Foundational machine-learning interatomic potentials (MLIPs) are being developed at a rapid pace, promising closer and closer approximation to ab initio accuracy. This unlocks the possibility to simulate much larger length and time scales. However, benchmarks for these MLIPs are usually limited to ordered, crystalline, and bulk materials. Hence, reported performance does not necessarily reflect MLIP performance accurately in real applications such as heterogeneous catalysis. Here, we systematically analyze zero-shot performance of 80 different MLIPs, evaluating tasks typical for heterogeneous catalysis across a range of different datasets, including adsorption and reaction on surfaces of alloyed metals, oxides, and metal-oxide interfacial systems. We demonstrate that current-generation foundational MLIPs can already perform with high accuracy for applications such as predicting vacancy formation energies of perovskite oxides or zero-point energies of supported nanoclusters. However, limitations also exist. We find that many MLIPs catastrophically fail when applied to magnetic materials, and structure relaxation in the MLIP generally increases the energy prediction error compared to single-point evaluation of a previously optimized structure. Comparing low-cost, task-specific models to foundational MLIPs, we highlight some core differences between these model approaches and show that-if considering only accuracy-these models can compete with the current generation of best-performing MLIPs. Furthermore, we show that no single MLIP universally performs best, requiring users to investigate MLIP suitability for their desired application.

Binding energies (BEs) of adsorbates on interstellar dust grains critically control adsorption, desorption, diffusion, and surface reactivity and therefore strongly influence astrochemical models of star- and planet-forming regions. While recent computational studies increasingly report full distributions of BEs rather than single representative values, these distributions are typically derived for either bare grain surfaces or thick water-ice mantles. In this work, we bridge these regimes by systematically investigating the BE distributions of water on partially and fully ice-covered dust grain surfaces. We employ machine learning interatomic potentials (MLIPs) based on graph neural networks to model water adsorption on graphene and on the Mg-terminated (010) surface of forsterite, representing carbonaceous and silicate grains, respectively. The models enable extensive sampling of adsorption sites on water clusters, monolayers, and bilayers generated under both crystalline (thermally processed) and amorphous (low temperature) growth conditions. At submonolayer coverage, the chemical nature of the underlying grain strongly affects both ice morphology and binding energies, with Mg-O interactions on silicate surfaces producing particularly deep binding sites. From monolayer coverage onward, adsorption on both substrates is dominated by hydrogen bonding within the ice, reducing the influence of the grain material. Across all coverages, amorphous ice structures systematically shift the BE distributions toward stronger binding compared to crystalline ice, introducing highly stable defect and pocket sites. These results demonstrate that BE distributions in the submonolayer-to-few-layer ice regime are broad and highly surface-dependent and provide physically motivated input for next-generation astrochemical models incorporating surface heterogeneity.

Determining transition states (TSs) of surface reactions is central to understanding and designing heterogeneous catalysts but remains computationally prohibitive with density functional theory (DFT). While machine learning potentials (MLPs) offer significant speedups, task-specific models have limited transferability across catalytic systems, and universal MLPs (uMLPs) lack the accuracy needed for reactive configurations. Here, we present a workflow based on active learning to iteratively fine-tune uMLPs for DFT-quality TS search. Using 250 TSs from the CO2 hydrogenation reaction network on metal and single-atom alloy surfaces, we first benchmark TS search algorithms, identifying the Sella algorithm as most robust, and propose a modification (Bond-Aware Sella) that substantially improves its success rate. We then explore sequential and batch active-learning strategies for fine-tuning and show that DFT-quality TS structures can be found using only 8 DFT single-point calculations on average per structure. This demonstrates the viability of fine-tuned uMLPs for high-throughput catalyst screening.

The conversion of CO2 into useful products such as methanol is a key strategy for abating climate change and our dependence on fossil fuels. Developing improved catalysts for this process is costly and time-consuming and can thus benefit from computational exploration of possible active sites. However, this is complicated by the complexity of the materials and reaction networks. Here, we present a workflow for exploring transition states of elementary reaction steps at inverse catalysts, which is based on the training of a neural network-based machine learning interatomic potential. We focus on the crucial formate intermediate and its formation over nanoclusters of indium oxide supported on Cu(111). The speedup compared to an approach purely based on density functional theory allows us to probe a wide variety of active sites found at nanoclusters of different sizes and stoichiometries. Analysis of the obtained set of transition state geometries reveals different structure-activity trends at the edge or interior of the nanoclusters. Furthermore, the identified geometries allow for the breaking of linear scaling relations, which could be a key underlying reason for the catalytic performance of inverse catalysts observed in experiments.

