Nanoscale tomography reveals the deactivation of automotive copper-exchanged zeolite catalysts

Copper-exchanged zeolite chabazite (Cu-SSZ-13) was recently commercialized for the selective catalytic reduction of NOX with ammonia in vehicle emissions as it exhibits superior reaction performance and stability compared to all other catalysts, notably Cu-ZSM-5. Herein, the 3D distributions of Cu as well as framework elements (Al, O, Si) in both fresh and aged Cu-SSZ-13 and Cu-ZSM-5 are determined with nanometer resolution using atom probe tomography (APT), and correlated with catalytic activity and other characterizations. Both fresh catalysts contain a heterogeneous Cu distribution, which is only identified due to the single atom sensitivity of APT. After the industry standard 135,000 mile simulation, Cu-SSZ-13 shows Cu and Al clustering, whereas Cu-ZSM-5 is characterized by severe Cu and Al aggregation into a copper aluminate phase (CuAl2O4 spinel). The application of APT as a sensitive and local characterization method provides identification of nanometer scale heterogeneities that lead to catalytic activity and material deactivation.


UV-Vis-NIR DRS spectra
The spectra of the Cu-exchanged zeolite samples are in Supplementary Figure 3 and the spectra of reference materials are in Supplementary Figure 4. All hydrated, Cu-exchanged zeolite samples contain two sharp features at 7140 and 5260 cm -1 , which are due the overtone and combination bands of physisorbed water 1 . These bands disappear upon dehydration, demonstrating that the samples are well-dehydrated after calcination. In hydrated Cu-SSZ-13 the adsorption bands near 12,500 cm -1 correspond to d-d transitions in Cu 2+ species. These species are also present in hydrated Cu-ZSM-5 and in dehydrated Cu-ZSM-5, although with a slight blue shift 1-3 . In dehydrated Cu-SSZ-13 it is well known that these d-d transitions split into four bands as removal of the water allows Cu 2+ to interact more strongly with the framework and be stabilized in a more distorted local environment. It is known that this only occurs in zeolite SSZ-13 due to its highly distorted environment, and not in zeolite ZSM-5 4 . Upon aging, there is a change in the d-d transition region in zeolite Cu-SSZ-13 and we believe this is due to Cu migration to different framework positions. In all hydrated Cu-exchanged zeolite samples there is a ligand to metal charge transfer (LMCT) band near 48,000 cm -1 associated with (O 2-Cu 2+ O -Cu + ). In dehydrated Cu-exchanged zeolite samples it is known that there is a red shift in the LMCT band due to changes in the local environment of Cu 2+ , which can also occur with aging, and in zeolite Cu-SSZ-13 the main shift is reported at 41,500 cm -1 with shoulders at 32,000 and 29,000 cm -1 , while in zeolite Cu-ZSM-5 there is a red shift to 41,000 cm -1 1 . Spectroscopic features associated with the presence of CuO species appear in the 35,000-40,0000 cm -1 range, and are present in aged zeolite Cu-ZSM-5, and to a lesser extent in zeolite Cu-SSZ-13 2,3,5 . Reference spectra of CuO and CuAl2O4 spinel are included in Supplementary Figure 4. In the dehydrated, oxygen activated Cu-exchanged zeolite samples [Cu-O-Cu] 2+ dimers are reported at 22,700 cm -1 , but we do not see this in our samples, excluding their presence 1 .
It is important to mention here that the reference spectra of CuAl2O4 spinel and CuO were taken after mixing each species to 6 weight % with Na-form ZSM-5 (CBV2314 was exchanged to Na-form using 1 M NaCl solution, full exchange was followed by pH and the zeolite was subsequently calcined) to dilute the species and prevent detector saturation. The sodium form of the zeolite was used to prevent Cu exchange. CuAl2O4 spinel was prepared following reference [6], while CuO was prepared by heating Cu(OH)2 in basic solution to ~80°C until a black solid was obtained. The difference spectra are also presented to remove the influence of the zeolite. Wavenumber (cm -1 ) 6

