Domestic and commercial vehicles are leading sources of global pollution, with vehicle emissions risking the health of communities near roads (1). Fine and ultrafine particulate matter, oxides of nitrogen, hydrocarbons and carbon monoxide are key road traffic pollutants that are associated with adverse health effects (2). Catalytic converters have been used since the 1970s to reduce the emission of these pollutants by catalysing their reaction into less-toxic substances, typically carbon dioxide, nitrogen and water (3). However, current catalytic converters are not 100% efficient in their reactions of pollutants and moreover have variable efficiency at different operating temperatures.

This work uses machine learning modelling to analyse current catalytic converter performance and identify which future experimental tests would add most value to the ongoing development of improved catalytic converters. Previous work using machine learning in the catalysis domain has tended to focus on either augmenting quantum mechanical models of catalyst function (4–8), screening potential new catalysts (7–11), or predicting properties from carefully-selected chemical descriptors of catalysts (6, 8, 12–14). In contrast, in this work we focus on modelling catalyst properties from the formulation ingredients and processing variables of the catalyst. The ingredients and processing conditions of samples are easily accessible during the development process, lowering the barrier to application of machine learning in active development projects. In the following section we discuss the project objectives, detail the machine learning methodology used and the results it delivers, before looking forward to potential future applications of machine learning for materials science in the automotive field and beyond.

We collated data on 612 catalytic converter test sets that have been manufactured and experimentally tested by Johnson Matthey as part of an ongoing catalyst development project. The data contained information on the formulation used for the catalysts, including amounts and properties of 34 ingredients; 10 test parameters describing the testing process for each catalyst; and 16 experimentally measured properties for each catalyst including target gas conversions and selectivities. These output properties consisted of eight sets of tests, with each test run at both a high (approx. 500°C) and low (approx. 225°C) temperature on different samples of the same catalyst formulation. Each experimental property was reported as a steady-state average over 50–100 s of gas stream.

Using this data, we aimed to build understanding of the performance of this class of catalyst, using a machine learning model trained on the data to extract information on which input features of the formulation and processing parameters have most impact on the performance. Using this model, we then designed catalysts that offer both high performance and also add value to the machine learning model, which once made and measured can be added to the training dataset to enable more accurate modelling of high performance catalysts.

To model the catalyst data we used the Alchemite^TM multi-target machine learning platform. This method is described in detail in the literature (15–17), but in brief consists of iteratively generating predictions for all data series, both input and output, and using these predictions to impute missing data on the input side, before the final iteration of predictions are reported as the predictions for the output series. This method is designed to handle sparse input data, as was found in this work where up to 10% of the catalysts were missing information on each of the input properties. As the method is multi-target, generating predictions for all output properties simultaneously, we trained a model to predict all 16 experimentally measured properties at once. Alchemite^TM also generates estimates of the uncertainty in each prediction, which is vital to prioritise suggestions for future experiments that are most likely to achieve specified objectives. To test the performance of the model, data on 61 catalysts (10% of the data) was randomly held back; the model was trained on data for the remaining 551 catalysts. Hyperparameters of the model were optimised using Bayesian Tree of Parzen Estimators via five-fold cross-validation within the training set only (17, 18).

To test the performance of the model we simultaneously predicted all 16 output properties for each of the 61 held-back catalysts and measured the coefficient of determination R², for each output property. The coefficient of determination is defined as Equation (i):

(i)

where i indexes each catalyst in the validation set; y_i are the true experimental values, with mean ȳ; and f_i are the model predictions. A value of 1 indicates a perfect fit between model and experimental values; a value of 0 indicates a fit that is no better than random chance; and negative values indicate predictions that are worse than random. The performance of the model is shown in light blue in Figure 1. The median R² across all the output properties is 0.71, indicating highly successful predictive accuracy. In Figure 1 we also compare to two robust standard machine learning approaches: support vector regression with radial basis function kernel and K nearest neighbours with 20 neighbours, implemented in scikit-learn (19), which were trained on a mean-imputed version of the ingredient and test parameter data and achieve baseline median R² values of 0.52 and 0.49 respectively.

