Establishing C-C electro-coupling mechanism via big data analyses of reaction networks

doi:10.21203/rs.3.rs-3950263/v1

Download PDF

Article

Establishing C-C electro-coupling mechanism via big data analyses of reaction networks

https://doi.org/10.21203/rs.3.rs-3950263/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

C-C coupling is important in electrocatalytic CO2 reduction for production of green chemicals. However, understanding of the underlying coupling mechanism(s) is poor because of the complexity of the reaction network. Here, we report a method to establish the structure-activity relationship in C-C coupling reactions via statistical analyses on a big dataset. The dataset contains a total of 45,738 adsorption data, which are generated through a combination of quantum chemical computations and machine-learning and encompasses 378 active site structures formed by 27 transition metals together with 6 potential coupling precursors. Analyses of these data establishes that, 1) asymmetric coupling mechanisms exhibit potential for greater efficiency compared with symmetric coupling, agreeing well with reported experiment, and 2) C-C coupling selectivity of Cu-based catalysts can be boosted through bimetallic doping. The three-component combination of CuAgNb exhibited superior coupling selectivity compared with both Cu and CuAg catalysts. Importantly, we experimentally substantiate the CuAgNb catalyst selected based on data screening, to demonstrate actual boosted performance in C-C coupling. We conclude that this finding evidences practicality of our theoretical model, and that combining big data with complex catalytic reaction mechanisms will become a basis or a new paradigm for accelerating optimal catalyst design for multi-carbon chemicals production.

Physical sciences/Chemistry/Catalysis/Electrocatalysis

Physical sciences/Chemistry/Theoretical chemistry/Computational chemistry

In electrocatalytic CO₂ reduction^,, C-C coupling is the requisite process to generate multi-carbon (C₂₊) products. However, the efficiency of current dominant Cu-based electrocatalysts for this remains problematic. The design of catalysts with high activity and efficiency is practically difficult because of limited understanding of the reaction mechanism. Reported reaction mechanisms involve different precursors, with symmetric couplings including CO^, and CHO^, dimerization, together with asymmetric couplings including CO coupling with COH^,, CHO^,, CH, CH₃ and COH coupling with CHO. Because of inherent limitations of experimental technique, particularly lack of in-situ characterizations, it is practically difficult to compare all possible precursor coupling mechanisms within the same system. The introduction of various transition metal elements into the Cu-based catalysts is reportedly practically promising in improving selectivity of C₂₊ products including inert metals such as Ag and Au, active metals such as Zr and Lanthanides, and main group elements such as Al, Ga and Bi. However, there is a lack of systematic comparison and understanding of the structure-activity relationships.

To solve such problems, theoretical research provides complementary capabilities by considering the complete reaction network mechanism and active site composition simultaneously in the same system. This approach offers meaningful insights to guide experimental mechanism exploration, enhancing our understanding of the structure-activity relationship. On the experimental side, the formation of few-atom clusters as active sites on metal surfaces^, under CO₂ reduction working conditions serves as a basis for theoretical active-site engineering^,^, at the molecular scale. However, the complexity of the reaction network and the diversity of active site compositions result in a substantial number of calculations. This challenge is compounded by the insufficient efficiency of computationally intensive quantum chemical methods, such as density functional theory (DFT).

Machine learning (ML) has significantly accelerated DFT computation efficiency. For instance, neural network models can be trained to obtain all adsorption energy data using only about 5% of DFT calculated adsorption energies while maintaining acceptable accuracy, thus accelerating computational screening of electrocatalysts¹⁸. Nevertheless, current ML-driven computational screening works^,^, heavily rely on the scaling relationship of the adsorption energies of reaction intermediates on the surfaces of metals and alloys^,, and they generally adopt the adsorption energy of certain intermediates as descriptors^,. Consequently, this methodology is primarily used for reactions with relatively simple intermediates or mechanisms, for example, using the adsorption energy of *CO and *H in CO₂ reduction^18,34,35 and *O and *OH in oxygen reduction³⁶. It faces challenges in generalizing the methodology to reactions with multiple possible precursors and complex networks, such as C-C coupling, making it difficult to meet the current needs of electrocatalytic production of multi-carbon products. In addition, although the early derivation process of these reaction mechanisms/descriptors has achieved certain successful prediction capabilities, the amount of original data is often small, usually a few dozen data points, which brings reliability concerns when migrating to more complex systems. In order to capture important information from complex reaction processes, some research groups have proposed integrating data-driven methods into the perspective of chemical reaction networks to find reaction mechanisms and guide the design of catalysts^,. However, practical applications of these ideas in electrocatalysis are still rare.

Here, we actualize these perspectives by incorporating big data analysis into the computational screening of electrocatalysts for complex C-C coupling reaction networks. We built a big dataset of over 45,000 data points, encompassing all possible coupling combinations of 6 precursor species, together with adsorption configurations on the active site. Using iterative sampling and 2D-3D ensemble ML techniques, we expand the initial 2,500 DFT computed data by about 20 times. Our big dataset analysis shows the adsorption energy scaling-relationship between adsorbed species before and after coupling, evidences the asymmetric coupling mechanism superior to symmetric, and predicts boosted performance of Cu-catalyzed C-C coupling through active site doping. Importantly, we experimentally confirm the big data-analysis suggested optimal CuAgNb catalyst’s superior performance towards C-C coupling. This finding demonstrates the practicality of big data analysis methods in understanding intricate reaction network mechanisms and designing novel catalysts.

