample was synthetised with a 75% ceria-zirconia and 25% alumina loading which is in good agreement with the observed measurement from the FIB-SEM dataset and provides confidence that the subregion analysed comprises a representative volume. Furthermore, 3D visualisation of the segmented dataset (Figure 14) reveals that the alumina phase (green) is mostly present as disconnected islands.

Since we are interested in understanding gas diffusion within the washcoat, we measured the tortuosity of the ceria-zirconia phase, which is both the major component of the washcoat and the material with higher porosity. Calculating the geometric tortuosity, Figure 16, of the ceria-zirconia phase can provide an indication of the diffusion pathways a gas could traverse within the sample. Our calculations show that the average tortuosity is close to 1 indicating that the alumina dispersion within the sample does not hinder the gas pathways drastically and that the ceria-zirconia phase is well connected for this sample. For further analysis one would have to segment the pores at a much higher resolution, possibly requiring TEM tomography.

Fig. 16.

Tortuosity profile of the ceria-zirconia phase of the washcoat through the sample starting from the surface of the washcoat

Ultimately to study the performance of the catalyst loaded within the washcoat, we need to study the atomic structure and its uniformity. This analysis is only useful if it is investigated as part of a bigger structure. In order to do imaging at atomic resolution the quality of the sample is paramount. The FIB preparation needs to restrict the amount of amorphous layer caused by FIB preparation and minimise the effect the resin embedding process would have on the imaging.

Here we have carried out some high-resolution imaging of the FIB slice, specifically, the ceria-zirconia component along with a series of atomic resolution images from a similar sample as mentioned in the previous section, without resin embedding. For the experiment, the high magnification imaging was performed on the same sample, however atomic resolution imaging with spectroscopy was performed on a similar sample.

Figure 17 conceptually demonstrates this by combining these images with increasing magnification illustrating the image and spectroscopic acquisition process for ceria-zirconia particles of 10–20 nm size. The EELS maps for cerium, zirconium and oxygen are also shown below the atomic resolution image. For such samples, the ceria-zirconia component is a very important part of the catalyst structure in determining the oxygen available for catalysis to progress. This component acts as an oxygen regulator within the structure (41). It is important that the structure and composition are uniform.

Fig. 17.

High magnification snapshot of the FIB sample conceptually demonstrating increasing magnification up to atomically resolved cerium-zirconium columns with simultaneous EELS mapping. Here the high-resolution images are from the same sample, but the atomic resolution imaging and spectroscopy are from a similar sample

4. Conclusions

At Johnson Matthey we are interested in understanding how our devices function at all length scales in order to understand the material properties that drive the performance of our materials. We show here how a sample can be decomposed at the macro and meso scales using XRM, FIB-SEM and TEM. Properties such as coating uniformity, location and pore network distribution can be determined with XRM. The strength of FIB‐SEM lies in observing the smaller features within the washcoat which are often missed by the resolution limit of the XRM. TEM provides information about the atomic structure of the catalyst coupled with simultaneous spectroscopic information with chemical and oxidation state mapping.

Where possible we have automated the analysis pipeline with a combination of state of the art proprietary, open source and in-house technologies. Integrated pipelines where we can characterise GPF filters correlatively with modelling provides a systematic approach to build understanding of our materials at scale. In the long term such ensemble approaches allow the building of statistically meaningful datasets for product development in the future.