Fig. 1.

We observed that the predictions for Property 6, at both high and low temperatures, were poor: we identified that although changes in Property 6 are observable, a key physical mechanism directly influencing the value of Property 6 is driven by a chemical species not easily measurable by any analytical method and so is not fully captured in the dataset used to train the models. This explains the poor performance of the models in this aspect. The addition of (perhaps heuristic) descriptors to capture the physical mechanism may improve the modelling performance (14), but at the cost of increasing the barrier to usage of the method compared to taking only ingredients and processes as input.

Because the experimental tests on the catalysts are each repeated, run first at high temperature and then at low temperature, these results can be correlated so there is the possibility of increasing the efficiency of the testing process by using machine learning to replace one of the rounds of testing. To validate this, we trained a machine learning model that took as inputs the formulation ingredients and test parameters as well as the experimentally measured results on all eight tests at high temperature, and predicted the results of the eight tests at low temperature. This order (using high temperature results as input to predict low temperature results) was selected to align with the current testing methodology.

The improved performance by using the high temperature measurements to help predict the low temperature performance is exemplified in dark blue in Figure 1. For five of the eight experimental properties the accuracy significantly increased (increase in R² of more than 0.1), and for Properties 1, 2 and 3 the resulting accuracy, with R²>0.95, is effectively equivalent to the experimental uncertainty in the measurement. For these three properties in particular, machine learning predictions could reliably replace experimental measurements, offering a saving in the time and effort required to run the experimental tests on new catalysts. The three experimental properties that were not improved by using the high temperature measurements are all related to the same target gas’ conversion rates, although it is not clear why these properties are not improved by access to increased experimental data. These three experimental properties are less commercially important than Property 1, which is the property with most commercial relevance.

Now that we have confirmed the accuracy of the model we are well-positioned to extract actionable insights. Therefore, we first analyse the relationships that the model identified between inputs and outputs. To do so we examined which input features are used by the model when making predictions for each of the output properties, by evaluating the overall relative weights assigned to each input feature by the trained model, i.e. what fraction of the model prediction for each output is attributable to each input feature, on average across the whole model. This is calculated using the information gain attributable to each input feature (20). The results are summarised in Figure 2, separately for the model trained to predict both high and low temperature properties and the model trained to predict low temperature properties only. Averaging across each of the output properties, we find that for the high and low temperature model the test parameters and formulation ingredients are utilised in the proportion 0.59:1, and for the low temperature only model the test parameters, formulation ingredients and experimental high temperature measurements are utilised in the proportion 0.60:1:1.19. The consistent ratio of 0.6:1 in utilisation of the test parameters and formulation ingredients between the two models indicates that the high temperature experimental measurements (especially Properties 1, 2 and 3) are adding distinct information to the model that it was not capable of identifying from either the test parameters or formulation ingredients.

Fig. 2.

The key operational insight derived from this analysis was that although the formulation ingredients provide important information for the simultaneous modelling of the high and low temperature results, the variation in the test parameters also provides a key contribution. Historically the test parameters have been controlled within specification ranges but the impact of variation within these ranges has not been considered. These results show that the test parameters have an impact on the resulting properties and that control and understanding of these parameters improves the value of the data.

With increased understanding of the importance of the test parameters for measured catalyst performance, we used the machine learning model to design catalyst formulations. For performance targets, we focussed on the most commercially important property (Property 1), aiming to maximise its value at both high and low temperatures, and for that value to be stable with temperature. Although Property 1 is the most commercially important property, the values of the other properties are also required for product success.