Construction of big data. By including all adsorption intermediates related to C-C coupling in

the big data, a coverage of both the precursors before coupling and products following can be assured. These data comprise all adsorption intermediates relevant to C-C coupling consisting of, 6 possible C₁ precursors, namely CO, COH, CHO, CH, CH₂ and CH₃, together with 6 symmetrically coupled, and 15 asymmetrically coupled, C₂ products derived from combinations of these precursors (Supplementary Fig. 1). Given the diverse symmetries of the adsorbed species and the necessity to determine the most stable adsorption energy as the optimal adsorption configuration, a total number of 121 unique adsorption structures on each catalyst surface need to be considered. Because of the important role of Cu in C-C coupling, all adsorption sites within the data consist of one Cu atom, whilst the other two atoms are chosen from a search space of 27 metal elements, leading to 378 unique ABCu (A,B = transition metals) tri-atomic active-sites, Fig. 1a. The combination of these adsorption structures with various sites generates big data comprising 45,738 adsorption energy data, Fig. 1a.

Because of the significant computations burden from DFT, building big data is practically time consuming. We developed therefore an iterative sampling method that balances computational burden and accuracy, Fig. 1b. This involved random sampling a subset (3d metals which are closest to Cu), and sequentially including the other two subsets (4d and 5d metals) to ensure in-distribution sampling of the whole dataset. This ensures all metal categories, permitting dynamic adjustments throughout the DFT computation and ML.

To overcome low ML accuracy, an application-oriented 2D-3D ensemble method was developed (Supplementary Fig. 2). The majority of reported graph neural network (GNN) models used for predicting adsorption energy rely solely on 3D atomic positions, as evidenced in the Initial Structure to Relaxed Energy task (IS2RE) in the Open Catalyst Project. Conversely, GNNs based on 2D connectivity graphs have been extensively reported for organic chemistry and tasks including drug discovery, however, use for heterogeneous catalysis is not widespread.

A practical difficulty in applying 2D-based GNNs to heterogeneous catalysis is the identification of bonds in inorganic materials. To obviate this, we identify the bonds within the adsorbates and the tri-atom catalytic active sites based on understanding of the chemical significance of surface adsorption structures, as illustrated in Supplementary Fig. 3. Figure 2 presents GNN predicted against DFT-computed adsorption energy using single 3D-based, or 2D-based GNN. For 3D-based GNNs, the DimeNet++^, algorithm exhibited greater accuracy compared with SchNet^, algorithm, with a mean absolute error (MAE) of 0.37 eV. In contrast, the graph convolutional network (GCN) algorithm exhibited greatest accuracy amongst all 2D-based GNN, with a MAE 0.36 eV. We attribute the low accuracy of single-modal models to data complex variation, including change in atomic positions because of substrate element compositions and alterations in connectivity caused by different adsorption species and sites. To obviate this, we combined top-performing DimeNet + + as the 3D-modal and GCN as the 2D-modal to create the 2D-3D ensemble. Significantly, the 2D-3D ensemble outperformed all single methods, reducing the MAE to ca. 0.32 eV, and importantly, met a practical standard for use. This provided an authoritative ML approach for predicting adsorption energy in complex reaction networks, effectively overcoming the limitations associated with the application of specific dimensional methods in the context of big data complexity.

The scaling relation. An important aspect of catalytic theory is the linear-correlation of adsorption energy between the same series of adsorbates on metal catalyst sites^30,31. This significantly simplifies theoretical computation of reaction intermediates, making it widely used therefore in catalyst screening. Related studies on C₂ species formed by C-C coupling are however limited because of increased computational complexity resulting from larger molecular size of C₂ species compared with C₁. Additionally, when dealing with few data, the resulting linear correlation is often weak, reducing reliability of conclusions. However, the big data used here for adsorption energy removes this. Figure 3a presents a strong linear correlation with R² coefficient generally > 0.8 between adsorption energy of the C₁ coupling reaction precursor and the C₂ coupling product, except however for saturated CH₃-CH₃ that exhibits a poor linear correlation because of physical adsorption. A good linear correlation is shown for the adsorption energies of the reaction precursors (Supplementary Fig. 4a) that is significantly different to the weaker linear correlation exhibited for pure metal surfaces (Supplementary Figs. 4b and 5). Because the big data (Fig. 3b, Supplementary Figs. 6 and 7) exceeds ordinary datasets (Supplementary Fig. 8), a conclusion of linear proportional correlation is reliable.

Because of the robust linear correlation between the adsorption energy of intermediates before and following C-C coupling, a high correlation between reaction properties and adsorption energy is demonstrated. It is concluded that this finding is a solid theoretical foundation on which to assess the ease of reactions via analyzing adsorption energies.

