Accelerated discovery of CO2 electrocatalysts using active machine learning.

The rapid increase in global energy demand and the need to replace carbon dioxide (CO2)-emitting fossil fuels with renewable sources have driven interest in chemical storage of intermittent solar and wind energy1,2. Particularly attractive is the electrochemical reduction of CO2 to chemical feedstocks, which uses both CO2 and renewable energy3-8. Copper has been the predominant electrocatalyst for this reaction when aiming for more valuable multi-carbon products9-16, and process improvements have been particularly notable when targeting ethylene. However, the energy efficiency and productivity (current density) achieved so far still fall below the values required to produce ethylene at cost-competitive prices. Here we describe Cu-Al electrocatalysts, identified using density functional theory calculations in combination with active machine learning, that efficiently reduce CO2 to ethylene with the highest Faradaic efficiency reported so far. This Faradaic efficiency of over 80 per cent (compared to about 66 per cent for pure Cu) is achieved at a current density of 400 milliamperes per square centimetre (at 1.5 volts versus a reversible hydrogen electrode) and a cathodic-side (half-cell) ethylene power conversion efficiency of 55 ± 2 per cent at 150 milliamperes per square centimetre. We perform computational studies that suggest that the Cu-Al alloys provide multiple sites and surface orientations with near-optimal CO binding for both efficient and selective CO2 reduction17. Furthermore, in situ X-ray absorption measurements reveal that Cu and Al enable a favourable Cu coordination environment that enhances C-C dimerization. These findings illustrate the value of computation and machine learning in guiding the experimental exploration of multi-metallic systems that go beyond the limitations of conventional single-metal electrocatalysts.

Pourbaix Machine Learning Framework Identifies Acidic Water Oxidation Catalysts Exhibiting Suppressed Ruthenium Dissolution.

The demand for green hydrogen has raised concerns over the availability of iridium used in oxygen evolution reaction catalysts. We identify catalysts with the aid of a machine learning-aided computational pipeline trained on more than 36,000 mixed metal oxides. The pipeline accurately predicts Pourbaix decomposition energy (Gpbx) from unrelaxed structures with a mean absolute error of 77 meV per atom, enabling us to screen 2070 new metallic oxides with respect to their prospective stability under acidic conditions. The search identifies Ru0.6Cr0.2Ti0.2O2 as a candidate having the promise of increased durability: experimentally, we find that it provides an overpotential of 267 mV at 100 mA cm-2 and that it operates at this current density for over 200 h and exhibits a rate of overpotential increase of 25 µV h-1. Surface density functional theory calculations reveal that Ti increases metal-oxygen covalency, a potential route to increased stability, while Cr lowers the energy barrier of the HOO* formation rate-determining step, increasing activity compared to RuO2 and reducing overpotential by 40 mV at 100 mA cm-2 while maintaining stability. In situ X-ray absorption spectroscopy and ex situ ptychography-scanning transmission X-ray microscopy show the evolution of a metastable structure during the reaction, slowing Ru mass dissolution by 20× and suppressing lattice oxygen participation by >60% compared to RuO2.

Cluster-MLP: An Active Learning Genetic Algorithm Framework for Accelerated Discovery of Global Minimum Configurations of Pure and Alloyed Nanoclusters.

Structural characterization of nanoclusters is one of the major challenges in nanocluster modeling owing to the multitude of possible configurations of arrangement of cluster atoms. The genetic algorithm (GA), a class of evolutionary algorithms based on the principles of natural evolution, is a commonly employed search method for locating the global minimum configuration of nanoclusters. Although a GA search at the DFT level is required for the accurate description of a potential energy surface to arrive at the correct global minimum configuration of nanoclusters, computationally expensive DFT evaluation of the significantly larger number of cluster geometries limits its practicability. Recently, machine learning potentials (MLP) that are learned from DFT calculations gained significant attention as computationally cheap alternative options that provide DFT level accuracy. As the accuracy of the MLP predictions is dependent on the quality and quantity of the training DFT data, active learning (AL) strategies have gained significant momentum to bypass the need of large and representative training data. In this application note, we present Cluster-MLP, an on-the-fly active learning genetic algorithm framework that employs the Flare++ machine learning potential (MLP) for accelerating the GA search for global minima of pure and alloyed nanoclusters. We have used a modified version the Birmingham parallel genetic algorithm (BPGA) for the nanocluster GA search which is then incorporated into distributed evolutionary algorithms in Python (DEAP), an evolutionary computational framework for fast prototyping or technical experiments. We have shown that the incorporation of the AL framework in the BPGA significantly reduced the computationally expensive DFT calculations. Moreover, we have shown that both the AL-GA and DFT-GA predict the same global minima for all the clusters we tested.

