Tracking heterogeneous structural motifs and the redox behaviour of copper–zinc nanocatalysts for the electrocatalytic CO2 reduction using operando time resolved spectroscopy and machine learning

Copper-based catalysts are established catalytic systems for the electrocatalytic CO2 reduction reaction (CO2RR), where the greenhouse gas CO2 is converted into valuable industrial chemicals, such as energy-dense C2+ products, using energy from renewable sources. However, better control over the catalyst selectivity, especially at industrially relevant high current density conditions, is needed to expedite the economic viability of the CO2RR. For this purpose, bimetallic materials, where copper is combined with a secondary metal, comprise a promising and a highly tunable catalyst for the CO2RR. Nevertheless, the synergy between copper and the selected secondary metal species, the evolution of the bimetallic structural motifs under working conditions and the effect of the secondary metal on the kinetics of the Cu redox behavior require careful investigation. Here, we employ operando quick X-ray absorption fine structure (QXAFS) spectroscopy coupled with machine-learning based data analysis and surface-enhanced Raman spectroscopy (SERS) to investigate the time-dependent chemical and structural changes in catalysts derived from shape-selected ZnO/Cu2O nanocubes under CO2RR conditions at current densities up to −500 mA cm−2. We furthermore relate the catalyst transformations observed under working conditions to the catalytic activity and selectivity and correlate potential-dependent surface and subsurface processes. We report that the addition of Zn to a Cu-based catalyst has a crucial impact on the kinetics of subsurface processes, while redox processes of the Cu surface layer remain largely unaffected. Interestingly, the presence of Zn was found to contribute to the stabilization of cationic Cu(i) species, which is of catalytic relevance since Cu(0)/Cu(i) interfaces have been reported to be beneficial for efficient electrocatalytic CO2 conversion to complex multicarbon products. At the same time, we attribute the increase of the C2+ product selectivity to the formation of Cu-rich CuZn alloys in samples with low Zn content, while Zn-rich alloy phases result in an increased formation of CO paralleled by an increase of the parasitic hydrogen evolution reaction.

The working electrodes were prepared by mixing 12.5 mg of the catalyst powder with 2 mL of methanol and 80 µL of a Nafion solution (Sigma-Aldrich, ~ 5 wt% in a mixture of alcohols and water). The mixture was ultrasonicated for at least 30 min, and 2 mL of the catalyst dispersion were spray-coated on the microporous layer of a 5 cm x 5 cm carbon-based gas diffusion electrode (GDE, Sigracet 28BC), which was placed on a hot plate at 120°C.

Supplementary Note 2: surface enhanced Raman spectroscopy measurements
For the SERS measurements, a Si(100) wafer (520.5 cm -1 ) served as calibration substrate. Further, a near-infrared laser (Renishaw, RL785, λ = 785 nm, P max = 500 mW) was used as excitation source. The backscattered light was Rayleigh-filtered, and the Raman scattering was collected in the range of 100-1200 cm -1 with a grating of 1200 lines mm -1 and guided to a CCD detector (Renishaw, Centrus). For the operando measurements, the excitation source was focused on the surface of the sample and Raman scattering signals were collected with a water immersion objective (Leica microsystems, 63×, numerical aperture 0.9) protected from the electrolyte by a Teflon film (DuPont, film thickness of 0.013 mm) wrapped around the objective. The acquisition of each spectrum was performed with 5 s of exposure time. The Raman data were processed using the Renishaw WiRE 5.2 software. The spectra were baseline-subtracted using the intelligent spline feature of eighth order, and cosmic rays were removed.

Supplementary Note 3: electrochemical characterization
The following equation (Eq. S1) was used to convert the measured potentials (V) to the reversible hydrogen (RHE) electrode scale: The pH of the bulk electrolyte was determined to be 7.8.

Supplementary Note 4: XAS measurements and data analysis
The QXAFS (Quick X-ray absorption fine structure) measurements were conducted at the SuperXAS -X10DA beamline at the SLS synchrotron facility of the Paul Scherrer Institute. For the energy selection a channel-cut LN-cooled Si(111) monochromator was used, which was oscillating with 1 Hz frequency. Rh-coated collimating mirrors and Rh-coated focusing mirrors were used to reject higher harmonics and reduce the heat load on the sample. The intensity of the incident radiation was measured using ionization chambers filled with N 2 . Operando XAS measurements were performed in fluorescence mode using a PIPS detector. The measurements were performed at the Cu K-edge (8979 eV) and the Zn K-edge (9659 eV).
For operando XAS measurements, we used our home-built gas-fed cell, which is suitable for measurements at high current densities. The cell design is described in Ref. 3 Samples were spray coated on a GDE with an area of 1 cm 2 . The sample loading was optimized and the sample amount for spray coating on the GDE was increased by a factor of 2.4, compared to the sample amount used for electrocatalytic measurements, to ensure a sufficient signal to noise ratio at the Cu K-edge and the Zn K-edge, but simultaneously avoid self-absorption effects.
For the calibration, a Cu foil spectrum was collected at the beginning of each QXAFS scan before moving the cell into the X-ray beam. The data calibration was performed with the Pro-QEXAFS software 4 . Further data processing, data reduction, spectra normalization, averaging and linear combination analysis of the XANES spectra were performed using a set of home-built Wolfram Mathematica scripts. The EXAFS data were extracted from averaged XAS spectra using the Athena software. 5 Advanced EXAFS data processing was carried out using a neural network approach as discussed in Supplementary Note 5.