Positioning Cu in big data. In addition to establishing theoretical trends, our big data permits for valuable comparisons with experiment. We introduce the concept of coupling energy (E_cplg), representing the energy difference before and following coupling. A more negative coupling energy evidences a more ready reaction. Given the strong linear correlation established between the energy of reaction intermediates, it is concluded that there is a similarly strong linear correlation between adsorption energy and E_cplg, Fig. 4 (Supplementary Fig. 9).

A rule can be established, namely, the more negative the adsorption energy, the more positive the coupling energy, and vice versa, evidencing an inverse relationship between the ease of precursor adsorption and the ease of coupling. This rule explains why C-C coupling reactions are practically difficult to achieve in experiment. It is analogous to a seesaw, with the difficulty of obtaining related precursors on one end and the difficulty of the coupling reaction on the other, making it difficult to achieve balance. Importantly the Cu element, known as the sole element for C-C coupling in experiment, is often situated in the middle of this seesaw, marked at the site composed of pure Cu in plots. This evidences that Cu exhibits a balanced ability to form precursors and facilitate C-C bond coupling, making it a best catalyst for C-C coupling. This finding significantly agrees well with experimental fact, confirming a benchmarking ability of big data with experiment.

Whether symmetric or asymmetric coupling? Big data permits consideration of a broader range of possibilities than conventional experiment. Whether the coupling mechanism is symmetric or asymmetric led to an assessment of coupling comparisons between specific precursor species and other precursors from big data. Figure 5 presents findings for 6 precursors. The black dashed-line is symmetric coupling (coupling to self), whilst the colored data represent asymmetric coupling to different, other species.

Analyses confirmed that asymmetric coupling of CO with other precursors is more ready than symmetric coupling with itself, with a preference for coupling with CH₃. Significantly, this finding agrees well with reported experiments of Yang et al.¹¹ where CO-CH₃ coupling was superior to CO-CO coupling. Additionally, amongst other precursors, COH, CHO and CH are most likely to couple with CH₂, whilst CH₂ exhibits preference to couple with CHO. Although the trend for CH₃ is less apparent, it shows better coupling with other precursors than self-symmetric coupling. Asymmetric coupling therefore has practical potential for greater C-C coupling efficiency compared with symmetric coupling. Given that not all coupling intermediates are easy to detect practically, this is useful theoretical insight into the mechanism.

Enhancing Cu. Big data is practically useful for valuable catalyst screening, eliminating need to search for catalytic performance descriptors separately. Importantly, catalysts can be directly screened from the data, streamlining screening. Although Cu is in a favorable position, the abundance of data near Cu seems to evidence the possibility of a C-C coupling that has the potential to surpass it. A potential better C-C coupling contender will need to, 1) exhibit low coupling energy, preferably negative, to facilitate coupling, 2) be readily for coupling precursor adsorption, preferably more readily than on Cu, and; 3) exhibit precursor formation ideally exothermic, given the mutual relationship between the precursors (Supplementary Fig. 10).

In addition, selectivity is important in CO₂ electroreduction as well because of side reaction of H₂ evolution. It is reported that this selectivity is linked to the adsorption energy of CO in pH and potential-mediated microkinetic computational simulations. Hence, the adsorption energy for CO will be restricted within the range of ± 0.1 eV of Cu to ensure selectivity for CO₂ reduction and prevent H₂ evolution side reaction. This gives rise to four screening criteria, Fig. 6a, for all compositions of 378 catalyst active sites.

According to the criteria above, 11 practically promising catalyst compositions, CuAlNb, CuZr, CuAuHf, CuMnW, CuAgNb, CuHfRh, CuTi, CuRuTi, CuAlV, CuIrV and CuPtV, were identified as exhibiting potentially superior performance compared with Cu in C-C coupling. In their reaction mechanism, the intermediate most likely to undergo coupling is CH-CHO or CH₂-CHO. To confirm the accuracy of the ML big dataset, DFT computations were used to validate the complete C-C coupling reaction network for the 11 selected catalyst structures, including generation of coupling precursors CHO, CH and CH₂ and coupling steps. Supplementary Fig. 11 illustrates comparison between the DFT-computed and ML-predicted adsorption energies of optimal adsorption configurations for CO, COH, CHO, CH, CH₂, CHO-CH and CHO-CH₂ species from the big dataset. It was found that there was excellent agreement, confirming reliability of ML findings.

Importantly, to establish practical usefulness of the big data theoretical model, actual experiment using the CuAgNb catalyst from the 11 predicted compositions was carried out. Detailed characterizations (Fig. 6b, Supplementary Fig. 12) from electron microscopy, energy-dispersive X-ray spectroscopy (EDS) and X–ray diffraction (XRD), confirmed the structure and presence of elemental Cu, Ag and Nb with this synthesized catalyst. The experimental Faradaic efficiency (FE) for C₁ and C₂₊ products for the CuAgNb catalyst, together with a comparative analysis against Cu and CuAg catalysts is given in Fig. 6c. It can be seen that for selected current density conditions, the CuAgNb catalyst exhibits higher C₂₊ FE compared with Cu and CuAg, whilst C₁ FE is lower than that for Cu and CuAg. Further analyses of the products (Fig. 6d, Supplementary Fig. 13) established the highly significant performance of the CuAgNb catalyst, especially exhibiting low CO yield. Conversely, the yield of C₂H₄ and ethanol is significantly increased, making this catalyst practically promising for generation of highly value-added C₂ products. Importantly, this confirms theoretical predictions that CuAgNb catalyst exhibits optimal C-C coupling catalytic performance, surpassing both mono-component Cu and binary-component CuAg in greater C₂₊ FE. It is concluded that ML-based big data are practically useful to explore catalysts.