APT data analysis
All error analyses were calculated from counting statistics using the method described in reference [8]. In all subsequent discussions, consistent with the terminology used in APT work, 'bulk' refers to the entire needle, 'matrix' refers to the atoms that remain after clusters are removed, and solute refers to the element(s) of interest in cluster analysis. See reference [9] for further details and definitions. In all APT analyses, a small amount of gallium was found to be present due to the FIB-cutting process, and our treatment of this potential influence is discussed in the subsection "Assessing Ga damage." All compositional percentages referred to in this work are atomic percentages. The non-conductive nature of the materials used in this study complicates the data collection. The pulsed laser heating helps to overcome these difficulties. Sample heating creates issues with processing the sample mass spectra as it creates thermal tails in the data. Due to this issue, not all collected data sets gave reliable, quantitative results, and only data sets that gave reliable results are presented. It is known that there are difficulties in quantitatively detecting oxygen or oxides as these tend to field evaporate as double detector hits, which will be registered as a single hit, and there is also evidence that molecular ions can dissociate after field evaporation, such that one ion becomes deionized and does not hit the detector within the appropriate time-of-flight window and is registered as background (even though we attempted to keep the detection rate low, 1 ion per 200 pulses) [9][10][11] . These issues can cause deviations from the actual composition, and hard to control experimental parameters can influence the severity of these artefacts. Therefore, we have concluded that APT has difficulty in quantifying the Al content in zeolite materials and we emphasized trends (isosurfaces, clustering, radial distribution function, etc.) rather than comparing absolute numbers. The Local Range-Assisted background estimate embedded in IVAS was used, which estimates the background based on the number of counts before and after the ranged peak.

TOF-MS analysis
Data from the APT TOF-MS for each material is shown below (samples 1, 3, 4 and 6) in Supplementary Figure 7 to Supplementary Figure 14 (bin width 0.03 Da), along with the peak assignments. For each sample two different spectra are included to demonstrate the change after exclusion of the Ga rich region from the data analysis. All peaks were ranged in the same location between samples, but the width varied based on the full width at half maximum (FWHM) of the peak and the thermal tails. The peaks were ranged to approximately two times the background unless another peak was present before this occurred. SEM images of samples 1, 3, 4 and 6 prior to the APT experiment are shown in Supplementary Figure 15 to illustrate differences in the tip shapes which may account for differences in the TOF-MS data.    prior to the APT experiment to illustrate the differences between the tip shapes that may account for differences in the mass spectra.

Isoconcentration surface analysis
Isoconcentration surface analysis is extensively discussed in reference [9] and the references contained therein. The isoconcentration surface analysis is initiated by creating a 3D grid in concentration space. After this grid is created, 3D surfaces connecting points of equal elemental concentration can be generated, so called isoconcentration surfaces. These surfaces allow internal features to be identified, and may be defined by a single element or multiple elements. The 3D grid is created by defining a voxel size and delocalization, with the tradeoff that smaller voxels will give a higher spatial resolution but a greater amount of noise. Once the 3D grids are generated, relevant isoconcentration surfaces can be found using concentration thresholds. The importance of the isoconcentration surface analysis comes in evaluating the resulting proximity histograms, which allow the border between concentration regions to be quantitatively evaluated. The exact boundary value chosen is not important as it can change with voxel size and delocalization, what is important is that the proximity histograms allow regions that are poor or rich in specific ion(s) to be identified and separated. We generally used the default settings in IVAS (3x3x3 nm delocalization and 1 nm voxel size). However, there were instances in which the voxel size was reduced to create a higher polygon density to increase the signal to noise ratio because the edge polygons are not included in the proximity histogram analysis. The delocalization remained the same, so the shape of the interfaces was not significantly altered with the reduced voxel size.

Assessing Ga damage
It is well known in APT that Ga damage can be problematic as it may influence the position of ions, making any findings the result of needle fabrication rather than an actual material property, potentially leading to incorrect conclusions. In order to minimize this during zeolite sample preparation, we coated the samples with Pt using ebeam deposition. A 30 kV beam was used for the initial milling, leaving an ~250 nm wide needle with several hundred nm's of Pt remaining on the surface. A 2 kV final milling step was used to remove the Pt cap. The zeolite material mills much faster than the Pt cap, and therefore, the tip shape was finally a very sharp needle with a smaller shank angle. We were also careful to minimize the ebeam exposure to the material because zeolites also suffer from damage resulting from the ebeam. Potential Ga damage has been an issue that we were also concerned with when analyzing the data, so we used a Ga isoconcentration surface to remove the Ga rich portions of the needle (normally the needle tip and surface) prior to any analyses by exporting the low gradient side of the Ga interface. The figures show the full datasets for aesthetic purposes. To demonstrate that Ga was not the cause of the major material features we observe, we show compositional histograms for all isosurfaces included in the figures of the main manuscript, but plotted in atomic % and with the Ga concentration included (Supplementary Figure 16), and the RDFs for needle 4 (Supplementary Figure 17.) as it did not have any isoconcentration surfaces. As these analyses show, there was no significant increase in Ga corresponding with Al and/or Cu agglomeration, demonstrating that these material features are not created during FIB milling. Figure 16. Compositional histograms including Ga concentration and plotted in atomic % for all isoconcentration surfaces discussed in the manuscript. a, 1.8% Cu isoconcentration surface in needle 1. b, 5% Cu isoconcentration surface in needle 3. c, 8% Al isoconcentration surface in needle 3. d, 1.4% Cu isoconcentration surface in needle 6. All error bars were calculated from counting statistics using the method described in reference [8].  4) indicating there is no affinity between Ga and Al or Cu. All error bars were calculated from counting statistics using the method described in reference [8].

