To meet government-mandated sustainability goals and greenhouse gas emission reductions, companies are being pressed to improve the efficiency of their production processes and shift away from petroleum-based fuels and feedstocks. Catalytic processes can play a key role in these efforts, but further improvements in efficiency will be necessary to meet greenhouse gas emission reduction targets.¹

As CO2 emissions are expected to continue to increase for the foreseeable future, it will become critical to use atmospheric CO2 as a feedstock for synthesizing non-petroleum-derived fuels and hydrocarbons such as HCOOH, CH4, and C2H4.1 The CO2 reduction reaction (CO2RR) can accomplish this, but it uses catalysts that contain toxic heavy metals or scarce precious metals.² Predicting alternative catalyst structures requires an understanding of how factors such as compositional changes affect the electronic structure and internal geometric strain of catalysts.

Reevaluating Experimental Approaches for Catalyst Discovery

Catalyst optimization and discovery activities typically require costly and time-consuming empirical operando approaches to identify a catalyst with desirable selectivity and activity. Such experimentation is often accompanied by detailed structural characterization, such as reaction intermediate analysis using Raman spectroscopy or catalyst crystallinity characterization using X-ray diffraction. High-throughput combinatorial screening can accelerate this process, although there are limitations in terms of data quality and the cost to establish such automated workflows.After a performant catalyst is discovered, validating its performance and incorporating it into an existing industrial process is a complex task because it is difficult to predict how the catalyst will interact with various formulation components under real-world operating conditions. The introduction of a new catalyst may also require reactor redesigns, and the interconnectedness of industrial processes makes it difficult to predict how the addition of a new catalyst will affect downstream processes.³ Pilot plants, requiring months to years of testing and significant capital investments, are a required step between lab experimentation and full-scale implementation. Ideally, it would be possible to limit the number of lab and pilot scale experiments required to proceed to each subsequent phase in the catalyst development process.Therefore, there is a need for more efficient use of the often-limited experimental data to allow a more guided approach to material discovery and experiment selection.

Traditional Computational Approaches for Accelerating Catalyst Discovery Provide Incomplete Insights

Standard computational chemistry approaches (e.g., DFT, MD, KMC) can provide additional insights with regard to property predictions of catalyst candidates, but there may be thousands of combinations of parameters to consider (solvents, precatalysts, additives, temperatures, pressures, etc.), creating a computationally intractable scenario.⁴ Similarly, such computational approaches can be used to provide insights into key kinetic parameters to guide the development of catalysts, but they require a priori knowledge about a reaction mechanism to determine the reaction coordinates of each step in a catalytic reaction. Designing catalysts requires obtaining clear structure-property relationships, which in turn relies on understanding the local electronegativity, ionic potential, and electron affinity of active sites on catalyst surfaces. Computational approaches often fail to account for the overwhelming complexity and interplay between these factors, including how they affect surface/adsorbate combinations, adsorption sites, and adsorbate conformations. This hinders efforts to determine underlying catalytic mechanisms, which are necessary to provide a complete understanding of a reaction.⁵ Ideally, one would want an approach that combines traditional computational chemistry with real-world experimental results.

Machine Learning Approaches for Developing Catalytic Processes

In a departure from traditional computational approaches, machine learning (ML) can handle more complex scenarios with large numbers of variables and can be applied for catalyst screening and the prediction of process-related conditions such as equilibria or transport.2 ML approaches can screen many catalyst structures very quickly, leading to as much as six times higher efficiency compared with human-led experimentation and greater accuracy.⁶ Rather than explore an entire design space using traditional experimental design of experiments (DOE) approaches, a subset of a complex DOE may be sufficient to serve as training data for a machine learning model that would then be able to interpret multi-component interactions and condition effects to provide new insights with regard to feature influence (i.e., what the key variables are), inform next steps, and thus streamline the experiment selection process. The ability to perform orders of magnitude more experiments in-silico using machine learning models trained with real-world data is a powerful way to accelerate new catalyst discovery. Selected validation experiments would be performed to progress to each step in the development process (lab to pilot to commercial scale). Significant time savings would be realized through such a streamlined process that combines real data with AI-based modeling.Additionally, machine learning models trained from DFT data can be leveraged to predict the properties of extremely large numbers of candidate materials, therefore providing a screening function.  Hybrid models that couple experimental data and computational property predictions have the potential to be very powerful catalyst development accelerators.

Use Case: Machine Learning to Develop Non-Noble Metal CO2RR Catalysts

Copper-based catalysts suffer from low product selectivity and produce a range of products such as HCOOH, CH4, and C2H4. To produce any one of these hydrocarbons with a high selectivity requires controlling which elementary steps occur during the overall catalytic process, which is governed by the adsorption mode of CO2 and subsequent intermediates (CO, COOH, and HCOO) according to the following relationship:

E_adsorption = E_ads/E_sur – (E_surface + E_adsorbate)

‍

Obtaining these values can help design catalysts with optimal CO2 and intermediate adsorption onto their surface to selectively target a specific reduction product, as shown in Figure 1.