We established an extensive dataset to unravel the intricacies of the complex C-C electro-coupling reaction network, which realizes the application of big data analysis approach to electrocatalysis. This dataset encompasses a comprehensive range of reaction precursors, coupling mechanisms, and active sites for Cu-based catalysts featuring various transition metal components. The analysis of the big dataset serves as a valuable tool for comprehending the reaction mechanism, establishing the structure-activity relationship between catalyst components and coupling reaction efficiency, and guiding the catalyst screening process.

To mitigate computational and time expenses associated with building the big dataset and ensure the accuracy and reliability, we integrated DFT calculations with an iterative sampling method and a 2D-3D ensemble ML algorithm. This resulted in the acquisition of 45,738 adsorption energy data for C₁ and C₂ species, with an acceptable error range. The big data analysis revealed the optimal Cu positioning, and asymmetric coupling consistently outperforms symmetric coupling, aligning well with experimental observations. Furthermore, it identified CH₂ species as the preferred coupling partner.

Importantly, the big data analysis not only enhances our understanding but also offers successful guidance and predictions for experiments. The identification of CuAgNb screened via big data as a promising catalyst was experimentally validated, showcasing superior C₂₊ selectivity compared to both Cu and CuAg. We conclude therefore that the integration of big dataset construction and analysis emerges as a fundamental approach for comprehending complex C-C coupling catalytic mechanisms, expediting optimal catalyst design for valuable multi-carbon product conversions.

Density functional theory computation (DFT). DFT computations were performed using the Vienna ab initio simulation package (VASP), projector-augmented wave (PAW) pseudopotentials, and the revised Perbew-Burke-Ernzerhof (RPBE)^, exchange-correlation functional, with the DFT-D3 method for van der Waals correction. The plane wave cutoff was set to 600 eV. All atoms were fully relaxed until convergence of energy of 1×10^− 5 eV and forces of 0.05 eV·Å⁻¹. Spin polarization was implemented for structures involving Fe, Co and Ni. The Brillouin zone was sampled with a (4×4×1) k-point grid. See Supplementary Fig. 14 for model construction. Test computations were performed to assess the influence of temperature and solvent (Supplementary Table 1), and electrochemical working potential (Supplementary Figs. 15 and 16, Supplementary Table 2) on the C-C coupling step. Importantly, findings evidenced that that these factors had negligible impact on coupling energy. These factors were therefore ignored in construction of the big dataset.

Adsorption energy and coupling energy. The C-C coupling big data consisted of 45,738 adsorption energies (E_ad) including, 6 precursors (CO, COH, CHO, CH, CH₂ and CH₃), and 21 C₂ combinations (CO-CO, CO-COH, CO-CHO, CO-CH, CO-CH₂, CO-CH₃, COH-COH, COH-CHO, COH-CH, COH-CH₂, COH-CH₃, CHO-CHO, CHO-CH, CHO-CH₂, CHO-CH₃, CH-CH, CH-CH₂, CH-CH₃, CH₂-CH₂, CH₂-CH₃ and CH₃-CH₃). E_ad was computed from:

E _ad = E_total - E_substrate - E_species (1)

where E_total is DFT total energy of the substrate with adsorbate, E_substrate the energy of corresponding pristine substrate and E_species energy of adsorbate species referenced to CO₂, H₂O and H₂ molecules.

To describe ‘difficulty’ of the coupling reaction, the coupling energy (E_cplg) is defined as the difference between the total energy following coupling, E_total(Product), and the total energy of the two precursors, E_total(Precursor1) and E_total(Precursor2), before coupling, namely:

E _cplg = [E_total(Product) + E_substrate] - [E_total(Precursor1) + E_total(Precursor2)] (2)

By substituting E_ad for Product, Precusor1, and Precusor2 into Eq. (2) and simplifying the following is obtained:

E _cplg = E_ad(Product) - [E_ad(Precursor1) + E_ad(Precursor2)] (3)

DFT calculations were used for the verification of the overall reaction pathway on selected catalyst surfaces with additional computation for intermediates C, CHOH, CH₂O, and CH₂OH.