WhereWulff: A Semiautonomous Workflow for Systematic Catalyst Surface Reactivity under Reaction Conditions.

This paper introduces WhereWulff, a semiautonomous workflow for modeling the reactivity of catalyst surfaces. The workflow begins with a bulk optimization task that takes an initial bulk structure and returns the optimized bulk geometry and magnetic state, including stability under reaction conditions. The stable bulk structure is the input to a surface chemistry task that enumerates surfaces up to a user-specified maximum Miller index, computes relaxed surface energies for those surfaces, and then prioritizes those for subsequent adsorption energy calculations based on their contribution to the Wulff construction shape. The workflow handles computational resource constraints such as limited wall-time as well as automated job submission and analysis. We illustrate the workflow for oxygen evolution reaction (OER) intermediates on two double perovskites. WhereWulff nearly halved the number of Density Functional Theory (DFT) calculations from ∼240 to ∼132 by prioritizing terminations, up to a maximum Miller index of 1, based on surface stability. Additionally, it automatically handled the 180 additional resubmission jobs required to successfully converge 120+ atoms systems under a 48-h wall-time cluster constraint. There are four main use cases that we envision for WhereWulff: (1) as a first-principles source of truth to validate and update a closed-loop self-sustaining materials discovery pipeline, (2) as a data generation tool, (3) as an educational tool, allowing users (e.g., experimentalists) unfamiliar with OER modeling to probe materials they might be interested in before doing further in-domain analyses, (4) and finally, as a starting point for users to extend with reactions other than the OER, as part of a collaborative software community.

Applying Large Graph Neural Networks to Predict Transition Metal Complex Energies Using the tmQM_wB97MV Data Set.

Machine learning (ML) methods have shown promise for discovering novel catalysts but are often restricted to specific chemical domains. Generalizable ML models require large and diverse training data sets, which exist for heterogeneous catalysis but not for homogeneous catalysis. The tmQM data set, which contains properties of 86,665 transition metal complexes calculated at the TPSSh/def2-SVP level of density functional theory (DFT), provided a promising training data set for homogeneous catalyst systems. However, we find that ML models trained on tmQM consistently underpredict the energies of a chemically distinct subset of the data. To address this, we present the tmQM_wB97MV data set, which filters out several structures in tmQM found to be missing hydrogens and recomputes the energies of all other structures at the ωB97M-V/def2-SVPD level of DFT. ML models trained on tmQM_wB97MV show no pattern of consistently incorrect predictions and much lower errors than those trained on tmQM. The ML models tested on tmQM_wB97MV were, from best to worst, GemNet-T > PaiNN ≈ SpinConv > SchNet. Performance consistently improves when using only neutral structures instead of the entire data set. However, while models saturate with only neutral structures, more data continue to improve the models when including charged species, indicating the importance of accurately capturing a range of oxidation states in future data generation and model development. Furthermore, a fine-tuning approach in which weights were initialized from models trained on OC20 led to drastic improvements in model performance, indicating transferability between ML strategies of heterogeneous and homogeneous systems.

Beyond independent error assumptions in large GNN atomistic models.