Due to the sheer number of possible adsorbates, transition states, and elementary steps during this and other chemical transformations, the problem quickly becomes computationally intractable. To facilitate high-throughput screening, ML and a microstructure model were used to develop seven active catalysts for forming methanol from CO2.⁷ Another approach proposed eight catalyst candidates by considering the work function, local electronegativity, electronegativity, interplanar spacing, and atomic number.⁸ Several of these candidates have been experimentally validated in other research, including Cu11In9, which showed a Faradaic efficiency of 92.8% when used for the electrocatalytic CO2RR to CO.⁹

NobleAI Accelerates the Development Cycle of Catalytic Processes

Because improvements to existing catalytic processes provide incremental improvements and diminishing returns for further optimization efforts,¹⁰ an approach is needed to optimize the catalyst design process to ensure that R&D efforts meet both production targets and sustainability goals.

The NobleAI Reactor Platform can model not only the individual material performance and properties but also the chemical and physical processes around it along with interactions in a real-world system, providing essentially a virtual experimentation space to explore catalyst candidates. NobleAI does this by using a Science-Based AI approach, in which appropriate ML models are combined with known scientific principles at different scales, to probe the thermodynamics and kinetics of catalytic processes. This approach allows users to leverage their existing experimental data and any available computational data to find viable candidate catalysts. The NobleAI Reactor platform utilizes deployed science-based AI models to provide additional insights into the chemistry of the materials, quantify uncertainty, and ultimately provide novel catalyst structures, shortening the screening process from months to only days, or even minutes.No matter where you are in your catalytic process development – from just beginning to explore a new catalytic process, to improving the efficiency of an existing process – the NobleAI Reactor Platform can help focus your R&D efforts where they’re likely to have the greatest impact, saving both time and money and ultimately improving the sustainability of your processes.

‍

References

(1) Thomas, J. M. Summarizing Comments on the Discussion and a Prospectus for Urgent Future Action. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 2016, 374 (2061), 20150226. https://doi.org/10.1098/rsta.2015.0226.

(2) Leonard, K. C.; Hasan, F.; Sneddon, H. F.; You, F. Can Artificial Intelligence and Machine Learning Be Used to Accelerate Sustainable Chemistry and Engineering? ACS Sustainable Chem. Eng. 2021, 9 (18), 6126–6129. https://doi.org/10.1021/acssuschemeng.1c02741.

(3) Optimizing Catalyst Performance to Support Sustainability Goals. https://www.aiche.org/resources/publications/cep/2021/january/optimizing-catalyst-performance-support-sustainability-goals (accessed 2023-11-30).

(4) Rosales, A. R.; Wahlers, J.; Limé, E.; Meadows, R. E.; Leslie, K. W.; Savin, R.; Bell, F.; Hansen, E.; Helquist, P.; Munday, R. H.; Wiest, O.; Norrby, P.-O. Rapid Virtual Screening of Enantioselective Catalysts Using CatVS. Nat Catal 2019, 2 (1), 41–45. https://doi.org/10.1038/s41929-018-0193-3.

(5) Margraf, J. T.; Jung, H.; Scheurer, C.; Reuter, K. Exploring Catalytic Reaction Networks with Machine Learning. Nat Catal 2023, 6 (2), 112–121. https://doi.org/10.1038/s41929-022-00896-y.

(6) Human versus Robots in the Discovery and Crystallization of Gigantic Polyoxometalates - Duros - 2017 - Angewandte Chemie International Edition - Wiley Online Library. https://onlinelibrary.wiley.com/doi/full/10.1002/anie.201705721 (accessed 2023-12-01).

(7) Roy, D.; Mandal, S. C.; Pathak, B. Machine Learning-Driven High-Throughput Screening of Alloy-Based Catalysts for Selective CO2 Hydrogenation to Methanol. ACS Appl. Mater. Interfaces 2021, 13 (47), 56151–56163. https://doi.org/10.1021/acsami.1c16696.

(8) Xing, M.; Zhang, Y.; Li, S.; He, H.; Sun, S. Prediction of Carbon Dioxide Reduction Catalyst Using Machine Learning with a Few-Feature Model: WLEDZ. J. Phys. Chem. C 2022, 126 (40), 17025–17035. https://doi.org/10.1021/acs.jpcc.2c02161.

(9) Wan, W.-B.; Dai, T.-Y.; Shi, H.; Zeng, S.-P.; Wen, Z.; Zhang, W.; Lang, X.-Y.; Jiang, Q. Intermetallic Cu11In9in Situ Formed on Hierarchical Nanoporous Cu for Highly Selective CO2 Electroreduction. J. Mater. Chem. A 2022, 10 (8), 4333–4343. https://doi.org/10.1039/D1TA10163C.

(10) Tokuriki, N.; Jackson, C. J.; Afriat-Jurnou, L.; Wyganowski, K. T.; Tang, R.; Tawfik, D. S. Diminishing Returns and Tradeoffs Constrain the Laboratory Optimization of an Enzyme. Nat Commun 2012, 3 (1), 1257. https://doi.org/10.1038/ncomms2246.

‍