Enumerating search space. Substrates: The active site comprises a tri-atom consisting of ABCu, where A and B can be substituted by 27 metals each, forming therefore 322 possible combinations. Adsorption sites: A total of 121 surface structures were constructed for the tri-atom active site. The hollow position of the active site accommodates adsorbates CO, COH, CH, CH₂, C, CHOH and CH₂O, each with a single adsorption configuration. On the bridge position, CH₃ adsorbs with 3 different configurations. For symmetric adsorbates CO-CO, COH-COH, CHO-CHO, CH-CH, CH₂-CH₂ and CH₃-CH₃, one C adsorbs on the bridge site whilst the other adsorbs on the top site, resulting in 3 adsorption configurations. Similarly, for asymmetric adsorbates CO-COH, CO-CHO, CO-CH, CO-CH₂, CO-CH₃, COH-CHO, COH-CH, COH-CH₂, COH-CH₃, CHO-CH, CHO-CH₂, CHO-CH₃, CH-CH₂, CH-CH₃, CH₂-CH₃, CHO and CH₂OH, one C or O adsorbs on the bridge site and the other C or O adsorbs on the top site, resulting in 6 adsorption configurations for each. In addition, the pristine substrate when there is no adsorbate also needs to be calculated.

Multiplication of 378 substrate compositions and 121 adsorption configurations gives 45,738 data in the set. The construction of these models and simulations was carried out in the atomic simulation environment (ASE) package.

Graph neural network training and prediction. A number of graph neural network (GNN) models were used to learn, and to predict adsorption energy for the big C-C coupling dataset. The training-set consisted of 2,500 adsorption energies obtained via iterative sampling, whilst the prediction set used all 45,738 adsorption energies. Features included the initial structure of each adsorption species, as depicted in Supplementary Fig. 4, in addition to ‘fingerprint’ corresponding to the atoms in the respective structures.

It is commonly reported that information from multiple sources including, images, texts and time-series data can be used to boost predictive performance of a model. We hypothesized that building GNN models based solely 3D information could be extended to incorporate other data sources to improve further performance.

To evaluate performance of the models and using 2D-3D ensemble information, SchNet, DimeNet++, GIN, GraphSAGE and GCN were assessed. A combination of DimeNet + + and GCN was used for to build the big data.

SchNet

SchNet^31,32 is a ‘deep-learning’ model designed to efficiently capture quantum interactions within molecular systems. It treats atoms as nodes and 2-body terms as edges and uses message passing and interaction terms to model the continuous spatial representation of atomic environments, enabling accurate predictions of molecular properties including, energy and force(s).

DimeNet++

DimeNet + + ^33,34 is an extension of the DimeNet, optimized to predict molecular properties. In contrast to SchNet, it treats 2-body terms as nodes and 3-body terms as edge. DimeNet + + incorporates interatomic distances and angles to boost representational capacity.

GIN

Graph Isomorphism Network (GIN) is designed to predict graph-level properties or labels. In GIN used here, the edges are constructed based on bond connectivity of the nodes. GIN has reportedly been applied widely including, social network analyses and bioinformatics.

GraphSAGE

Graph Sample and Aggregator (GraphSAGE) is a scalable, inductive-learning framework that enables efficient representation learning on large graphs. Via sampling and aggregating node-level features from local neighborhoods, GraphSAGE captures structural information of graphs.

GCN

Graph Convolutional Network (GCN)³⁵ is a fundamental graph-based deep-learning model. It introduces graph convolutions to iteratively update node representations by aggregating information from neighboring nodes.

To leverage the complementary strengths of 3D-based and 2D-based GNN, we devised a 2D-3D ensemble model that integrated predictions from both modalities. This ensemble boosted predictive performance and robustness (a model’s ability to maintain its performance and generalization capabilities across diverse data sets) in predictions.

Chemicals. Copper (II) nitrate hexahydrate (Cu(NO₃)₂.3H₂O, AR), silver (I) nitrate (AgNO₃, AR), ammonium niobate(V) oxalate hydrate (C₂H₅NNbO₄, AR), potassium hydroxide (KOH, AR), sodium hydroxide (NaOH, AR) and ascorbic acid (AA, AR) were purchased from Sigma-Aldrich. Carbon dioxide (CO₂, 99.999%) was purchased from BOC Gas (Australia). All chemicals were used as received without further modification. Water (18 MΩ*cm) used in experiment was prepared via passing through an ultra-pure purification system.

Preparation of catalysts. The CuAgNb catalyst was prepared via electroreduction of pre-synthesized Cu₂O-AgNb catalysts on carbon paper during CO₂RR as follows: 0.5 mL NaOH solution (1 M), 0.5 mL Cu(NO₃)₂ solution (0.1 M) and 0.025 mL C₂H₅NNbO₄ (0.03M) were added to a 20 mL vial with 9 mL water under vigorous stirring for 5 min at room temperature. 18 mg AA was added to the vial under stirring. Following stirring for 30 min, 0.1 mL AgNO₃ solution (0.01M) was added to the vial, and stirring continued for 30 min. Resulting products were obtained via centrifugation, washed with ethanol and dispersed in 0.8 mL ethanol. The ethanol solutions were mixed with 20 µL 5 m/m% Nafion solution and sprayed using an airbrush onto a carbon paper. The loading mass of catalyst on carbon paper was controlled to ca. 0.5 mg cm^− 2. Carbon paper sprayed with pre-synthesized Cu₂O-AgNb catalysts was used as working electrode, and reduced at a current density of 200 mA cm^− 2 during CO₂RR to obtain the CuAgNb catalyst. Preparation of Cu and CuAg catalysts was similar however without addition of corresponding precursors AgNO₃ or C₂H₅NNbO₄ during pre-synthesis.