Supplementary Note 5: Machine learning -description, validation, and analysis at different current densities
To develop the neural network for the analysis of our EXAFS spectra for mixtures of Zn species, we followed a similar procedure as discussed in our previous works. 6,7 Here, the important difference is that instead of extracting the partial radial distribution functions (RDFs) from the EXAFS data for atoms of different types, we extracted the partial radial distributions corresponding to the different Zncontaining phases (see the discussion in the main text). Briefly, we designed a neural network (NN), a composite mathematical function , that ℎ( ( ), ⃗ )→ oxide phase, metallic fcc-like phase and metallic non-fcc-like phase of Zn, respectively. , and are the relative weights of the corresponding phases. As in our previous works, the EXAFS spectrum was discretized by performing the Morlet wavelet transformation, 6,8 and providing ( ) selected wavelet coefficients as an input for the NN. For the analysis in this work, we used a relatively short spectral range in k-space (between 3 and 8 Å -1 ). The total input vector for NN contained 738 points. The RDFs were approximated with a histogram of interatomic distances, calculated in the range between R min = 1.0 Å and R max = 5.5 Å, with the histogram bin width 0.04 Å. The NN output vector thus contains 336 points. In-between input and output layers, our NN contained three hidden layers with 1000 nodes each to ensure sufficient flexibility of our model.
The parameters of the NN nodes (weights and biases) were found during the training step, where we provided the theoretical spectra with a known structure as an input for the NN and tuned the parameters to ensure a good agreement between the NN outputs and the true RDFs for our models. For the training, we used linear combinations of EXAFS spectra, which were obtained from molecular dynamics (MD) and Monte Carlo (MC) simulations using empirical force field models. These simulations have been carried out as described in our previous work. 6 As we have demonstrated in Figure S14a, the obtained theoretical spectra resemble well the experimental data of reference compounds, giving us confidence that after the training using theoretical data, the NN will be able to recognize features in our experimental data of the investigated samples as well. Note that the RDFs ( Figure S14b), obtained in these simulations for ZnO, metallic Zn and CuZn brass alloy, are clearly distinct. For instance, while the contribution of the first coordination shell appears to be similar for metallic Zn and CuZn brass structures (featuring distorted hcp-type structure and fcc-type structure, respectively, both having 12 nearest neighbours around the central atom, see Figure S14c), the peaks in the RDF region between 4 and 6 Å are very different. This suggests that there is sufficient contrast between these structures so that their amounts in the mixture can be quantified reliably by our EXAFS analysis, if it is extended to the contributions of these distant coordination shells. We would like to stress here, that this is challenging to do in conventional EXAFS fitting, but it is straightforward in NN-EXAFS approach.
For the NN training, the numbers of the theoretical spectra that we used are the following: 1512 spectra calculated for metallic Cu, 1804 spectra for Cu 2 O, 1708 spectra for CuO, 1231 spectra for Cu(OH) 2 , 1512 spectra for hcp-type Zn, 1320 spectra for wurtzite-type ZnO, 772 spectra for rocksalttype ZnO. Different spectra were obtained by changing the temperature and lattice spacings in the MD/MC models to extend the applicability of our approach to a broad range of metals, oxides, and alloys with different degree of disorder and different lattice constants. These unique spectra were further combined linearly to form 60000 model spectra for mixtures, which then were used for the NN training. We explained the technical details of the NN implementation and training in our previous work 5 . To estimate the uncertainties, we created 10 NNs, which were trained independently on different training sets. The results that are reported in this study are an average result, given by these 10 NNs, and the standard deviation of their predictions is used as an estimate of the random uncertainty of the NN-EXAFS method.
After the training was completed, the NN parameters were fixed, and the NN could be used to process experimental EXAFS spectra. Before applying the NN to the interpretation of our operando data, we first tested the NN on experimental data for well-defined reference compounds, for which the particular RDFs can be obtained independently from the EXAFS data using the reverse Monte Carlo (RMC) approach. [9][10][11] As shown in Figure S15ab, the RDFs, which were obtained by the NN-EXAFS method, are in an excellent agreement with the RMC results in the cases where we applied the method to EXAFS data for fcc metals with different lattice constants (metallic Cu, Ni, CuZn brass), EXAFS data for metals with distorted hcp-structure (metallic Zn), as well as to EXAFS spectra in different oxides (Cu 2 O, CuO, ZnO). Furthermore, to test the applicability of our method for the speciation of mixtures, we constructed artificial linear combinations of experimental Zn K-edge EXAFS spectra for Zn metal, wurtzite-type ZnO, and CuZn, where the spectra were added with random weights. We then provided these spectra as an input to our NN and estimated the mixing weights of these three components by using Eq. (2) in the main text. As one can see from Figure S15cde, the NN results are in a good agreement with the known true values of the weights for oxide, fcc-and hcp-phases in the mixtures. We estimated that the average deviation between the true value of the species' concentration and the NN result is ca. 4.4% for the oxide phase, 7.7% for the fcc phase, and 7.6% for the hcp phase. These numbers represent the expected systematic uncertainty of our NN-EXAFS method for the speciation of Zn phases.
Examples of RDFs that were reconstructed for these mixtures of reference spectra, are shown in Figure S16. The obtained RDFs are compared here with the linear combination of RDFs, obtained for the reference structures by RMC method, where the combined RDFs are weighted with the same weights as reference spectra in the analyzed mixture spectrum. From this figure, one can confirm that not only the weights of different phases are reproduced with reasonable accuracy, but also the RDFs themselves. This indicates that we can reliably extract information on the distribution of interatomic distances from our NN-EXAFS method. Therefore, our validation procedure thoroughly demonstrates that the changes in interatomic distances due to alloying/dealloying processes can be tracked accurately by our approach.
After the described training and validation of the NN with experimental data for reference compounds, this method then can be applied to the interpretation of operando QXAFS spectra for our catalysts.