As well as looking for the formulations that would be most likely to succeed against these performance targets (‘exploitation’ of the model) we also searched for formulations that, when measured, would increase the model’s understanding of the formulation landscape and so improve future rounds of predictive modelling and formulation design (‘exploration’ of the model), as well as a balanced mix of the two objectives. We used a Bayesian search of the formulation space using Tree of Parzen Estimators (18) built into the Alchemite^TM platform, taking as the cost function the probability of simultaneously achieving all the performance targets, including a contribution from the uncertainty in each formulation’s predicted performance calculated as standard errors across the Alchemite^TM platform’s internal ensemble of sub-models (21). This cost function is the commercially relevant metric to aim to propose successful and useful new formulations. Exploitation-focused suggestions prioritise formulations with high probability of success, while exploration-focused suggestions prioritise formulations whose predictions are currently uncertain and will also help improve predictions over a wide range of formulation space.

A two-dimensional Uniform Manifold Approximation and Projection (UMAP) embedding (22) of the formulations is shown in Figure 3. The dark blue points show the historic experimental results, with more opaque points having higher performance against Property 1 and more transparent points having lower performance. We observe that there are several clusters of dissimilar formulations that had previously been measured, but that most of the formulations were relatively similar and are clustered in the centre of the plot (this clustering analysis being a key strength of the UMAP approach). Figure 3 also shows the formulations proposed by the machine learning approach, labelled by whether they are focused on exploration, exploitation or a balanced mixture. We observe that, as expected, the exploitation-focused suggestions are clustered more tightly at the centre of the plot, demonstrating that they are attempting to exploit a class of formulations with a high probability (up to 60%) of achieving all of the design targets simultaneously. In contrast, the exploration-focused suggestions are more varied, focusing particularly on gaps in the existing coverage of the formulation space where additional information will improve the model. The balanced suggestions show aspects of both behaviours. A subset of the formulations suggested by the machine learning, including samples from the exploration, exploitation and balanced suggestions, are currently undergoing experimental validation.

Fig. 3.

In this work we have shown how machine learning analysis of catalyst formulations enables new insights into the factors that affect catalyst performance, including particularly that the test parameters more strongly impact the eventual performance than was initially anticipated: this will have operational significance for the future of this product development. We have also shown how the use of a machine learning platform, rather than a single predictive tool, can enable full design workflows, including prioritising exploration of the formulation space or exploitation of a model to achieve high product performance, accelerating the design process by enabling a holistic view of the formulation opportunities. Future progress in this project could focus on achieving multiple target properties simultaneously, beyond only Property 1, or utilising the accurate predictions of low temperature measurements based on experimental high temperature measurements to halve the amount of experimental effort required when screening new formulations.

The machine learning approach here is applicable beyond catalytic converters, including the design of metal alloys (15, 23), batteries (24), and pharmaceutical drugs (21). A machine learning platform that can carry out the full cycle of formulation development, handling sparse real-world experimental data, building predictive models and proposing and interpreting new formulation designs adds value in each of these areas, with reduced barrier to entry by working directly on the composition and processing variables immediately accessible to project scientists.

FAIR data principles have been held as the gold standard for ensuring data across the sciences and across individual institutions is generated and kept in as sustainable a way as possible (1). FAIR data principles unlock powerful ‘data lake’ workflows that allow for multiple interactions, machine learning and deep insight to be gained, adding value to already collected data (2). Reports and peer reviewed publications are needed to share knowledge with others at both an inter- and intra-institution level.

One nemesis to this approach is the use of proprietary software and proprietary data standards. It has been suggested that all research software should be free open source software (FOSS) and that closed source software should be the exception (3). The use of FOSS and open source hardware has been shown to offer flexibility and insight in a range of practical applications within chemical R&D (4–6).

A wealth of new software is available every year including productivity tools, document management, data analysis suites and code produced via individuals or research groups. One recent report showed that ~51,000 publications in the life sciences had 25,900 unique pieces of software cited (7). In addition to the wealth of new software offerings humans are keenly biased towards additive problem solving (8). Adding to an existing system rather than taking away in order to solve a problem is seen across sectors, job roles and in the digital tools used to enable science. An exemplar of this type of approach in software was seen with the introduction of the ribbon into Microsoft Office. Those more experienced with the software were more likely to be dissatisfied and impeded by the addition of the ribbon into the Office suite (9). Frustration stemming from unclear error messages, poor wording and lack of training lead to a loss of as much as 40% of a user’s time trying to solve software related issues (10).