Characterizations. High-angle annular dark-field imaging and EDS mapping were determined on a FEI Titan Themis 80–200 operating at 200 kV. XRD data were determined on a Rigaku MiniFlex 600 X-Ray Diffractometer. SEM images were collected using a FEI QUANTA 450 FEG Environmental SEM OPERATING at 10 kV.

Electro-coupling performance tests. Electroreduction of CO₂ was tested in a microfluidics flow cell that consisted of two electrolyte chambers, and one gas chamber. An anion exchange membrane (Fumasep FAB-PK-130) was placed between the two electrolyte chambers for products separation and ionic conduction. Catalyst-deposited carbon paper (YSL-30T), micro Ag/AgCl electrode (4.0 M KCl) and Ni-foam (0.5 mm thickness), respectively, were working electrode, reference electrode and anode. The working electrode (active area: 0.5 x 2 cm^− 2) was placed between gas and catholyte chambers to ensure gaseous CO₂ diffusion and reaction at the catholyte/catalyst interface. An electrochemical workstation (CHI760) with a current amplifier was used to perform the CO₂RR test. 1 M KOH (20 mL) was circulated through the electrolyte chambers under a constant flow of 20 mL min^–1 via peristaltic pumping. CO₂ was supplied into the gas chambers via a mass flow controller at a constant flow of 30 mL min^–1. Reactions were tested via chronopotentiometry at differing current for 1 h without iR correction. Gaseous products were analyzed via GC equipped with a PLOT MolSieve 5A column and a Q-bond PLOT column. Liquid products were characterized via HPLC (Thermo Scientific RefractoMax 520) with a Bio-Rad Aminex HPX-87H column. All potentials were converted against RHE and iR correction based on, E_RHE = Eversus Ag/AgCl + 0.059 × pH + 0.210 + 0.85 × iR, where i is the current at each applied potential and R the equivalent series resistance measured via electrochemical impedance spectroscopy in the frequency. FE for reaction products was computed from, FE = eF × n/Q = eF × n/(I × t), where e is the number of electrons transferred, F Faraday constant, Q charge, I current, t running time and n total of product (mole).

Data availability

Data that support findings from this study are available from the corresponding author on reasonable request.

Acknowledgements

This work was financially supported by the Australian Research Council (DP230102027, CE230100032, LP210301397 and DE240100661). Computations were undertaken with resources from Phoenix High Performance Computing, which is supported by The University of Adelaide. H.L., Z.Z. and X.L. acknowledge financial support from the Center for Augmented Reasoning, Australian Institute for Machine Learning. H.L. gratefully acknowledges Dr Jodie A. Yuwono and Mr Yiran Jiao for technical support.

Author contributions

S.-Z.Q. supervised the project. H.L. and S.-Z.Q. conceived the idea. H.L. designed and performed the theoretical computations. X.L. and Z.Z. performed the machine learning studies. P.W. designed and carried out the experiments. H.L., X.L., Z.Z. and J.Q.S. discussed the machine learning results. H.L., K.D. and S.-Z.Q. wrote the manuscript. All authors discussed the results and commented on the manuscript.

Competing interests

The authors declare no competing interests.