Supplementary Note 6: Interpretation of CV features corresponding to catalyst oxidation
In the performed cyclic voltammetry (CV) scans, we observed several features that can be attributed to the oxidation of the catalyst species (see Figure 7 in the main text). The main oxidation features for Cu 2 O catalyst are marked as B1 and B2 and correspond to the CV peaks visible at ca. 0.7 V in the anodic scan (B1), and at ca. 0.5 V (B2). The main features in the anodic part of CVs for the bimetallic catalysts are marked as A1 and A2. A1 corresponds to the region between -0.4 V and 0 V in the anodic scan, where two small split peaks are observed in the CVs for samples with 4% and 7% Zn loading, and one broad peak for the 15% Zn containing sample. A2 denotes the intense peak followed by a current plateau from 0.5 V until the upper vertex in the upward scan. To identify these features, we used XAS and LCA-XANES and NN-EXAFS analyses.
Feature A1 in the CVs for bimetallic samples appears simultaneously with the contribution of Zn(II) in LCA-XANES data as well as with the appearance of the oxide structure contribution in the NN-EXAFS data of the Zn K-edge. This feature can thus be linked to the oxidation of Zn species in a CuZn alloy and a metallic phase. Indeed, this feature is not present in the CV for pure Cu 2 O NCs, and its intensity increases systematically with the Zn loading ( Figure S20). Moreover, the appearance of this feature also coincides with the onset of the decrease of the Zn fluorescence intensity ( Figure S21). We attribute the decrease of the Zn fluorescence signal to the dissolution of newly formed cationic Zn species, which are expected to be unstable at these potentials. 12 Both, the metallic Zn phase and the CuZn alloy phases with different Cu to Zn ratios are affected by the oxidation of the sample. Thus, Zn species with different local structure and charge state will be present, which likely explains why the A1 peak is broad and split in several small peaks for samples with lower Zn loading. The NN-EXAFS analysis also shows that the oxidation of Zn is accompanied by a shift of the RDF peaks in both, fcc and nonfcc phases, to lower interatomic distances ( Figure S25), suggesting that the oxidation is accompanied by dealloying processes, forming increasingly copper-rich phases. Note that at these potentials the Cu species remain metallic.
Features B2 and A2 are among the most pronounced features in the CVs for pure Cu 2 O, and for the Zn-decorated samples, respectively. The onset potentials of B2 and A2 peaks are similar, but the A2 peaks are much broader for Zn-decorated samples. In both cases, we attribute this feature to the oxidation of metallic Cu to the Cu(I) state. We note, nonetheless, that the oxidation of Cu to the Cu(II) state also contributes to this feature. Indeed, our LCA-XANES results show that for the monometallic Cu sample, the signatures of Cu(II) species appear almost simultaneously with those of Cu(I) species, suggesting that oxidation of Cu(0) species to Cu(I) and Cu(II) takes place simultaneously. At the same potential, we also observe the onset of the decrease in the Cu fluorescence signal ( Figure S27), which based on our previous works can be assigned to the dissolution of newly formed, unstable Cu(II) species. 3,12 We note that the decrease of Cu fluorescence continues during the CV scan, until the potential is lowered and Cu(II) species are reduced. The remaining Cu(I) and Cu(0) species, unlike Cu(II) species, are stable and are not dissolving.