References

  1. 1.
    M. Shelef and R. W. McCabe, Catal. Today, 2000, 62, (1), 35 LINK https://doi.org/10.1016/s0920-5861(00)00407-7
  2. 2.
    L. Tao, E. Garnsey, D. Probert and T. Ridgman, ‘Innovation as Response to Emissions Legislation: Revisiting the Automatic Catalytic Converter at Johnson Matthey’, Centre for Technology Management Working Paper Series No. 2009/3, Institute for Manufacturing, University of Cambridge, Cambridge, UK, 1st February, 2009, 25 pp LINK https://doi.org/10.2139/ssrn.1923084
  3. 3.
    R. E. Hayes, S. T. Kolaczkowskib, P. K. C. Li and S. Awdry, Appl. Catal. B: Environ., 2000, 25, (2–3), 93 LINK https://doi.org/10.1016/s0926-3373(99)00122-8
  4. 4.
    Y.-D. Kim, S.-J. Jeong and W.-S. Kim, Environ. Eng. Sci., 2009, 26, (7), 1171 LINK https://doi.org/10.1089/ees.2008.0088
  5. 5.
    K. Jirátová, J. Balabánová, F. Kovanda, A. Klegová, L. Obalová and R. Fajgar, Catal. Lett., 2017, 147, (6), 1379 LINK https://doi.org/10.1007/s10562-017-2047-z
  6. 6.
    K. Hormann, V. Baranau, D. Hlushkou, A. Höltzel and U. Tallarek, New J. Chem., 2016, 40, (5), 4187 LINK https://doi.org/10.1039/c5nj02814k
  7. 7.
    X. Wang, Z. Zhao, C. Zhang, Q. Li and X. Liang, Catalysts, 2020, 10, (11), 1298 LINK https://doi.org/10.3390/catal10111298
  8. 8.
    T. Rasheed, M. Bilal, C. Li, F. Nabeel, M. Khalid and H. M. N. Iqbal, J. Photochem. Photobiol. B: Biol., 2018, 181, 44 LINK https://doi.org/10.1016/j.jphotobiol.2018.02.024
  9. 9.
    L. L. Lavery, J. Gelb, A. P. Merkle and A. Steinbach, Micros. Today, 2014, 22, (3), 16 LINK https://doi.org/10.1017/s155192951400056x
  10. 10.
    J. Goral, I. Miskovic, J. Gelb and M. Andrew, ‘Correlative XRM and FIB-SEM for (Non)Organic Pore Network Modeling in Woodford Shale Rock Matrix’, International Petroleum Technology Conference (IPTC), Doha, Qatar, 6th–9th December, Paper No. IPTC-18477-MS, 2015, OnePetro Physical, Richardson, USA, 2015 LINK https://doi.org/10.2523/iptc-18477-ms
  11. 11.
    F. Schmidt, M. Kühbacher, U. Gross, A. Kyriakopoulos, H. Schubert and R. Zehbe, Ultramicroscopy, 2011, 111, (4), 259 LINK https://doi.org/10.1016/j.ultramic.2010.12.017
  12. 12.
    G. Kothleitner, M. J. Neish, N. R. Lugg, S. D. Findlay, W. Grogger, F. Hofer and L. J. Allen, Phys. Rev. Lett., 2014, 112, (8), 085501 LINK https://doi.org/10.1103/physrevlett.112.085501
  13. 13.
    M. Watanabe, M. Kanno and E. Okunishi, JEOL News, 2010, 45, (1), 8 LINK https://admin.jeol.com.cn/admin/static/uploadfiles/20170419/929268fb-cce7-4dc8-a572-271a98f51976.pdf
  14. 14.
    H. Tan, S. Turner, E. Yücelen, J. Verbeeck and G. Van Tendeloo, Phys. Rev. Lett., 2011, 107, (10), 107602 LINK https://doi.org/10.1103/physrevlett.107.107602
  15. 15.
    A. Choudhury, Arch. Comput. Methods Eng., 2021, 28, (5), 3361 LINK https://doi.org/10.1007/s11831-020-09503-4
  16. 16.
    P. Iassonov, T. Gebrenegus and M. Tuller, Water Resour. Res., 2009, 45, (9), W09415 LINK https://doi.org/10.1029/2009wr008087
  17. 17.
    A. Ioannidou, E. Chatzilari, S. Nikolopoulos and I. Kompatsiaris, ACM Comput. Surv., 2018, 50, (2), 20 LINK https://doi.org/10.1145/3042064
  18. 18.
    M. Čalkovský, E. Müller, M. Meffert, N. Firman, F. Mayer, M. Wegener and D. Gerthsen, Mater. Charact., 2021, 171, 110806 LINK https://doi.org/10.1016/j.matchar.2020.110806
  19. 19.
    ‘Importing a CZMODEL into ZEN blue or ZEN core’, Carl Zeiss AG, Jena, Germany, GitHub, 20th March, 2020 LINK https://github.com/zeiss-microscopy/OAD/tree/f47a5970fb710ba8066444b8c71f5781755fca64/Machine_Learning#importing-a-czmodel-into-zen-blue-or-zen-core
  20. 20.
    I. Arganda-Carreras, V. Kaynig, C. Rueden, K. W. Eliceiri, J. Schindelin, A. Cardona and H. S. Seung, Bioinformatics, 2017, 33, (15), 2424 LINK https://doi.org/10.1093/bioinformatics/btx180
  21. 21.
    S. Berg, D. Kutra, T. Kroeger, C. N. Straehle, B. X. Kausler, C. Haubold, M. Schiegg, J. Ales, T. Beier, M. Rudy, K. Eren, J. I. Cervantes, B. Xu, F. Beuttenmueller, A. Wolny, C. Zhang, U. Koethe, F. A. Hamprecht and A. Kreshuk, Nat. Methods, 2019, 16, (12), 1226 LINK https://doi.org/10.1038/s41592-019-0582-9
  22. 22.
    M. Andrew and B. Hornberger, Microsc. Microanal., 2018, 24, (S2), 118 LINK https://doi.org/10.1017/s1431927618012989
  23. 23.
    K. Simonyan and A. Zisserman, arXiv:1409.1556, 10th April, 2015 LINK https://doi.org/10.48550/arXiv.1409.1556