As we train the next generation of scientists, and during the course of professional development, it is imperative that individuals reflect on and take control of the digital tools used to plan, conduct and share work. Frustration can be avoided if the tool being used is understood. Ideally, any skills learned during any part of an individual’s scientific career should be transferable. This is not possible if proprietary software solutions are used as there is no guarantee the software will be available at a new role either due to funding, dropped support or incompatibility with other systems.

One part of the solution to this, as demonstrated clearly by software projects like GNU/Linux (referred to herein as Linux), is the use of open source plain file formats like text files. Text files are human and computer readable, have demonstrable longevity and, crucially, are free and open source. Coupling this with tools that allow users to build, maintain and deploy their own solutions could resolve many of the frustrations seen with modern computer use.

Herein a demonstration of a workflow using a single tool, working with just text files, that can be used to radically change the workflow of a modern, flexible and agile scientist. The key benefits are increased productivity, return on investment, cost and environment, health and safety via improved ergonomics. In this viewpoint it will be demonstrated that such a solution exists and how it can be used in the context of corporate R&D.

Figure 1 shows two simplified workflows. Figure 1(a) shows the current state for many scientists. Each box in this flow represents a separate piece of software. These often have different shortcut keys, require many open programs and limit the user in terms of customisability and automatic flows. Each box may represent a different piece of software with separate associated upkeep costs, adding to both R&D expenditure and cost to monitor and ensure compliance with licenses. Figure 1(b) shows one possible solution where a single software solution replaces all the programs in a digital workflow. This workflow is possible with the open source and free program: Emacs.

Fig. 1.

Emacs is a fully programmable and extensible text editor. It is used widely in the IT and programming fields. Originally developed in the 1970s, the version used today (GNU Emacs – referred to herein as Emacs) was developed in the 1980s by Richard Stallman. It may seem retrograde that a decades old software solution can compete with newer offerings, but its longevity speaks to its utility. Emacs has been maintained and updated throughout this period with versions available across Windows, macOS and Linux.

Out of the box Emacs is a blank canvas. The decades of use mean that many contributors have written, maintained and updated a large number of packages that can be downloaded and used for free. These packages are completely user customisable and self-documenting. Emacs allows the user to employ these packages to build what is needed from the ground up. The below examples demonstrate how this approach can be used in a range of tasks in corporate R&D. This was built and personalised in-house with speed and ease of use being key. By building this tool from the ground up there is no bloat or incompatibilities that come with other, long lived, commercial solutions.

Figure 2(a) and 2(b) shows the software loaded in either its unmodified form or after the application of one of the many distributions, in this case Spacemacs. These distributions come preconfigured for ease of use and with many quality of life features. It is possible for a user to use one of these or to build their own version.

Fig. 2.

Because the below use cases can be achieved from within one piece of software, productivity and focus can be retained with the use of suite-wide shortcuts and hot keys. This reduces the possibility of fragmented work which can reduce productivity (11). Emacs is also fully controllable from the keyboard, again improving speed, productivity and ergonomics.

Org-mode is a major mode (a set of instructions for how certain files should be handled) for Emacs which was developed in 2003 by astrophysicist Carsten Dominik. Initially as a way to organise Dominik’s work, it has grown into a full suite. Allowing for everything from ‘todo’ task management to note taking and scientific manuscript preparation.

Importantly, it allows for a single document to contain data, working code and prose (12). Org mode has several minor modes (options that can be turned on or off) which can unlock advanced features impossible with other free or commercial solutions. These will be discussed in the following sections.

Data generation does not happen in a vacuum. A scientist’s work day includes ‘scientific overheads’ that can dramatically lower the time spent by an individual on the act of conducting high quality science (13). Indeed only ~40% of young researchers’ time in academia is spent on research, with the majority of the remaining time spent on writing and administration (14).