The calculation of relative energy difference has significant practical applications, such as determining adsorption energy, screening for optimal catalysts with volcano plots, and calculating reaction energies. Although Density Functional Theory (DFT) is effective in calculating relative energies through systematic error cancellation, the accuracy of Graph Neural Networks (GNNs) in this regard remains uncertain. To address this, we analyzed ∼483 × 106 pairs of energy differences predicted by DFT and GNNs using the Open Catalyst 2020-Dense dataset. Our analysis revealed that GNNs exhibit a correlated error that can be reduced through subtraction, challenging the assumption of independent errors in GNN predictions and leading to more precise energy difference predictions. To assess the magnitude of error cancellation in chemically similar pairs, we introduced a new metric, the subgroup error cancellation ratio. Our findings suggest that state-of-the-art GNN models can achieve error reduction of up to 77% in these subgroups, which is comparable to the error cancellation observed with DFT. This significant error cancellation allows GNNs to achieve higher accuracy than individual energy predictions and distinguish subtle energy differences. We propose the marginal correct sign ratio as a metric to evaluate this performance. Additionally, our results show that the similarity in local embeddings is related to the magnitude of error cancellation, indicating the need for a proper training method that can augment the embedding similarity for chemically similar adsorbate-catalyst systems.

Screening of bimetallic electrocatalysts for water purification with machine learning.

Electrocatalysis provides a potential solution to NO3 - pollution in wastewater by converting it to innocuous N2 gas. However, materials with excellent catalytic activity are typically limited to expensive precious metals, hindering their commercial viability. In response to this challenge, we have conducted the most extensive computational search to date for electrocatalysts that can facilitate NO3 - reduction reaction, starting with 59 390 candidate bimetallic alloys from the Materials Project and Automatic-Flow databases. Using a joint machine learning- and computation-based screening strategy, we evaluated our candidates based on corrosion resistance, catalytic activity, N2 selectivity, cost, and the ability to synthesize. We found that only 20 materials will satisfy all criteria in our screening strategy, all of which contain varying amounts of Cu. Our proposed list of candidates is consistent with previous materials investigated in the literature, with the exception of Cu-Co and Cu-Ag based compounds that merit further investigation.

Transfer learning using attentions across atomic systems with graph neural networks (TAAG).

Recent advances in Graph Neural Networks (GNNs) have transformed the space of molecular and catalyst discovery. Despite the fact that the underlying physics across these domains remain the same, most prior work has focused on building domain-specific models either in small molecules or in materials. However, building large datasets across all domains is computationally expensive; therefore, the use of transfer learning (TL) to generalize to different domains is a promising but under-explored approach to this problem. To evaluate this hypothesis, we use a model that is pretrained on the Open Catalyst Dataset (OC20), and we study the model's behavior when fine-tuned for a set of different datasets and tasks. This includes MD17, the *CO adsorbate dataset, and OC20 across different tasks. Through extensive TL experiments, we demonstrate that the initial layers of GNNs learn a more basic representation that is consistent across domains, whereas the final layers learn more task-specific features. Moreover, these well-known strategies show significant improvement over the non-pretrained models for in-domain tasks with improvements of 53% and 17% for the *CO dataset and across the Open Catalyst Project (OCP) task, respectively. TL approaches result in up to 4× speedup in model training depending on the target data and task. However, these do not perform well for the MD17 dataset, resulting in worse performance than the non-pretrained model for few molecules. Based on these observations, we propose transfer learning using attentions across atomic systems with graph Neural Networks (TAAG), an attention-based approach that adapts to prioritize and transfer important features from the interaction layers of GNNs. The proposed method outperforms the best TL approach for out-of-domain datasets, such as MD17, and gives a mean improvement of 6% over a model trained from scratch.

Computational catalyst discovery: Active classification through myopic multiscale sampling.

The recent boom in computational chemistry has enabled several projects aimed at discovering useful materials or catalysts. We acknowledge and address two recurring issues in the field of computational catalyst discovery. First, calculating macro-scale catalyst properties is not straightforward when using ensembles of atomic-scale calculations [e.g., density functional theory (DFT)]. We attempt to address this issue by creating a multi-scale model that estimates bulk catalyst activity using adsorption energy predictions from both DFT and machine learning models. The second issue is that many catalyst discovery efforts seek to optimize catalyst properties, but optimization is an inherently exploitative objective that is in tension with the explorative nature of early-stage discovery projects. In other words, why invest so much time finding a "best" catalyst when it is likely to fail for some other, unforeseen problem? We address this issue by relaxing the catalyst discovery goal into a classification problem: "What is the set of catalysts that is worth testing experimentally?" Here, we present a catalyst discovery method called myopic multiscale sampling, which combines multiscale modeling with automated selection of DFT calculations. It is an active classification strategy that seeks to classify catalysts as "worth investigating" or "not worth investigating" experimentally. Our results show an ∼7-16 times speedup in catalyst classification relative to random sampling. These results were based on offline simulations of our algorithm on two different datasets: a larger, synthesized dataset and a smaller, real dataset.

