One of the best ways to reduce smog and associated pollution from vehicle exhausts is to use autocatalysts in use since the 1970s.1 The majority of cities are now cleaner places to live as a direct result of this reduction.
The auto catalyst achieves this by lowering emissions from vehicles, including carbon monoxide, nitrous oxides, and soot, to legislated levels.
A porous ceramic monolith wall with multiple channels and a honeycombed structure frequently coats the catalyst channels. While the catalyst coating (typically porous) is made of alumina and/or cerium zirconia supported with platinum group metal (PGM) nanoparticles, the honeycomb structure enhances the surface area contact with the gas.
While there are various other catalyst and support configurations, a research-grade “gasoline particulate filter” (GPF) is discussed in this article. For this GPF, the pores within the walls, along with the surface of the monolith channels, are coated with the catalyst.
Two factors affect how well the GPF performs:
- Gas transportation to the catalyst’s active site via channels and connecting pores
- To guarantee that there are sufficient active sites available for the desired performance selectivity to reduce pollution, the morphology of the catalyst nanoparticles must be considered
During operation, both criteria occur at the macro, meso, and nanoscales. X-Ray Microscopy (XRM) and Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) are needed to investigate the performance criteria listed.
The advantage of the XRM is that it can observe the characteristics of the monolith at a large scale. This method can identify characteristics such as coating uniformity, location, and pore network distribution. The main factors to examine are pore size distribution and catalyst coating thickness.
The advantage of the FIB-SEM over the XRM’s resolution limit (~0.5 µm)2 is its ability to see smaller features within the catalyst coating, also known as the washcoat. The distribution of elements within the washcoat, pore size distributions, and percentage changes of different elemental species within the washcoat are the main parameters that need to be analyzed.
Multiple two-dimensional transmission images of an object from various orientations are examined in the XRM. Slices can be made from any direction of the sample after reconstruction.
This is demonstrated in Figure 1, where a ZEISS Xradia 520 Versa was used to scan a single channel from a four-channel GPF monolith.
Figure 1. XRM scan from a ZEISS Xradia 520 Versa, recorded at a voxel size of 4.74 µm, showing 4 channels from a monolith. A single channel was selected for analysis, the slice view directions along x, y, and z provide an overview of the washcoat coating along the channel. Image Credit: Carl Zeiss Raw Materials
By selecting a more focused area of interest from the XRM analysis and using the ZEISS Atlas software for cross-correlation, analysis in the FIB-SEM is carried out. The experimental data was captured on the ZEISS Crossbeam 550 with automated slice and view and sequential elemental mapping capability.3
By carefully milling the sample using the ion beam and imaging the polished cross-section up to a predetermined sample depth, representative areas can be studied correlatively to comprehend regions of interest, such as the inner washcoat to monolith interface.
Analysis Methods and Discussion
It often takes several observations, measurements, and iterative refinements to link performance to rational catalyst design. Due to the high turnover in the industry, this must be completed quickly. This agile approach necessitates quick correlative characterization of materials using a methodology supported by automation and machine learning.
The most typical workflow for a monolith sample involves segmenting features such as the substrate, catalyst, and pores from images like Figure 1 and then performing measurements on these features.
Both workflow components have their share of difficulties, but the first has emerged as one of the biggest in the field of microscopy.
The most important step in the workflow is segmentation, as it establishes the basis for all subsequent quantitative analyses. Deep learning methods have recently been demonstrated to be very effective in several computer vision tasks, particularly where more traditional methods (such as thresholding) have previously failed.4
The machine learning image segmentation software from ZEISS ZEN Intellesis offers one such option right out of the box. This method “paints” various pixels in the image connected to various classes using a deep learning model trained on pixels characterized by the user or microscope.
The main benefit of labeling pixels in this way, as demonstrated by Figure 2, is that the microscopist’s experience is taken into account. The pixels are vectorized as inputs for training the neural network after associated classes (such as catalyst, substrate, etc.) have been labeled, while the named classes are offered as training outputs.
Figure 2. Intellesis trainable segmentation and the subsequent image processing workflow. Once segmentation was performed, the segmented labels were cleaned and masked to obtain the pore labels. After propagating the image processing to all image slices using software developed at Johnson Matthey, measurements such as percentage change and mean sample decomposition can be calculated. Image Credit: Carl Zeiss Raw Materials
XRM Analysis and Discussion
As shown in Figure 2, the segmentation method’s classification approach makes it impossible to distinguish pores from empty channel space for this sample. However, a mask of pore segments can be produced by combining label masks and cleaning edge features.