The conversion of CO2 to value-added compounds is an important part of the effort to store and reuse atmospheric CO2 emissions. Here, we focus on CO2 hydrogenation over so-called inverse catalysts: transition metal oxide clusters supported on metal surfaces. The conventional approach for computational screening of such candidate catalyst materials involves a reliance on density functional theory (DFT) to obtain accurate adsorption energies at a significant computational cost. Here, we present a machine learning (ML)-accelerated workflow for obtaining adsorption energies at the metal-oxide interface. We enumerate possible binding sites at the clusters and use DFT to sample a subset of these with diverse local adsorbate environments. The dataset is used to explore interpretable and black-box ML models with the aim of revealing the electronic and structural factors controlling adsorption at metal-oxide interfaces. Furthermore, the explored ML models can be used for low-cost prediction of adsorption energies on structures outside of the original training dataset. The workflow presented here, along with the insights into trends in adsorption energies at metal-oxide interfaces, will be useful for identifying active sites, predicting parameters required for microkinetic modeling of reactions on complex catalyst materials, and accelerating data-driven catalyst design.

Computational modeling of metal–oxide interfaces is challenging due to the large search space of compositions and structures and the complexity of catalyst materials under operating conditions in general. In this work, we develop an efficient structure search workflow to discover chemically unique and relevant nanocluster geometries of inverse catalysts and apply it to ZnyOx and InyOx on Cu(111), Pd(111), and Au(111). We show that the workflow is successful in obtaining a large range of chemically distinct structures. Structural geometry trends are identified, including stable motifs such as tripod, rhombus, and pyramidal motifs. Using ab initio thermodynamics, we explore the in situ stability of the structures, including single-atom alloys, at a range of oxygen availabilities. This approach allows us to find trends such as the susceptibility to oxidation of the different systems and the range of stability of different cluster motifs. Our analysis highlights the importance of taking the diversity of sites exposed by metal–oxide interfaces into account in catalyst design studies.

Obtaining accurate transition state (TS) energies is a bottleneck in computational screening of complex materials and reaction networks due to the high cost of TS search methods and first-principles methods such as density functional theory (DFT). Here we propose a machine learning (ML) model for predicting TS energies based on Gaussian process regression with the Wasserstein Weisfeiler-Lehman graph kernel (WWL-GPR). Applying the model to predict adsorption and TS energies for the reverse water-gas shift (RWGS) reaction on single-atom alloy (SAA) catalysts, we show that it can significantly improve the accuracy compared to traditional approaches based on scaling relations or ML models without a graph representation. Further benefitting from the low cost of model training, we train an ensemble of WWL-GPR models to obtain uncertainties through subsampling of the training data and show how these uncertainties propagate to turnover frequency (TOF) predictions through the construction of an ensemble of microkinetic models. Comparing the errors in model-based vs DFT-based TOF predictions, we show that the WWL-GPR model reduces errors by almost an order of magnitude compared to scaling relations. This demonstrates the critical impact of accurate energy predictions on catalytic activity estimation. Finally, we apply our model to screen new materials, identifying promising catalysts for RWGS. This work highlights the power of combining advanced ML techniques with DFT and microkinetic modeling for screening catalysts for complex reactions like RWGS, providing a robust framework for future catalyst design.

We investigate the interaction of CO 2 with metallic and oxidized Cu(110) surfaces using a combination of near-ambient pressure scanning tunneling microscopy (NAP-STM) and theoretical calculations. While the Cu(110) and full CuO films are inert, the interface between bare Cu(110) and the CuO film is observed to react instantly with CO 2 at a 10 mbar pressure. The reaction is observed to proceed from the interfacial sites of CuO/Cu(110). During reaction with CO 2 , the CuO/Cu(110) interface releases Cu adatoms which combine with CO 3 to produce a variety of added Cu−CO 3 structures, whose stability depends on the gas pressure of CO 2 . A main implication for the reactivity of Cu(110) is that Cu adatoms and highly undercoordinated CuO segments are created on the Cu(110) surface through the interaction with CO 2 , which may act as reaction-induced active sites. In the case of CO 2 hydrogenation to methanol, our theoretical assessment of such sites indicates that their presence may significantly promote CH 3 OH formation. Our study thus implies that the CuO/Cu(110) interfacial system is highly dynamic in the presence of CO 2 , and it suggests a possible strong importance of reaction-induced Cu and CuO sites for the surface chemistry of Cu(110) in CO 2 -related catalysis.