Differentiable Optimization for the Prediction of Ground State Structures (DOGSS).

Ground state or relaxed inorganic structures are the starting point for most computational materials science or surface science analyses. Many of these structure relaxations represent systematic changes to the structure, but there are currently no general methods to improve the initial structure guess based on past calculations. Here we present a method to directly predict the ground state configuration using differentiable optimization and graph neural networks to learn the properties of a simple harmonic force field that approximates the ground state structure and properties. We demonstrate this flexible open source tool for improving the initial configurations for large datasets of inorganic multicomponent surface relaxations across 32 elements and the relaxation of adsorbates (H and CO) on these surfaces. Using these improved initial configurations reduces the expensive adsorbate-covered surface relaxations by approximately 50% and is complementary to other approaches to accelerate the relaxation process.

Capturing Structural Transitions in Surfactant Adsorption Isotherms at Solid/Solution Interfaces.

Although adsorption isotherms of surfactants are critical in determining the relationship between interfacial properties and structures of surfactants, providing quantitative predictions of the isotherms remains challenging. This is especially true for adsorption at hard interfaces such as on two-dimensional (2D) layered materials or on nanoparticles where simulation techniques developed for fluid-fluid interfaces that dynamically change surface properties by adjusting unit cells do not apply. In this work, we predict nonideal adsorption at a solid-solution interface with a molecular thermodynamic theory (MTT) model that utilizes molecular dynamics (MD) simulations for the determination of free-energy parameters in the MTT. Furthermore, the MD/MTT model provides atomistic insights into the nonideal behavior of surfactants by capturing structural phases of the surfactants at the interface. Our approach captures structural transitions from the ideal state at low concentrations and then to the critical surface aggregation concentration (CSAC) and finally through the critical micelle concentration (CMC). We validate our model against the original MTT model by comparing predicted adsorption isotherms of a simplified surfactant system from both approaches. We further substantiate the applicability of our model in complex systems by providing adsorption isotherms in an aqueous sodium dodecyl sulfate (SDS)-graphene system, in good agreement with the experimental observations of the CSAC for the same system.

In silico discovery of active, stable, CO-tolerant and cost-effective electrocatalysts for hydrogen evolution and oxidation.

Various databases of density functional theory (DFT) calculations for materials and adsorption properties are currently available. Using the Materials Project and GASpy databases of material stability and binding energies (H* and CO*), respectively, we evaluate multiple aspects of catalysts to discover active, stable, CO-tolerant, and cost-effective hydrogen evolution and oxidation catalysts. Finally, we suggest a few candidate materials for future experimental validations. We highlight that the stability analysis is easily obtainable but provides invaluable information to assess thermodynamic and electrochemical stability, bridging the gap between simulations and experiments. Furthermore, it reduces the number of expensive DFT calculations required to predict catalytic activities of surfaces by filtering out unstable materials.

Toward Predicting Intermetallics Surface Properties with High-Throughput DFT and Convolutional Neural Networks.