This is represented pictographically as the first set of software in Figure 1(a). This can be thought of as everything up to the act of experimentation along with all the administration tasks associated with modern knowledge work. Emails, meetings and conferences all add to the overhead workers face. The following section is not an exclusive list of what can be done but aims to demonstrate a few case studies of how Emacs can remove the burden of scientific overheads by consolidation of tasks with Emacs and Org-mode.

The act of producing, reviewing and executing a plan is an essential component of problem solving (15). Time management behaviours improve job satisfaction and health while negatively impacting stress (16). Org-mode allows for easy task management and planning from within the Emacs environment.

By setting up ‘Org-capture’, a package that works with Org-mode, todos can be captured and stored centrally from anywhere within Emacs. This makes capturing and recording tasks without interruption to flow trivial. Agendas and todo lists can be automatically populated from multiple sources (for example, reading list, meeting notes, project files). Importantly this approach works well with systems like ‘getting things done’ while staying flexible enough to allow for individual customisation (17). Examples of todo management as well as automatically generated agenda views can be found in Figure 3(a) and 3(b) respectively.

Fig. 3.

Additionally other tedious tasks can be automated. The use of tools like ‘Yasnippet’ allow for chunks of text to be stored and pasted into a document with only a few key presses. The production of meeting notes, for example, can be sped up by producing a template which can be imported. These can be exported via a .tex file and rendered into a PDF using LaTeX. This may seem arduous but, once set up, this is completely automated.

Macros can also be recorded and called when needed. If any task is done repeatedly then tools with Emacs can be used to automate that process. This reduction of overheads frees up a scientist to allow them to do what generates value for companies and academic institutions alike.

In a world where scientists are not just expected to produce data but be fully fledged knowledge workers, tools like this are invaluable. Their flexibility and utility can be tailored to the user’s workflow, enabling high productivity work to be conducted.

The act of collecting, reading and making notes on reference materials is a key aspect of scientific work. Importantly any possible solution to digitalise this should allow for citations to be placed within documents as well as easy access to referencing styles. This is possible with commercial solutions and even some open source options. Where an Emacs workflow outshines all is that the reference manager, note taking, citation tools and writing program are all one.

Packages like ‘Org-ref’ allow for import of PDFs from digital object identifiers (DOIs) allowing for fast import and conversion into a defined bibtex file (the plain text file used by LaTeX to generate citations). Notes can be accessed quickly using a package like ‘Interleave’ or ‘Org-noter’ which allows for automated note taking during the reading of a document, Figure 4.

Fig. 4.

Linking of notes and PDFs is extremely powerful and a rarity in the reference manager space. Due to the notes being in plain text they are also searchable unlike PDF highlighting or other, non-text or paper based, approaches.

One of the benefits of multidisciplinary teams is learning about best practices outside of one’s field. One concept that has taken hold in the computer science world is that of literate programming. Literate programming is the idea that written code should not just tell a computer what to do but that it is imperative that the code also informs a human about what is running (18).

This approach should be common to scientists. The aim of written reports, manuscripts and presentations is to display complex data and analysis in an easy to understand form for humans. The problem, as we approach more complex analysis, is that: (a) the analysis is split from the final report or manuscript which leads to loss of reproducibility; or (b) that the analysis is hidden in proprietary software that does not conform to FAIR principles nor the longevity principles a large corporate or academic institution may expect.

Org-mode, by utilising ‘Org-babel’, allows for chunks of code to be written and executed from within a single document. Variables can be extracted from these code blocks and then embedded in the text or fed into other code blocks. There are clear parallels between this type of approach and that of the IPython/Jupyter notebooks. These notebooks offer similar advantages in combining prose and code, allowing for reproducibility in data analytics. Both Emacs Org-mode and IPython/Jupyter notebooks offer parallelisation as a feature within the language. These notebooks do, however, suffer from the same issues described above as they form part of a fragmented software solution. As will be described below, they also lack the ability to embed analysis to a final manuscript.