The doping of zirconia to enhance its activity and selectivity for the hydrogenation of CO2 to methanol has been studied intensively in experiments, but a thorough theoretical understanding of the factors that decide whether a dopant has a positive or negative influence on the reactivity is lacking. In this work, we conduct a mechanistic investigation using density functional theory and microkinetic modeling, considering the ZrO2(101) surface doped with 16 different metals. This analysis elucidates the following two criteria for enhanced reactivity. One, the ability of the surface to facilitate the dissociation of H-2 and provide the H* species necessary for the catalytic reaction is deemed a necessary but not sufficient criterion. Two, dopants that are thermodynamically stable under reaction conditions in a 2+ or 3+ oxidation state are beneficial, since this entails the introduction of O vacancies, which stabilize O-containing reaction intermediates such as formate and lower key transition states. We construct linear scaling relations that can reliably predict transition state energies in terms of less computationally costly adsorption energies. It is revealed that dopants that are stable in the 4+ state (e.g., Ti), and thereby lack O vacancies, follow a different scaling relation with a higher intercept for formate formation, which can explain their reduced reactivity. Overall, our microkinetic models can successfully predict the trends for dopants that have been found active in experiments (Zn2+, Cd2+, Ga3+, In3+) and not. Together with the established reactivity criteria, this paves the way for computational screening of oxides for the important CO2-to-methanol process.

An accurate understanding of the physicochemical properties of bimetallic heterogeneous catalysts relies on atomic-scale knowledge of the surface morphology and the atomic distribution. Alloys of Cu and Zn created during catalyst operation are frequently studied and debated in relation to a description of the active phase of Cu/ZnO/Al2O3 methanol synthesis catalysts. This makes it relevant to build a better understanding of Zn dissolution pathways in Cu surfaces and the resulting surface morphology. Herein, we use scanning tunneling microscopy to investigate surface morphology and the distinct atom site configurations of Zn and Cu on Cu(111) resulting from room-temperature Zn exchange from a Zn monolayer into the topmost layer of Cu(111). A gradual dissolution of Zn islands induces an extensive element intermixing at room temperature, resulting in Zn alloying at Cu terrace lattice sites. In addition, we observe and address an interlayer element exchange between the Zn submonolayers in direct contact with Cu. The exchange process is driven by lattice strain and is strongly facilitated at the perimeter of Zn edges. The STM contrast associated with the resulting intermixed sites is reported together with the simulation of these sites based on density functional theory, showing that imaging of isolated Zn sites in Cu(111) is sensitive to the STM tip state. The findings provide new insight into the atomic-scale exchange for Zn/Cu bimetallic surfaces, which may be used onward for understanding the debated surface morphology that develops during reductive activation and alloy formation in the Cu/ZnO/Al2O3 methanol synthesis catalyst.

The discovery of new inorganic magnesium electrolytes may act as a foundation for the rational design of novel types of solid-state batteries. Here we investigated a new type of organic-inorganic metal hydride, isopropylamine magnesium borohydride, Mg(BH4)2∙(CH3)2CHNH2, with hydrophobic domains in the solid state, which appear to promote fast Mg2+ ionic conductivity. A new synthetic strategy was designed by combination of solvent-based methods and mechanochemistry. The orthorhombic structure of Mg(BH4)2∙(CH3)2CHNH2 was solved ab initio by the Rietveld refinement of synchrotron X-ray powder diffraction data and density functional theory (DFT) structural optimization in space group I212121 (unit cell, a = 9.8019(1) Å, b = 12.1799(2) Å and c = 17.3386(2) Å). The DFT calculations reveal that the three-dimensional structure may be stabilized by weak dispersive interactions between apolar moieties and that these may be disordered. Nanoparticles and heat treatment (at T > 56 °C) produce a highly conductive composite, σ(Mg2+) = 2.86 × 10−7, and 2.85 × 10−5 S cm−1 at −10 and 40 °C, respectively, with a low activation energy, Ea = 0.65 eV. Nanoparticles stabilize the partially eutectic molten state and prevent recrystallization even at low temperatures and provide a high mechanical stability of the composite.

