Following the dynamics of industrial catalysts under operando conditions

Catalytic reactions taking place in industrial processes are often performed under extreme conditions of temperatures and pressures. A typical example is the Haber–Bosch process to industrially synthesize ammonia from nitrogen and hydrogen which operates under reaction pressures from 10 to 15 MPa and temperatures higher than 400 °C (1). Following the dynamics of heterogeneous catalysts under such extreme conditions is highly challenging both from experimental and theoretical point of view. In their paper, Bonati et al. give unique molecular insights into the dynamics, adsorption, and bond breakage of the N₂ molecule when interacting with the (111) iron surface at high temperatures relevant for the Haber–Bosch catalytic system (2). The simulations reveal that the surface is much more dynamic than anticipated from low-temperature experiments or simulations. Active sites are continuously formed and disrupted, and this behavior is instrumental for driving the catalytic process. To follow the weakening of the nitrogen–nitrogen bond, the degree of charge transfer from the metallic surface to the triple bond was followed during various steps of the catalytic process.

The study of Bonati et al. is an important proof-of-concept study, showcasing that reaction mechanisms may be highly dependent on the reaction conditions and that an evaluation of the dynamics of the system at industrially relevant conditions is of utmost importance to obtain molecular insights. Such insights can not be obtained from low-temperature investigations. It is notoriously difficult to simulate the dynamics of industrially relevant catalytic reactions under operating conditions. Hence, Bonati et al. had to combine various innovations in the field of machine learning and enhanced sampling molecular dynamics to follow the adsorption and reaction on the fly at operating conditions during sufficiently long time scales. A summary of some essential ingredients of their workflow is schematically shown in Fig. 1 and discussed further below. The impact of the methodological advances presented in their study is of much greater importance than the specific case study discussed in the PNAS paper and opens perspectives to follow industrially relevant catalytic reactions on the fly at the conditions where the catalyst does the work.

Fig. 1.

Schematic illustration on essential components of the methodological workflow necessary to study complex catalytic phenomena at operating conditions. QM stands for quantum mechanical, ML for Machine Learning and MLP for Machine Learning Potential.

The study of Bonati et al. is an important proof-of-concept study, showcasing that reaction mechanisms may be highly dependent on the reaction conditions and that an evaluation of the dynamics of the system at industrially relevant conditions is of utmost importance to obtain molecular insights.

In their study, Bonati et al. studied the impact of temperature on a crucial step of the Haber–Bosch process (1). In the broader field of heterogeneous catalysis, also other process parameters crucially determine the selectivity, activity, or lifetime of the catalyst, like the pH in the field of electrocatalysis or water content when converting complex feedstocks (3). Remarkably, industrially used catalytic solids remain stable under these extreme conditions where they are continuously exposed to a flow of reactants, products, and reaction intermediates (4). However, the catalytic surface is very dynamic and reactive intermediates may completely change nature at operating conditions, as convincingly shown in the PNAS paper. Such effects have also been observed in other fields of heterogeneous catalysis. Striking examples within zeolite catalysis comprise the change in nature of an adsorbed alkene on a Brønsted acidic zeolite at high temperatures relevant for catalytic cracking or mobility of metals coordinated to the zeolite active sites upon interaction with reactive species (5–7).

Following the dynamics of an industrial catalyst at operando conditions poses a tremendous challenge both for experimentalists and theoreticians (8). Experimental characterization is often performed under much lower temperatures and pressures than the industrially relevant conditions. Around 2000, the need to further develop spectroscopic and microscopic techniques to investigate experimentally how an industrial catalyst was working was recognized and the field of operando characterization was launched (9, 10). The aim is to use one or more spectroscopy or microscopy techniques to interrogate the catalytic behavior under realistic conditions with real-time online analysis of products. Since its inception, the field has flourished and systematically better advanced spectroscopic methods were developed to follow the dynamics of heterogeneous catalytic solids with systematically better spatial and temporal resolution. Yet today, it is certainly not possible from experimental point of view to make a molecular movie of a catalyst in action under operando conditions (11–13). In contrast to experiments, modeling follows a bottom-up approach starting from the atomistic scale by building realistic structures for the catalyst with the aim to evaluate the catalyst behavior when interrogated with reactive species under operando conditions. Operando modeling of industrially relevant catalytic systems is very challenging given the vastly different lengths and time scales involved and the complexity of the realistic catalyst (8, 14, 15). With the advancement of machine learning methods and their infiltration in the field of chemistry, we are entering a new era in simulating catalytic reactions under operating conditions as shown by Bonati et al. However, as shown by the authors, such methods need to be combined with other methodological advances and their integration in a computational catalytic workflow is critically dependent on a thorough insight into the system at hand.