Seh, Z. W., Kibsgaard, J., Dickens, C. F., Chorkendorff, I., Norskov, J. K. & Jaramillo, T. F. Combining theory and experiment in electrocatalysis: Insights into materials design. Science 355, eaad4998 (2017).
Tang, C., Zheng, Y., Jaroniec, M. & Qiao, S.-Z. Electrocatalytic refinery for sustainable production of fuels and chemicals. Angew. Chem. Int. Ed. 60, 19572-19590 (2021).
Zhou, Y., Che, F., Liu, M. et al. Dopant-induced electron localization drives CO₂ reduction to C₂ hydrocarbons. Nat. Chem. 10, 974-980 (2018).
Liu, X., Schlexer, P., Xiao, J., Ji, Y., Wang, L., Sandberg, R. B., Tang, M., Brown, K. S., Peng, H., Ringe, S., Hahn, C., Jaramillo, T. F., Nørskov, J. K. & Chan, K. pH effects on the electrochemical reduction of CO₍₂₎ towards C₂ products on stepped copper. Nat. Commun. 32, 10 (2019).
Deng, H., Guo, C., Shi, P. & Zhao, G. Amino assisted protonation for carbon-carbon coupling during electroreduction of carbon dioxide to ethylene on copper(I) oxide, ChemSusChem 13, 4325-4333 (2021).
Li, H., Jiang, Y., Li, X., Davey, K., Zheng, Y., Jiao, Y. & Qiao, S.-Z. C₂₊ selectivity for CO₂ electroreduction on oxidized Cu-based catalysts. J. Am. Chem. Soc. 145, 14335-14344 (2023).
Yao, K., Li, J., Wang, H., Lu, R., Yang, X., Luo, M., Wang, N., Wang, Z., Liu, C., Jing, T., Chen, S., Cortés, E., Maier, S. A., Zhang, S., Li, T., Yu, Y., Liu, Y., Kang, X. & Liang, H. Mechanistic insights into OC-COH coupling in CO₂ electroreduction on fragmented copper. J. Am. Chem. Soc. 144, 14005-14011 (2022).
Chen, L., Tang, C., Jiao, Y. & Qiao, S.-Z. Anomalous C-C coupling on under-coordinated Cu (111): A case study of Cu nanopyramids for CO₂ reduction reaction by molecular modelling. ChemSusChem 14, 671-678 (2021).
Shen, H., Wang, Y., Chakraborty, T., Zhou, G., Wang, C., Fu, X., Wang, Y., Zhang, J., Li, C., Xu, F., Cao, L., Mueller, T. & Wang, C. Asymmetrical C-C coupling for electroreduction of CO on bimetallic Cu-Pd catalysts. ACS Catal. 12, 5275-5283 (2022).
Zhang, X. Y., Lou, Z. X., Chen, J. et al. Direct OC-CHO coupling towards highly C₂₊ products selective electroreduction over stable Cu⁰/Cu²⁺ interface. Nat. Commun. 14, 7681 (2023).
Pei, W., Zhou, S., Zhao, J., Xu, X., Du, Y. & Dou, S. X. Immobilized trimeric metal clusters: A family of the smallest catalysts for selective CO₂ reduction toward multi-carbon products. Nano Energy 76, 105049 (2020).
Chen, C., Yu, S., Yang, Y. et al. Exploration of the bio-analogous asymmetric C-C coupling mechanism in tandem CO₂ electroreduction. Nat. Catal. 5, 878-887 (2022).
Zhao, Q., Martirez, J. M. P. & Emily A. Carter, E. A. Charting C-C coupling pathways in electrochemical CO₂ reduction on Cu(111) using embedded correlated wave function theory. Proc. Natl. Acad. Sci. U.S.A. 119, e2202931119 (2022).
Chen, P.-C., Chen, C., Yang, Y., Maulana, A. L., Jin, J., Feijoo, J. & Yang, P. Chemical and structural evolution of AgCu catalysts in electrochemical CO₂ Reduction. J. Am. Chem. Soc. 145, 10116-10125 (2023).
Kuang, S., Su, Y., Li, M., Liu, H., Chuai, H., Chen, X., Hensen, E. J. M., Meyer, T. J., Zhang, S. & Ma, X. Asymmetrical electrohydrogenation of CO₂ to ethanol with copper-gold heterojunctions. Proc. Natl. Acad. Sci. U.S.A. 120, e2214175120 (2023).
Guo, P.-P., He, Z.-H., Yang, S.-Y., Wang, W., Wang, K., Li, C.-C., Wei, Y.-Y., Liu, Z.-T. & Han, B. Electrocatalytic CO₂ reduction to ethylene over ZrO₂/Cu-Cu₂O catalysts in aqueous electrolytes. Green Chem. 24, 1527-1533 (2022).
Feng, J., Wu, L., Liu, S., Xu, L., Song, X., Zhang, L., Zhu, Q., Kang, X., Sun, X. & Han, B. Improving CO₂-to-C₂₊ product electroreduction efficiency via atomic lanthanide dopant-induced tensile-strained CuO_x catalysts. J. Am. Chem. Soc. 145, 9857-9866 (2023).
Zhong, M., Tran, K., Min, Y., Wang, C., Wang, Z., Dinh, C. T., De Luna, P., Yu, Z., Rasouli, A. S., Brodersen, P., Sun, S., Voznyy, O., Tan, C. S., Askerka, M., Che, F., Liu, M., Seifitokaldani, A., Pang, Y., Lo, S. C., Ip, A., Ulissi, Z. & Sargent, E. H. Accelerated discovery of CO₂ electrocatalysts using active machine learning. Nature 581, 178-183 (2020).
Li, P., Bi, J., Liu, J., Wang, Y., Kang, X., Sun, X., Zhang, J., Liu, Z., Zhu, Q. & Han, B. p-d orbital hybridization induced by p-block metal-doped Cu promotes the formation of C₂₊ products in ampere-level CO₂ electroreduction. J. Am. Chem. Soc. 145, 4675-4682 (2023).
Cao, Y., Chen, S., Bo, S., Fan, W., Li, J., Jia, C., Zhou, Z., Liu, Q., Zheng, L. & Zhang, F. Single atom Bi decorated copper alloy enables C-C coupling for electrocatalytic reduction of CO₂ into C₂₊ products. Angew. Chem. Int. Ed. 62, e202303048 (2023).
Mou, T., Pillai, H. S., Wang, S. et al. Bridging the complexity gap in computational heterogeneous catalysis with machine learning. Nat. Catal. 6, 122-136 (2023).
Xu, L., Papanikolaou, K. G., Lechner, B. A. J., Je, L., Somorjai, G. A., Salmeron, M., Mavrikakis, M. Formation of active sites on transition metals through reaction-driven migration of surface atoms. Science 380, 70-76 (2023).