Using proprietary software created by Johnson Matthey, further image processing, including segmentation, was carried out on all the image slices.
Catalyst, substrate, and pores are each represented by a different color in the segmented images of Figure 2. Following segmentation, the remaining analysis, such as calculating the percentage change along a monolith channel, is simple.
Calculations can also be made for analysis, such as pore size distributions, pore volumes, and catalyst wall thickness. Such quantification can be incorporated as inputs into gas flow simulations to link pressure drop to efficiency.
Figure 2 illustrates the catalyst coverage of the monolith along the z-direction. It is possible to see a uniform distribution of substrate, pores, and catalyst along the channel, which shows careful process engineering management. The x and y directions also enable similar observations.
Figure 2 depicts how the sample can be dissected into a Sankey plot for a mean three-dimensional quantification of the sample. The sample is broken down into parts such as catalyst, pores, and substrate in the mean sample decomposition plot. These are further divided along the x, y, and z axes. Once more, consistent coverage can be seen in the mean sample decomposition along the different axial directions.
By modifying the processing strategy and fine-tuning the monolith's performance, the performance can be precisely controlled due to the coverage's consistency and accuracy. Such a workflow can be improved to create a logical GPF design with the help of iterative testing, rapid analysis, and simulation.
After the segmentation was completed, the segmented labels were cleaned and masked to produce the pore labels. Using software created by Johnson Matthey, measurements such as percentage change and mean sample decomposition can be computed after the image processing has been propagated to all image slices.
FIB-SEM Analysis and Discussion
It is necessary to use a FIB-SEM to observe the sample to see the monolith’s micro- and nano-scale features. Segmentation is the main issue in this case.
A representative area of interest chosen for FIB-SEM analysis is shown in Figure 3. Using the Atlas software, a top-down scan of the same region in the XRM scan is superimposed on the FIB-SEM image.
Figure 3. Intellesis trainable segmentation and the subsequent image processing workflow. Once segmentation was performed, the segmented labels were cleaned and masked to obtain the pore labels. After propagating the image processing to all image slices using software developed at Johnson Matthey, measurements such as percentage change and mean sample decomposition can be calculated. Image Credit: Carl Zeiss Raw Materials
A large field of view area of the washcoat was chosen from this area of interest, as seen in Figure 3. The area was cut with the ion beam in 20 nm steps along the z-axis. The image stack collected along the z-direction was aligned for segmentation and analysis after finishing the FIB tomography routine.
“Curtaining,” or vertical lines that can be seen on the polished cross-section due to variations in sputtering rates across the sample, is one of the milling artifacts frequently present in inhomogeneous samples. The curtaining effect severely hampers any segmentation procedure.
Before Intellesis segmentation, a curtaining removal and median filtering routine developed at Johnson Matthey was used to streamline the process.
The pink areas of the washcoat in the segmented image correspond to the ceria-zirconia, the green areas to the alumina, and all other colors of the remaining monolith. After further analysis, the image’s composition is revealed to be 76% ceria zirconia and 24% alumina.
After the workflow was established, the analysis was applied to all the image slices using the spectroscopic elemental maps produced using energy-dispersive X-Ray mapping. The results at the nanoscale are also examined in the subsequent publication.
Conclusions
One of the best methods for lowering air pollution is using autocatalysts, which filter harmful gases from vehicle exhaust. A multiscale study from the macro to the nanoscale is necessary to understand them and enhance their performance.
However, accurate measurements, automation, and agile engineering are necessary to receive useful results. This is possible with the ZEISS Xradia and Crossbeam, as well as pre-built deep learning techniques and Johnson Matthey’s domain knowledge.
References
- Twigg. M. (1999) Twenty-Five Years of Autocatalysts. Platinum Metals Rev., 43, (4), 168-171
- https://www.zeiss.com/content/dam/Microscopy/us/download/pdf/Products/xradia520versa/xradia-520-versa-product-information.pdf
- https://blogs.zeiss.com/microscopy/news/en/zeiss-crossbeam-550-sets-new-standards-in-3d-analytics-and-sample-preparation
- Rachmadi et al. (2017) Deep Learning vs. Conventional Machine Learning: Pilot Study of WMH Segmentation in Brain MRI with Absence or Mild Vascular Pathology. J. Imaging., 3, 66
This information has been sourced, reviewed and adapted from materials provided by Carl Zeiss Raw Materials.
For more information on this source, please visit Carl Zeiss Raw Materials.