Plotting can be done in the same way with direct output to a number of image formats that can, in turn, be embedded into the Org file. If one simply wants a way to record one’s work in an easy to follow format which is completely human readable then Org-mode makes that a simple task. Where the power of this approach becomes evident is when this is linked with manuscript or report production.

Org files are human readable with any text editor but Emacs unlocks many ways to quickly access the myriad of features not available outside Emacs. Importantly Org files can be exported in a range of formats including PDFs, markdown and open document formats. This manuscript was prepared as a Org file which was automatically processed into a .tex file and rendered into a PDF. Tools like ‘Writeroom-mode’ format documents to allow for a distraction-free writing experience, Figure 5.

Fig. 5.

When it comes to reports and manuscripts written in Emacs and Org-mode it is trivial to produce literate documents. Data and analysis can all be included within the manuscript which is also machine accessible. This works well with FAIR principles allowing for a human readable document to also act as metadata and a repository for computer readable data. To demonstrate this Figure 6(a) is a plot rendered by Python code embedded in this document. The values have been calculated from data within the file. The code snippet for this can be seen in Figure 6(b). If any changes are made to the analysis or the data, the plot is updated. This means that a single Org file can be provided and all data and analytics can be reproduced. It also makes the process of data analytics and report writing much easier. Any changes to the analysis will be updated in the text, either via plots or by embedded variables. This reduces the cognitive load associated with making requested changes, either during the peer review cycle or due to feedback from colleagues.

Fig. 6.

Previous reports have demonstrated how experimental data can be embedded into PDFs produced from Emacs allowing for a manuscript or report to contain all the data reported (19). The benefits of this are clear for both scientific integrity and rigour but also as a way to ensure a report or manuscript can be understood fully if an employee were to leave an institution, retaining the value of that work indefinitely.

Emacs has a reputation of being difficult to learn and this should not be ignored. Emacs has a learning curve however this can be as steep or as shallow as the user needs. Emacs distributions like ‘Spacemacs’ or ‘Doom Emacs’ allow for mnemonics key bindings and other quality of life features. Vanilla Emacs has many of the graphical user interface aspects you would expect, such as menus, which allows for most of the functionality to be explored. Becoming proficient takes time however this comes slowly as utility is unlocked. As summarised by John Kitchin:

“Scientific publishing is a career-long activity, and one should not shy away from learning a tool that can have an impact over this time scale.” (19)

While this still holds true, the author feels it is imperative to add the same is true of all aspects of a scientist’s workflow including productivity, reference management and data analytics.

Additionally, despite best efforts, all aspects of an Emacs workflow may not be possible. Email is possible within Emacs. However due to some institutions’ policies, such as Azure Information Protection, it may not be possible to set up due to issues with accessing confidential information without support from the host organisation. In this case it would not be possible to utilise such a tool. Similarly, while FOSS software allows for flexibility and the ability to create one’s own code, a user will be dependent on the software being correctly maintained. This lack of warranty is an inherent issue with FOSS. With repositories like GitHub (and similar), it is possible to access, fork and publish or maintain one’s own repositories for tools at a personal or institutional level, providing licensing conditions allow.

The maintenance overhead should not be underestimated, especially when considering issues with business continuity. However, this is not a new problem and, if the value is seen, institutions can add resource to deliver long lasting FOSS solutions. Parallels can be drawn to the development of the Linux kernel. Here private companies contribute extensively to the FOSS development because there is an understanding of the value of that project to their business interest (20).

While FOSS approaches offer great benefits, the use of proprietary or closed source software is preferable when that software offers utility not possible by other routes. Complex analysis using statistical software, complex peak fitting or databases requiring subscriptions are still a reality of the profession. When these tools are needed the approach outlined above still works providing the data can be exported from such a program into a plain text format. If this is not possible and FAIR principles cannot be upheld, the use of such a tool should be re-evaluated to determine if its use can facilitate long term and sustainable analysis.