Amirbeigiarab, R., Tian, J., Herzog, A. et al. Atomic-scale surface restructuring of copper electrodes under CO₂ electroreduction conditions. Nat. Catal. 6, 837-846 (2023).
Li, H. & Reuter, K. Active-site computational screening: Role of structural and compositional diversity for the electrochemical CO₂ reduction at Mo carbide catalysts. ACS Catal. 10, 11814-11821 (2020).
Zhang, D., Li, H., Lu, H. et al. Unlocking the performance of ternary metal (hydro)oxide amorphous catalysts via data-driven active-site engineering. Energy Environ. Sci. 16, 5065-5075 (2023).
Li, H., Jiao, Y., Davey, K. & Qiao, S.-Z. Data-driven machine learning for understanding surface structures of heterogeneous catalysts. Angew. Chem., Int. Ed. 62, e202216383 (2023).
Tran, K., Ulissi, Z.W. Active learning across intermetallics to guide discovery of electrocatalysts for CO₂ reduction and H₂ evolution. Nat. Catal. 1, 696-703 (2018).
Pedersen, J. K., Batchelor, T. A., Bagger, A. & Rossmeisl, J. ACS Catal. 10, 2169-2176 (2020).
Batchelor, T. A. A., Pedersen, J. K., Winther, S. H. et al. Joule 3, 834-845 (2019).
Abild-Pedersen, F., Greeley, J., Studt, F. et al. Scaling properties of adsorption energies for hydrogen-containing molecules on transition-metal surfaces. Phys. Rev. Lett. 99, 016105 (2007).
Studt, F., Abild-Pedersen, F., Bligaard, T., Sorensen, R. Z., Christensen, C. H., Norskov, J. K. Identification of non-precious metal alloy catalysts for selective hydrogenation of acetylene. Science 320, 1320 (2008).
Zhao, Z.-J., Liu, S., Zha, S. et al. Theory-guided design of catalytic materials using scaling relationships and reactivity descriptors. Nat. Rev. Mater. 4, 792-804 (2019).
Greeley, J. Theoretical heterogeneous catalysis: Scaling relationship and computational catalyst design. Annu. Rev. Chem. Biomol. Eng. 7, 605-635 (2016).
Wen, M., Spotte-Smith, E. W. C., Blau, S. M. et al. Chemical reaction networks and opportunities for machine learning. Nat. Comput. Sci. 3, 12-24 (2023).
Margraf, J. T., Jung, H., Scheurer, C. et al. Exploring catalytic reaction networks with machine learning. Nat. Catal. 6, 112-121 (2023).
Chanussot, L., Das, A., Goyal, S., Lavril, T., Shuaibi, M., Riviere, M., Tran, K., Heras-Domingo, J., Ho, C., Hu, W. H., Palizhati, A., Sriram, A., Wood, B., Yoon, J., Parikh, D., Zitnick, C. L. & Ulissi, Z. Open catalyst 2020 (OC20) dataset and community challenges. ACS Catal. 11, 6059-6072 (2021).
Klicpera, J., Groß, J. & Günnemann, S. Directional message passing for molecular graphs. International Conference on Learning Representations (ICLR 2020) (2020).
Klicpera, J., Giri, S., Margraf, J. T. & Günnemann, S. Fast and uncertainty-aware directional message passing for non-equilibrium molecules. Machine Learning for Molecules Workshop NeurIPS (2020).
Schutt, K. T., Kindermans, P. J., Sauceda, H. E., Chmiela, S., Tkatchenko, A. & Muller, K. R. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in Neural Information Processing Systems 30 (NeurIPS 2017) 30 (2017).
Schutt, K. T., Sauceda, H. E., Kindermans, P. J., Tkatchenko, A. & Muller, K. R. SchNet - A deep learning architecture for molecules and materials. J. Chem. Phys. 148, 241722 (2018).
Kipf, T. N. & Welling, M. Semi-supervised classification with graph convolutional networks. International Conference on Learning Representations (ICLR 2017) (2017).
Su, D.-J., Xiang, S.-Q., Gao, S.-T., Jiang, Y., Liu, X., Zhang, W., Zhao, L.-B. & Tian, Z.-Q. Kinetic understanding of catalytic selectivity and product distribution of electrochemical carbon dioxide reduction reaction. JACS Au 3, 905-918 (2023).
Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 54, 11169 (1996).
Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 50, 17953 (1994).
Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 78, 1396 (1997).
Zhang, Y. & Yang, W. Comment on ‘Generalized gradient approximation made simple’. Phys. Rev. Lett. 80, 890 (1998).
Grimme, S., Antony, J., Ehrlich, S. & Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. J. Chem. Phys. 132, 154104 (2010).
Larsen, A. H., Mortensen, J. J., Blomqvist, J., et al. The atomic simulation environment - A Python library for working with atoms. J. Phys.: Condens. Matter 29, 273002 (2017).
Xu, K., Hu, W., Leskovec, J. & Jegelka, S. How powerful are graph neural networks? International Conference on Learning Representations (ICLR 2019) (2019).
Hamilton, W. L., Ying, Z. & Leskovec, J. Inductive representation learning on large graphs. Advances in Neural Information Processing Systems 30 (NIPS 2017) 1024-1034 (2017).

There is NO Competing Interest.

SIsubmitted.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Establishing C-C electro-coupling mechanism via big data analyses of reaction networks

Status:

Version 1

Abstract

Figures

Introduction

Results

Discussion

Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1