Cluster analysis
The purpose of cluster analysis is to identify regions where the spacing between solute atoms in APT data is smaller than that in the bulk, effectively determining regions that are locally enriched in a specified element. For a detailed discussion of the procedure the reader is referred to reference [9]. As a summary, cluster analysis is initialized by choosing a Nmin value, which is the minimum number of solute atoms that can form a cluster. With Nmin set, different values of Dmax can be evaluated. Dmax defines the maximum distance in which another solute atom must be found to form a cluster (one additional solute atom must be found for order = 1, higher orders are also possible with more than one solute atom required to fall within Dmax). A minimum value of Dmax is established such that clusters containing at least Nmin atoms are identified. There are additional parameters that are used in the analysis, and matrix atoms contained in the cluster can also be identified. The key element of cluster analysis is how the parameters are determined and significant clustering is identified. The significance test is accomplished by first taking the number of solute atoms and volume being considered, and randomizing their position using a normal distribution in space. Then, the same cluster analysis is carried out on the randomized data. Once these two data sets are established, the cluster parameters are determined using an iterative process where Nmin is first set and then the cluster count is plotted as a function of Dmax. If the solute data is found to be significantly separated from the randomized data (normally chosen as a point where clusters exist in the collected data and zero, or a very minimal amount, such that >95% of the clusters would be non-random, exist in the randomized data), then Dmax can be fixed. A fixed Dmax can then be used to plot cluster count versus cluster size to set an optimal Nmin. By iterating through these parameters, the optimal cluster Dmax and Nmin can be determined, where significant clusters can be found relative to the randomized data. This iterative process as applied through the cluster count distribution (CCD) and cluster size distribution (CSD) as applied to the data sets in this manuscript are shown with the analyses for each individual parameter along with the other parameters determined for each specific cluster analysis. A simple method that is useful in identifying if clusters are present and a cluster analysis should be initiated, is by directly comparing the nearest neighbor distributions (NNDs) of the solute and randomized data 12 . In doing this, the number of counts is plotted as a function of the atom pair distance. For the randomized data this will, by definition, give a normal distribution. However, if clustering is present, then the solute NNDs will be shifted to a smaller atom pair distance. The NND for the solute data is then described best by at least two Gaussians, one centered at a smaller atom pair distance, representing the clusters, and a second centered at an atom pair distance greater than the normal distribution of the randomized data, representing the matrix ion spacing. This makes plotting the NNDs an invaluable tool in cluster analysis.
In this work, when performing the Al cluster analysis, we used all Al related peaks (Al and AlO). For Cu, there were only metallic Cu peaks and all of those were used in the Cu cluster analysis.

Radial distribution function
The radial distribution function (RDF) is a powerful tool to examine affinity between small numbers of atoms and test homogeneity, and it was applied using CAMECA's IVAS software. The method is extensively discussed in references [13] and [14]. The RDF is conducted by normalizing the local concentration of a selected ion by the bulk concentration, and this is done radially outward from the center of the ion(s) of interest. Error analyses were conducted using counting statistics; because few ions will be counted near each ion of interest, the error bars are high for the first few data points. As the RDF is shown as a bulk normalized concentration, values between pairs of atoms will not be equal. For example, Cu-Al is not equal to Al-Cu. The number of counts is equal for both pairs, but not the bulk normalized concentrations since they are defined over different volumes. This is illustrated in 2D below in Supplementary Figure 18. In our analyses, all molecular ions were decomposed prior to performing the RDF.
Supplementary Figure 18. 2D schematic of the radial distribution function (RDF) for analyzing the spatial distribution of different elements in the atom probe tomography (APT) data.