Some key ingredients of the methodological workflow are illustrated in Fig. 1. First, as reactive events are studied, one needs to rely on a quantum mechanical (QM) based evaluation of energies and forces of the system. Currently, Density Functional Theory (DFT) is the method of choice for this purpose, however even with the expansion of computational power attainable time scales with DFT-based MD simulations are limited to hundreds of picoseconds, which is way too short to make a meaningful comparison with experimentally observed reactions (14). Second, reactive events are rare events, meaning that the probability of observing these in a regular MD run is very low. To sample in a more efficient way activated regions of phase space-enhanced sampling MD methods have become very popular (16). Still, even with usage of such advanced sampling techniques, the computational cost remains too expensive to directly access experimental time scales.

To mitigate this, Bonati et al. derived a machine learning potential (MLP) from underlying QM data that yields energies and forces of the system with almost equal accuracy than the underlying QM on which it was trained. As such, much longer simulations of tens of nanoseconds could be performed. Training a reactive MLP for industrially relevant systems is very challenging, since one needs to include all essential configurations to capture the chemistry of the system in the underlying QM training dataset. Such information is not automatically obtained from regular MD simulations since the probability to sample activated regions is very low and subsequent configurations in a MD run are highly correlated. In their paper, Bonati et al. combine machine learning procedures with enhanced sampling molecular dynamics methods to efficiently explore regions in the reactive event configuration space and combined it with an active learning procedure to automatically detect configurations which are not yet captured by the MLP. The challenge on how to efficiently generate a QM training dataset for complex systems in chemistry and material physics is an active branch of current research. Active learning strategies typically follow an iterative approach where an already trained MLP is used to sample phase space and generate configurations with new information to add to the QM training dataset, leading to a next-generation MLP. The final MLP should capture all relevant chemistry of the system of interest (17–19). The field is quickly evolving and future strategies might also use path sampling methods for adding new configurations to the training data. MLPs might even open the possibility to derive the kinetics with path sampling methods for complex chemical phenomena (20). Despite their potential, the accuracy of the final MLPs is critically dependent on the quality of the underlying QM data. In the paper of Bonati et al., training data could safely be obtained from DFT, but for more challenging systems such as metal-loaded zeolites with open shell structures or metal-organic frameworks with partially filled d- or f- shells, DFT itself may not be sufficiently accurate and high-level QM data might be needed to train the MLP. This further illustrates the need for advances at the cross road of machine learning techniques, active learning strategies, and advanced electronic structure methods to further progress in this field.

Armed with these methodological advances, intriguing new insights were obtained for the N₂ dissociation step. At high temperatures (700K), the surface does not remain flat, instead, a dynamic creation of hills and holes takes place and the diffusivity of surface atoms is substantially increased. At low temperatures (300K), a series of well-defined adsorption states are found for the N₂ molecule; however, at high temperatures, such distinction becomes less obvious and an ensemble of mobile horizontal N₂ configurations on disordered surface structures is found. To study the charge transfer from the metallic surface to the N₂ molecule, a second machine-learning model was developed to estimate the partial charges for a fixed atomic configurations without having to perform expensive QM calculations. At low temperatures, one finds a typical prereactive state corresponding to a pocket where N₂ can make a sevenfold coordination with iron to optimize the donation of electrons to the molecule. At high temperatures, such cavities are continuously formed and broken and N₂ must find a critical configuration to transfer all the necessary charge to weaken its bond. It is clear that the reactive behavior at high temperatures can not be extrapolated from observations at low temperatures.

The work presented by Bonati et al. focuses on one crucial step of the Haber–Bosh catalytic process. The overall reaction cycle operative in an industrial set-up is much more complex, with more reactions, lateral interactions between active sites and other agents like promotors. However, the presented methodological workflow gives critical insights into the interaction of reagents with the catalyst surface at high temperatures. Most importantly the work opens perspectives for studying the dynamics of many other heterogeneous catalytic systems like catalysis in nanoporous solids, on metal surfaces under a broad variety of operating conditions.