  24. 24.
    J. Ali, R. Khan, N. Ahmad and I. Maqsood, Int. J. of Comp. Sci. Iss., 2012, 9, (5), 272 LINK http://ijcsi.org/articles/Random-forests-and-decision-trees.php
  25. 25.
    T. Chen and C. Guestrin, ‘XGBoost: A Scalable Tree Boosting System’, in KDD ‘16: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August, 2016, pp. 785–794 LINK https://doi.org/10.1145/2939672.2939785
  26. 26.
    J. Lafferty, A. McCallum, and F. C. Pereira, ‘Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data’, ICML 2001, 28th June–1st July, 2001, San Francisco, USA, “Proceedings of the 18th International Conference on machine Learning 2001”, ed. C. E. Brodley, Morgan Kauffman Publishers, Burlington, USA, 2001, pp. 282–298 LINK https://dl.acm.org/doi/abs/10.5555/645530.655813
  27. 27.
    P. Tahmasebi, F. Javadpour and M. Sahimi, Transp. Porous Media, 2015, 110, (3), 521 LINK https://doi.org/10.1007/s11242-015-0570-1
  28. 28.
    S. Soyer, ‘czmodel 3.1.0: A Conversion Tool for TensorFlow or ONNX ANNs to CZANN’, Carl Zeiss AG, Jena, Germany, 26th April, 2022 LINK https://pypi.org/project/czmodel/
  29. 29.
    J. A. Taillon, C. Pellegrinelli, Y.-L. Huang, E. D. Wachsman and L. G. Salamanca-Riba, Ultramicroscopy, 2018, 184, (A), 24 LINK https://doi.org/10.1016/j.ultramic.2017.07.017
  30. 30.
    G. S. Armatas, Chem. Eng. Sci., 2006, 61, (14), 4662 LINK https://doi.org/10.1016/j.ces.2006.02.036
  31. 31.
    S. van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner, N. Yager, E. Gouillart, T. Yu and the scikit-image contributors, PeerJ, 2014, 2, e453 LINK https://doi.org/10.7717/peerj.453
  32. 32.
    K. Wu, E. Otoo, and A. Shoshani, ‘Optimizing Connected Component Labeling Algorithms’, Medical Imaging, 12th–17th February, 2005, San Diego, USA, “Proceedings Medical Imaging 2005: Image Processing”, eds. J. M. Fitzpatrick and J. M. Reinhardt, Vol. 5747, SPIE, Bellingham, USA, 2005, pp. 1965–1976 LINK https://doi.org/10.1117/12.596105
  33. 33.
    A. Hammoumi, M. Moreaud, E. Jolimaitre, T. Chevalier, A. Novikov, and M. Klotz, ‘Efficient Pore Network Extraction Method Based on the Distance Transform’, in “Artificial Intelligence and Industrial Applications: Artificial Intelligence Techniques for Cyber-Physical, Digital Twin Systems and Engineering Applications”, eds. T. Masrour, A. Cherrafi, I. El Hassani, Lecture Notes in Networks and Systems Book Series, Vol. 144, Springer Nature Switzerland AG, Cham, Switzerland, 2021, pp. 1–13 LINK https://doi.org/10.1007/978-3-030-53970-2_1
  34. 34.
    T. H. Loeber, B. Laegel, S. Wolff, S. Schuff, F. Balle, T. Beck, D. Eifler, J. H. Fitschen and G. Steidl, J. Vac. Sci. Technol. B, 2017, 35, (6), 06GK01 LINK https://doi.org/10.1116/1.4991638
  35. 35.
    J. Schwartz, Y. Jiang, Y. Wang, A. Aiello, P. Bhattacharya, H. Yuan, Z. Mi, N. Bassim and R. Hovden, Microsc. Microanal., 2019, 25, (3), 705 LINK https://doi.org/10.1017/s1431927619000254
  36. 36.
    M. B. Clennell, Geol. Soc. London: Spec. Publ., 1997, 122, (1), 299 LINK https://doi.org/10.1144/gsl.sp.1997.122.01.18
  37. 37.
    P. Moldrup, T. Olesen, T. Komatsu, P. Schjønning and D. E. Rolston, Soil Sci. Soc. Am. J., 2001, 65, (3), 613 LINK https://doi.org/10.2136/sssaj2001.653613x
  38. 38.
    J. Joos, T. Carraro, A. Weber and E. Ivers-Tiffée, J. Power Sources, 2011, 196, (17), 7302 LINK https://doi.org/10.1016/j.jpowsour.2010.10.006
  39. 39.
    R. Cárdenes, C. Alberola-López and J. Ruiz-Alzola, Image Vis. Comput., 2010, 28, (3), 307 LINK https://doi.org/10.1016/j.imavis.2009.05.013
  40. 40.
    D. Gostovic, N. J. Vito, K. A. O’Hara, K. S. Jones and E. D. Wachsman, J. Am. Ceram. Soc., 2011, 94, (2), 620 LINK https://doi.org/10.1111/j.1551-2916.2010.04111.x
  41. 41.
    J. Kašpar, R. Di Monte, P. Fornasiero, M. Graziani, H. Bradshaw and C. Norman, Top. Catal., 2001, 16, (1–4), 83 LINK https://doi.org/10.1023/a:1016682831177
  42. 42.
    R. D. Leapman, P. Rez and D. F. Mayers, J. Chem. Phys., 1980, 72, (2), 1232 LINK https://doi.org/10.1063/1.439184
  43. 43.
    C. C. Ahn and P. Rez, Ultramicroscopy, 1985, 17, (2), 105 LINK https://doi.org/10.1016/0304-3991(85)90003-8
  44. 44.
    S. P. Sutera and R. Skalak, Annu. Rev. Fluid Mech., 1993, 25, 1 LINK https://doi.org/10.1146/annurev.fl.25.010193.000245