The d -band model was proposed by Bjørk Hammer and Jens Nørskov almost 30 years ago to explain trends in the interaction of adsorbates with transition-metal surfaces. It remains a cornerstone in heterogeneous catalysis research and has inspired a wealth of later models.

It is difficult to identify stable surface reconstructions of complex materials. Now a Monte Carlo sampling strategy is coupled with a machine learning interatomic potential that is iteratively improved via active learning during the search.

Binding energies of radicals and molecules at dust grain surfaces are important parameters for understanding and modeling the chemical inventory of interstellar gas clouds. While first-principles methods can reliably be used to compute such binding energies, the complex structure and varying sizes and stoichiometries of realistic dust grains make a complete characterization of all adsorption sites exposed by their surfaces challenging. Here, we focus on nanoclusters composed of Mg-rich silicates as models of interstellar dust grains and two adsorbates of particular astrochemical relevance; H and CO. We employ a compressed sensing method to identify descriptors for the binding energies, which are expressed as analytical functions of intrinsic properties of the clusters, obtainable through a single first-principles calculation of the cluster. The descriptors are identified based on a diverse training dataset of binding energies at low-energy structures of nanosilicate clusters, where the latter structures were obtained using a first-principles-based global optimization method. The composition of the descriptors reveals how electronic, electrostatic, and geometric properties of the nanosilicates control the binding energies and demonstrates distinct physical origins of the bond formation for H and CO. The predictive performance of the descriptors is found to be limited by cluster reconstruction, e.g., breaking of internal metal-oxygen bonds, upon the adsorption event, and strategies to account for this phenomenon are discussed. The identified descriptors and the computed datasets of stable nanosilicate clusters along with their binding energies are expected to find use in astrochemical models of reaction networks occurring at silicate grain surfaces.

Room temperature oxygen hydrogenation below graphene flakes supported by Ir(111) is investigated through a combination of X-ray photoelectron spectroscopy, scanning tunneling microscopy, and density functional theory calculations using an evolutionary search algorithm. We demonstrate how the graphene cover and its doping level can be used to trap and characterize dense mixed O-OH-H2O phases that otherwise would not exist. Our study of these graphene-stabilized phases and their response to oxygen or hydrogen exposure reveals that additional oxygen can be dissolved into them at room temperature creating mixed O-OH-H2O phases with an increased areal coverage underneath graphene. In contrast, additional hydrogen exposure converts the mixed O-OH-H2O phases back to pure OH-H2O with a reduced areal coverage underneath graphene.

A very important parameter for the electrocatalytic performance of a material is its potential of zero free charge (pzfc). For Cu(111) at pH 13 it was identified at −0.73 VSHE in the apparent double layer region. It shifts by (88±4) mV to more positive potentials per decreasing pH unit.1 At the pzfc, Cu(111) starts to restructure. At higher potentials, full reconstruction and electric field dependent OH adsorption occur, causing a remarkable decrease in the atomic density of the first Cu layer.1 It is this restructuring that enables Cu(111) to efficiently oxidize CO and to reduce water. Therefore, knowledge of the surface structure and the position of the pzfc is of paramount importance for the understanding of copper’s catalytic properties and for the rational design of electrocatalysts. CO is a key intermediate in the electro-oxidation of energy carrying fuels and known to act as a catalyst poison. Single-crystal Cu(111) model catalysts can efficiently electro-oxidize CO in alkaline media,2 where strong surface structural changes are observed under reaction conditions with electrochemical scanning tunneling microscopy (EC-STM). Supported by first-principles microkinetic modelling, we have shown that the concomitant presence of high-energy undercoordinated Cu structures at the surface is a prerequisite for the high activity. In water electrolyzers, it is possible to produce H2 in the course of the hydrogen evolution reaction (HER), which was studied with Ni(OH)2 and Co(OH)2 modified Cu(111) electrodes in alkaline media. 3 Strong morphological changes upon adatom modification lead to a significant HER rate enhancement. Intriguingly, this is induced through a decrease of the electric field strength negative of the pzfc. This implies an easier reorganization of the interfacial water molecules facilitating charge transfer through the double layer, and thus enhancing the efficiency of electrocatalytic reactions. The tendency of Cu(111) to restructure is found to dominate its electrochemical properties. The structural changes of the electrode surface are intimately related to the electric field at the solid/liquid interface and to its electrocatalytic activity, in general. References: [1] A. Auer, X. Ding, A. S. Bandarenka, J. Kunze-Liebhäuser, J. Phys. Chem. C 125 (2021) 5020. [2] A. Auer, M. Andersen, E.M. Wernig, N.G. Hörmann, N. Buller, K. Reuter, J. Kunze-Liebhäuser, Nature Catal. 3 (2020) 797. [3] A. Auer, F.J. Sarabia, D. Winkler, C. Griesser, V. Climent, J.M. Feliu, J. Kunze-Liebhäuser, ACS Catal. 11 (2021) 10324. Figure 1