APT data analysis workflow
The processing and analysis of the data obtained in an APT experiment is far from trivial and is the subject of active research and debate 9,15-17 . Below is a brief outline of the normal workflow used in the processing of APT data as performed for the experiments described herein. This is not meant to be an exhaustive or definitive guide, but rather a brief overview in order to give the reader an idea of the process used, and the reader is referred to our previous manuscript for further details 10 . 1. Data collection with an APT instrument. 2. Data processing and reconstruction: When processing zeolite samples it is especially important that all peaks are ranged correctly. The 3D distribution of all ions can then be generated. 3. Identification of isoconcentration surfaces: Isoconcentration surface analysis is used to separate large compositional inhomogenities. The isoconcentration surfaces are quantified using proximity histograms. 4. Cluster identification: Within subvolumes (if any exist, isolated using isoconcentration surface analysis), both the nearest neighbor distributions (NNDs) and frequency distribution analysis (FDA) are used to examine compositional heterogeneities that are not as obvious as isoconcentration surfaces. The identification of clusters is an iterative process between comparing the cluster count distribution (CCD) and cluster size distribution (CSD) so that statistically relevant clusters can be identified. 5. The radial distribution function can be used to examine compositional heterogeneities that exist that may not have been identified using isoconcentration surface analysis or cluster analysis. Fresh Cu-ZSM-5 (Sample 1). Bounding box 140x43x40 nm 3 .

Individual analyses
Supplementary Figure 22. Analysis of fresh Cu-ZSM-5, sample 1: a, Al nearest neighbor distribution (NND, order = 10). b, Al radial distribution function (RDF). c, Cu NND. d, Cu RDF. e, 1.8% Cu isoconcentration surface shown with all Al (blue) and Cu (red) ions. Bounding box dimensions of 140x43x40 nm 3 . This image is also included as Supplementary Movie 1. f, Bulk normalized compositional histograms across the 1.8% Cu isoconcentration surface. All error bars were calculated from counting statistics using the method described in reference [8].    Bounding box dimensions of 43x15x15 nm 3 . These images are also included as Supplementary Movie 3. f, Bulk normalized compositional histograms across the 8% Al isoconcentration surface. g, Bulk normalized compositional histograms across the 5% Cu isoconcentration surface. The percentage of Cu and Al atoms within the region with Cu/Al = 2 stoichiometry (in the 8% Al isoconcentration surface) were considered to give an estimate of the portion of Cu and Al atoms forming the phase that resembles CuAl2O4.This region was determined to contain ~20 % of all Cu ions and ~35% of all Al ions present in the needle. All error bars were calculated from counting statistics using the method described in reference [8]. Values agree with the stoichiometry of Cu aluminate spinel, CuAl2O4, which is indicated on each graph by the dotted line. All error bars were calculated from counting statistics using the method described in reference [8].   . f, Compositional histograms across the 1.4% Cu isoconcentration surface g, Bulk versus cluster composition. All error bars were calculated from counting statistics using the method described in reference [8].   These images are also included as Supplementary Movie 10. f, Compositional histograms across the 4% Al isoconcentration surface. g, Compositional histograms across the 1.3% Cu isoconcentration surface. All error bars were calculated from counting statistics using the method described in reference [8]. Additional sample of aged Cu-ZSM-5 Data was successfully collected for a second sample of aged Cu-ZSM-5. However, during the analysis of this material it was apparent that it had significant Ga damage, so we chose not to present this data in the main manuscript. The analysis of the material is shown below and demonstrates that the data collected from the other specimen of aged Cu-ZSM-5 is indeed reproducible, mainly the highly heterogeneous Cu and Al distributions. Ion maps for Al, Cu and Ga are shown in Supplementary Figure 35 and it is evident that the tip of the needle contains a region high in Al and Cu, similar to what was found in needle 3. However, the tip of the needle contains a significant amount of Ga, which was found in all needles due to the final FIB milling step. Since it was impossible to exclude the influence of Ga in creating the region rich in Al and Cu, the Ga rich region was removed as discussed in the "Assessing Ga damage" section. In the case of this needle, this volume also contained the Al and Cu rich regions. Based on the discussions in the "Assessing Ga damage" section the Cu and Al agglomeration would not be caused by the FIB milling, but since we did not feel comfortable analyzing the Ga rich volume we chose to exclude this data set. The two views in Supplementary Figure 35, along with the 2D ion maps, show that these regions are close they are not identical. However, even after removal of the Ga rich region, the RDFs for Al and Cu (Supplementary Figure 36) still show significant agglomerations between these species, consistent with the results from needle 3.