‘ZEISS ZEN Intellesis for Image Segmentation in Microscopy: Use the Power of Deep Learning to Easily Segment Your Images’, Carl Zeiss AG, Jena, Germany LINK https://www.zeiss.com/microscopy/int/products/microscope-software/zen-intellesis-image-segmentation-by-deep-learning.html

‘Scout-and-Scan Instrument Control System’, Carl Zeiss AG, Jena, Germany LINK https://www.zeiss.com/microscopy/us/local/scout-and-scan.html

‘Software for CT and Microscopy Image Data Visualization and Analysis: Avizo Software: Materials Characterization and Quality Control Software for Imaging Data Analysis and Quality Assurance’, ThermoFisher Scientific, Waltham, USA LINK https://www.thermofisher.com/uk/en/home/industrial/electron-microscopy/electron-microscopy-instruments-workflow-solutions/3d-visualization-analysis-software.html

‘scikit-fmm 0.0.9 documentation’ LINK https://pythonhosted.org/scikit-fmm/

‘OpenFOAM and The OpenFOAM Foundation’, The OpenFOAM Foundation Ltd, London, UK LINK https://openfoam.org/

‘ZEISS Atlas 5: Master Your Multi-scale Challenge’, Carl Zeiss AG, Jena, Germany LINK https://www.zeiss.com/microscopy/int/products/microscope-software/atlas.html

Acknowledgements

The authors would like to thank Johnson Matthey colleagues for provision of samples and useful discussions for interpretation of the results.