Computational screening in heterogeneous catalysis relies increasingly on machine learning models for predicting key input parameters due to the high cost of computing these directly using first-principles methods. This becomes especially relevant when considering complex materials spaces such as alloys, or complex reaction mechanisms with adsorbates that may exhibit bi- or higher-dentate adsorption motifs. Here we present a data-efficient approach to the prediction of binding motifs and associated adsorption enthalpies of complex adsorbates at transition metals and their alloys based on a customized Wasserstein Weisfeiler-Lehman graph kernel and Gaussian process regression. The model shows good predictive performance, not only for the elemental transition metals on which it was trained, but also for an alloy based on these transition metals. Furthermore, incorporation of minimal new training data allows for predicting an out-of-domain transition metal. We believe the model may be useful in active learning approaches, for which we present an ensemble uncertainty estimation approach.

Context. The behaviour of molecules in space is to a large extent governed by where they freeze out or sublimate. The molecular binding energy is therefore an important parameter for many astrochemical studies. This parameter is usually determined with time-consuming experiments, computationally expensive quantum chemical calculations, or the inexpensive yet relatively inaccurate linear addition method. Aims. In this work, we propose a new method for predicting binding energies (BEs) based on machine learning that is accurate, yet computationally inexpensive. Methods. We created a machine-learning (ML) model based on Gaussian process regression (GPR) and trained it on a database of BEs of molecules collected from laboratory experiments presented in the literature. The molecules in the database are categorised by their features, such as mono- or multilayer coverage, binding surface, functional groups, valence electrons, and H-bond acceptors and donors. Results. We assessed the performance of the model with five-fold and leave-one-molecule-out cross validation. Predictions are generally accurate, with differences between predicted binding energies and values from the literature of less than ±20%. We used the validated model to predict the binding energies of 21 molecules that were recently detected in the interstellar medium, but for which binding energy values are unknown. We used a simplified model to visualise where the snow lines of these molecules would be located in a protoplanetary disk. Conclusions. This work demonstrates that ML can be employed to accurately and rapidly predict BEs of molecules. Machine learning complements current laboratory experiments and quantum chemical computational studies. The predicted BEs will find use in the modelling of astrochemical and planet-forming environments.

Liquid metal catalysts have recently attracted attention for synthesizing high-quality 2D materials facilitated via the catalysts' perfectly smooth surface. However, the microscopic catalytic processes occurring at the surface are still largely unclear because liquid metals escape the accessibility of traditional experimental and computational surface science approaches. Hence, numerous controversies are found regarding different applications, with graphene (Gr) growth on liquid copper (Cu) as a prominent prototype. In this work, novel in situ and in silico techniques are employed to achieve an atomic-level characterization of the graphene adsorption height above liquid Cu, reaching quantitative agreement within 0.1 Å between experiment and theory. The results are obtained via in situ synchrotron X-ray reflectivity (XRR) measurements over wide-range q-vectors and large-scale molecular dynamics simulations based on efficient machine-learning (ML) potentials trained to first-principles density functional theory (DFT) data. The computational insight is demonstrated to be robust against inherent DFT errors and reveals the nature of graphene binding to be highly comparable at liquid Cu and solid Cu(111). Transporting the predictive first-principles quality via ML potentials to the scales required for liquid metal catalysis thus provides a powerful approach to reach microscopic understanding, analogous to the established computational approaches for catalysis at solid surfaces.