Emacs is a powerful and versatile tool for modern science. It facilitates the production, handling and analysis of data in a FAIR fashion while allowing modern scientists to be as agile as possible. By using tools under one FOSS umbrella huge productivity gains can be realised along with improvements in ergonomics and associated cost benefits with the removal of proprietary software tools. The learning curve should be viewed in the context of a lifelong scientific career. With institutions understanding the value of data beyond a single scientist, applying (or supporting individuals who wish to apply) this type of workflow more widely would have a profound and long last effect beyond the career of just one scientist.

Johnson Matthey’s surface spectroscopy team focuses on providing essential information on the surface chemistry and composition of different materials for Johnson Matthey businesses and their customers. The team develops in situ, ex situ multi-technique surface analysis methods to deliver a more in-depth surface characterisation (6). Using laboratory-based X-ray photoelectron spectroscopy (XPS) as the main surface analysis technique, the team works on the applications of several complementary spectroscopic techniques such as ion scattering spectroscopy (ISS), reflection electron energy loss spectroscopy (REELS), ultraviolet photoelectron spectroscopy (UPS) and Raman.

The surface spectroscopy team is also involved in developing synchrotron-based near-ambient pressure (NAP)-XPS applications to study materials under reaction conditions. The team has been supporting fundamental surface science investigations and has sponsored several PhD projects that involved NAP-XPS characterisation of catalysts under reaction conditions. Two recent PhD projects with the University of Reading involved synchrotron-based NAP-XPS measurements to study supported platinum group metal (pgm) catalysts under methane oxidation reaction conditions in situ.

The first PhD project focused on investigating the chemical and compositional changes in alumina supported palladium catalysts with different particle sizes (4 nm to 10 nm) under reaction conditions similar to those used in the partial oxidation of methane to synthesis gas (syngas) (7). Surface adsorbates, palladium oxidation states and partial pressures of reactants and products were simultaneously tracked using mass spectrometry and NAP-XPS. NAP‐XPS data showed how the oxidation state of the palladium changes with increasing temperature (from Pd[0] to PdO and back to Pd[0]). NAP-XPS data analysis was further enhanced using mass spectrometry which showed an increase in carbon monoxide production over the Pd[II] oxide phase. In this study, a particle size effect was revealed for the catalysts demonstrating that methane conversion starts at lower temperatures with larger sized particles (Figure 1) (8).

Fig. 1

For palladium catalysts on different supports such as alumina, silica and a mixture of alumina and silica, NAP-XPS showed that on all the supports studied PdO is the dominant oxidation state and is the active site for complete methane oxidation which occurs at 500–600 K. As the oxygen is consumed and the temperature increases to >650 K, PdO is found to reduce to PdO_x, where 0 ≤ x < 1. Mass spectrometry showed a decrease in the partial pressures of complete methane oxidation products (carbon dioxide and water). Syngas formation (hydrogen and carbon monoxide), the product of partial methane oxidation, is dominant, suggesting reduced palladium is the active state for partial methane oxidation. The reactivity of alumina supported palladium materials is found to increase in the order: SiO₂ < SiO₂-Al₂O₃ < Al₂O₃ (Figure 2) (8).

Fig. 2

Another collaborative PhD project (Johnson Matthey; Diamond Light Source, UK; and the University of Reading) involved studying the effect of pgm composition and reaction conditions (dry and wet) on the catalytic behaviour of a range of alumina supported monometallic palladium and bimetallic palladium-platinum nanocatalysts under methane oxidation conditions. NAP-XPS and in situ mass spectrometry were combined to correlate the product formation and the chemical state of the catalyst throughout the temperature ramps under methane and oxygen gas mixture at elevated temperatures under dry and wet conditions (Figure 3). NAP-XPS was used to study the chemical states of monometallic palladium and bimetallic palladium-platinum nanocatalysts, demonstrating that there is a clear link between platinum presence, palladium oxidation and catalyst activity under stoichiometric reaction conditions. Under oxygen-rich conditions this behaviour is found to be less clear, as all of the palladium tends to be oxidised, but there are still benefits to the addition of platinum in place of palladium for complete oxidation of methane (9).

Fig. 3