The surface energy of inorganic crystals is important in understanding experimentally relevant surface properties and designing materials for many applications. Predictive methods and data sets exist for surface energies of monometallic crystals. However, predicting these properties for bimetallic or more complicated surfaces is an open challenge. Computing cleavage energy is the first step in calculating surface energy across a large space. Here, we present a workflow to predict cleavage energies ab initio using high-throughput DFT and a machine learning framework. We calculated the cleavage energy of 3033 intermetallic alloys with combinations of 36 elements and 47 space groups. This high-throughput workflow was used to seed a database of cleavage energies. The database was used to train a crystal graph convolutional neural network (CGCNN). The CGCNN model provides an accurate prediction of cleavage energy with a mean absolute test error of 0.0071 eV/Å2. It can also qualitatively reproduce nanoparticle surface distributions (Wulff constructions). Our workflow provides quantitative insights into unexplored chemical space by predicting which surfaces are relatively stable and therefore more realistic. The insights allow us to down-select interesting candidates that we can study with robust theoretical and experimental methods for applications such as catalyst screening and nanomaterials synthesis.

Dynamic Workflows for Routine Materials Discovery in Surface Science.

The rising application of informatics and data science tools for studying inorganic crystals and small molecules has revolutionized approaches to materials discovery and driven the development of accurate machine learning structure/property relationships. We discuss how informatics tools can accelerate research, and we present various combinations of workflows, databases, and surrogate models in the literature. This paradigm has been slower to infiltrate the catalysis community due to larger configuration spaces, difficulty in describing necessary calculations, and thermodynamic/kinetic quantities that require many interdependent calculations. We present our own informatics tool that uses dynamic dependency graphs to share, organize, and schedule calculations to enable new, flexible research workflows in surface science. This approach is illustrated for the large-scale screening of intermetallic surfaces for electrochemical catalyst activity. Similar approaches will be important to bring the benefits of informatics and data science to surface science research. Lastly, we provide our perspective on when to use these tools and considerations when creating them.

2D equation-of-state model for corona phase molecular recognition on single-walled carbon nanotube and graphene surfaces.

Corona phase molecular recognition (CoPhMoRe) has been recently introduced as a means of generating synthetic molecular recognition sites on nanoparticle surfaces. A synthetic heteropolymer is adsorbed and confined to the surface of a nanoparticle, forming a corona phase capable of highly selective molecular recognition due to the conformational imposition of the particle surface on the polymer. In this work, we develop a computationally predictive model for analytes adsorbing onto one type of polymer corona phase composed of hydrophobic anchors on hydrophilic loops around a single-walled carbon nanotube (SWCNT) surface using a 2D equation of state that takes into consideration the analyte-polymer, analyte-nanoparticle, and polymer-nanoparticle interactions using parameters determined independently from molecular simulation. The SWCNT curvature is found to contribute weakly to the overall interaction energy, exhibiting no correlation for three of the corona phases considered, and differences of less than 5% and 20% over a larger curvature range for two other corona phases, respectively. Overall, the resulting model for this anchor-loop CoPhMoRe is able to correctly predict 83% of an experimental 374 analyte-polymer library, generating experimental fluorescence responses within 20% error of the experimental values. The modeling framework presented here represents an important step forward in the design of suitable polymers to target specific analytes.

Spatiotemporal intracellular nitric oxide signaling captured using internalized, near-infrared fluorescent carbon nanotube nanosensors.

Fluorescent nanosensor probes have suffered from limited molecular recognition and a dearth of strategies for spatial-temporal operation in cell culture. In this work, we spatially imaged the dynamics of nitric oxide (NO) signaling, important in numerous pathologies and physiological functions, using intracellular near-infrared fluorescent single-walled carbon nanotubes. The observed spatial-temporal NO signaling gradients clarify and refine the existing paradigm of NO signaling based on averaged local concentrations. This work enables the study of transient intracellular phenomena associated with signaling and therapeutics.

The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture.

Direct air capture (DAC) of CO2 with porous adsorbents such as metal-organic frameworks (MOFs) has the potential to aid large-scale decarbonization. Previous screening of MOFs for DAC relied on empirical force fields and ignored adsorbed H2O and MOF deformation. We performed quantum chemistry calculations overcoming these restrictions for thousands of MOFs. The resulting data enable efficient descriptions using machine learning.

Charge transfer at junctions of a single layer of graphene and a metallic single walled carbon nanotube.