The Authors

Aakash Varambhia is a Data Scientist in the Advanced Characterisation team at Johnson Matthey, UK. His work consists of developing specialised tools for data analysis from a wide range of instruments such as X-ray tomography, focused ion beam (FIB) and transmission electron microscopy (TEM). He is also an honorary researcher at Diamond Light Source, UK, and collaborates with the data science and microscopy teams at the facility. Before working for Johnson Matthey, Aakash completed a DPhil project at the University of Oxford, UK, in the Nellist research group where he developed quantitative experimental and data processing techniques to study catalyst nanoparticles.

Angela E. Goode received a MSci in Natural Sciences from the University of Cambridge, UK, in 2008. Following this she undertook a PhD, postdoctoral and research fellowship at Imperial College London, UK, employing correlative electron and X-ray microspectroscopies on nanoscale systems. She joined the Advanced Characterisation group at Johnson Matthey in 2018, where she specialises in FIB-assisted quantitative analysis.

Ryutaro Sato is Senior Scientist in the Johnson Matthey emission control research reaction engineering group. In 2013, he started working in Johnson Matthey, Japan as Product/Application Engineer after 11 years academic research on hydrogen storage materials. During gasoline particulate filter development for Japanese customers he began to take an interest in computational fluid dynamics (CFD) analysis, relocated to Johnson Matthey, UK, in 2019 to work on CFD simulation after the customer business win.

Trung Tran did his PhD study (2007–2010) at the University of Birmingham, UK, on synthesis, atomistic modelling and particularly TEM of bimetallic nanoparticles. From 2011–2013, he focused more on (scanning) TEM-based tomography of catalysts when working as a postdoctoral researcher at the Norwegian University of Science and Technology (NTNU). During 2013–2016, he was a visiting researcher at Stockholm University, Sweden, where he developed the software SUePDF for atomic pair distribution functions obtained from electron diffraction. Since June 2016, Trung joined the Electron Microscopy analytical group at Johnson Matthey where he does advanced characterisation with various electron microscopy techniques: electron energy loss spectroscopy, aberration-corrected atomic imaging, nano-tomography and electron diffraction-pair distribution functions.

Alissa Stratulat studied at Smith College in the USA, where her undergraduate thesis focused on the synthesis and characterisation of carbon compounds. She gained experience working with different materials during many projects in collaboration with Princeton University, USA, and Clemson University, USA. Then she continued her doctoral studies at the University of Oxford, UK. Prior to joining ZEISS Microscopy, Germany, as R&D Applications Developer – Advanced Materials, Alissa worked as a researcher in Honeywell’s research and development laboratory in Romania, where she developed and improved sensor materials. Currently she is working as a Program Manager for MIPAR Image Analysis, USA.

Markus Boese is an application specialist for the materials science marketing team at ZEISS Microscopy. He holds a PhD in Chemistry from the University of Bonn, Germany. For his postdoctoral research in electron microscopy he joined several research facilities: Ernst Ruska Center, Germany, Trinity College, Ireland and the National Center of Electron Microscopy, USA. At ZEISS Microscopy he started 2012 working in product management, supporting the development of the latest electron microscope products. His current activities at ZEISS Microscopy are focusing on nanomaterials and nanoscience relevant applications.

Gareth Hatton received his BSc in Archaeological Sciences at the University of Bradford, UK, in 2000. Subsequently he undertook a DPhil at the University of Oxford working on the analysis and replication of ancient vitreous materials. He joined the electron microscopy group at Johnson Matthey in 2005 where he specialises in the application of electron probe microanalysis and 3D analysis.

Dogan Ozkaya works as a Scientific Consultant and is in charge of the electron microscopy team in the Analytical Department at the Johnson Matthey Technology Centre, Sonning Common. He holds a PhD in Materials Science and Metallurgy from the University of Cambridge, UK. He carried out postdoctoral research in electron microscopy of various materials in several university departments, including the Cavendish Laboratory, University of Cambridge, and the Materials Department, University of Oxford, before joining Johnson Matthey in 2003.

Related articles

Erratum: Data-Driven Modelling of a Pelleting Process and Prediction of Pellet Physical Properties

Guest Editorial: Applications of Modelling at Johnson Matthey

Find an article

ArticleSearch