Junctions between a single walled carbon nanotube (SWNT) and a monolayer of graphene are fabricated and studied for the first time. A single layer graphene (SLG) sheet grown by chemical vapor deposition (CVD) is transferred onto a SiO2/Si wafer with aligned CVD-grown SWNTs. Raman spectroscopy is used to identify metallic-SWNT/SLG junctions, and a method for spectroscopic deconvolution of the overlapping G peaks of the SWNT and the SLG is reported, making use of the polarization dependence of the SWNT. A comparison of the Raman peak positions and intensities of the individual SWNT and graphene to those of the SWNT-graphene junction indicates an electron transfer of 1.12 × 10¹³ cm⁻² from the SWNT to the graphene. This direction of charge transfer is in agreement with the work functions of the SWNT and graphene. The compression of the SWNT by the graphene increases the broadening of the radial breathing mode (RBM) peak from 3.6 ± 0.3 to 4.6 ± 0.5 cm⁻¹ and of the G peak from 13 ± 1 to 18 ± 1 cm⁻¹, in reasonable agreement with molecular dynamics simulations. However, the RBM and G peak position shifts are primarily due to charge transfer with minimal contributions from strain. With this method, the ability to dope graphene with nanometer resolution is demonstrated.

The chemical dynamics of nanosensors capable of single-molecule detection.

Recent advances in nanotechnology have produced the first sensor transducers capable of resolving the adsorption and desorption of single molecules. Examples include near infrared fluorescent single-walled carbon nanotubes that report single-molecule binding via stochastic quenching. A central question for the theory of such sensors is how to analyze stochastic adsorption events and extract the local concentration or flux of the analyte near the sensor. In this work, we compare algorithms of varying complexity for accomplishing this by first constructing a kinetic Monte Carlo model of molecular binding and unbinding to the sensor substrate and simulating the dynamics over wide ranges of forward and reverse rate constants. Methods involving single-site probability calculations, first and second moment analysis, and birth-and-death population modeling are compared for their accuracy in reconstructing model parameters in the presence and absence of noise over a large dynamic range. Overall, birth-and-death population modeling was the most robust in recovering the forward rate constants, with the first and second order moment analysis very efficient when the forward rate is large (>10(-3) s(-1)). The precision decreases with increasing noise, which we show masks the existence of underlying states. Precision is also diminished with very large forward rate constants, since the sensor surface quickly and persistently saturates.

Elimination of Multidrug-Resistant Bacteria by Transition Metal Dichalcogenides Encapsulated by Synthetic Single-Stranded DNA.

Antibiotic-resistant bacteria are a significant and growing threat to human health. Recently, two-dimensional (2D) nanomaterials have shown antimicrobial activity and have the potential to be used as new approaches to treating antibiotic resistant bacteria. In this Research Article, we exfoliate transition metal dichalcogenide (TMDC) nanosheets using synthetic single-stranded DNA (ssDNA) sequences, and demonstrate the broad-spectrum antibacterial activity of MoSe2 encapsulated by the T20 ssDNA sequence in eliminating several multidrug-resistant (MDR) bacteria. The MoSe2/T20 is able to eradicate Gram-positive Escherichia coli and Gram-positive Staphylococcus aureus at much lower concentrations than graphene-based nanomaterials. Eradication of MDR strains of methicillin-resistant S. aureus (MRSA), Enterococcus faecalis, Pseudomonas aeruginosa, Klebsiella pneumoniae, and Acinetobacter baumannii are shown to occur at at 75 µg mL-1 concentration of MoSe2/T20, and E. coli at 150 µg mL-1. Molecular dynamics simulations show that the thymine bases in the T20 sequence lie flat on the MoSe2 surface and can, thus, form a very good conformal coating and allow the MoSe2 to act as a sharp nanoknife. Electron microscopy shows the MoSe2 nanosheets cutting through the cell membranes, resulting in significant cellular damage and the formation of interior voids. Further assays show the change in membrane potential and reactive oxygen species (ROS) formation as mechanisms of antimicrobial activity of MoSe2/T20. The cellular death pathways are also examined by mRNA expression. This work shows that biocompatible TMDCs, specifically MoSe2/T20, is a potent antimicrobial agent against MDR bacteria and has potential for clinical settings.