One can note that the oxidation of Cu proceeds clearly differently in the bimetallic samples than in the monometallic sample. First of all, the overall amount of Cu(I) species formed during the CV scan at potentials above 0.5 V is much lower in the bimetallic catalysts than for the pure Cu 2 O NCs. An increased Zn content, however, increases the amount of Cu(I) species formed in the bimetallic samples. This can be related to the higher contribution of residual Cu(I) species with increasing Zn loading at more negative potentials before and after the oxidation cycle (for pure Cu 2 O NCs: 4% Cu(I), while for the sample containing 15% of Zn: 12% Cu(I)). Moreover, the appearance of the Cu(II) contribution is slightly delayed in the bimetallic samples. We attribute both these effects to the competition between the oxidation of Zn species and Cu species. Indeed, in bimetallic catalysts, the Cu oxidation takes place on the background of continuous Zn oxidation. Moreover, the feature A2 further marks the onset of a significant boost in the oxidation of Zn species, as visible from our LCA-XANES and NN-EXAFS results. We explain this sharp increase in the Zn oxidation by the fact that the oxidation of Cu exposes the Zn species to oxidative conditions, which were so far protected from the oxidation by being incorporated in the bulk of CuZn alloy phases. Simultaneously, the oxidation of these Zn species could delay the oxidation of Cu to the Cu(II) state. Interestingly, while the appearance of Cu(II) species is shifted to a higher potential in the bimetallic catalysts, the amount of Cu(II) generated is larger than in the monometallic catalyst and systematically increases with Zn loading. This effect can be tentatively attributed to the fact that the lower contribution of Cu(I) species observed in the bimetallic catalysts leaves more metallic Cu sites to be directly oxidized to the Cu(II) state. The combination of these three aforementioned effects in our bimetallic catalysts: (i) a more pronounced Cu oxidation to Cu(II) state at higher Zn loadings, (ii) the shift of the onset of Cu oxidation to Cu(II) species to higher potential with respect to that in the monometallic catalysts, and (iii) the additional contribution of Zn species oxidation explains why the A2 feature in our bimetallic catalysts is much broader than the corresponding B2 feature in the monometallic Cu 2 O catalyst.
For the discussion of CV features observed during the cathodic part of CV scan the discussion can be found in the main text. Figure S1. STEM-EDX image of a Cu 2 O NC sample in as-prepared state with the area indicated that was used for the quantification of the Cu:O ratio.  Table S1. Edge lengths and the corresponding size distribution of Cu 2 O NCs obtained from the analysis of STEM-HAADF images and the dimension of the Zn shells/islands obtained from the analysis of the corresponding EDX maps of the bimetallic samples containing 4%, 7% and 15% Zn. Data for each sample are included in their as-prepared state as well as after CO 2 RR at -500 mA/cm². For the CuZn sample containing 15% Zn, no formation of Zn islands was observed after the exposure to CO 2 RR conditions.           The analyzed spectra were collected during the first 10 seconds of the CO 2 RR at a current density of -500 mA/cm² (e-h). The Cu K-edge XANES spectra of the sample with 4% Zn loading may suffer from self-absorption. The analyzed spectra were collected during 1 hour of CO 2 RR at a current density of -10 mA/cm² (a-d) and -500 mA/cm² (e-h). The Cu K-edge XANES spectra of the sample with 4% Zn loading that was collected at a current density of -500mA/cm² may suffer from self-absorption, and therefore, was not included in this figure.   Figure S17. The time-dependent evolution of the RDFs obtained by the NN-EXAFS analysis from operando Zn K-edge EXAFS data for bimetallic CuZn catalysts with (a) 15%, (b) 7% and (c) 4% Zn content during CO 2 RR at a current density of -10 mA/cm 2 . The partial RDFs corresponding to an oxide phase, a metallic phase with fcc-type structure and a metallic phase with non-fcc type structure are shown. The RDFs are shifted vertically for clarity and the RDFs for reference materials (ZnO, brass foil and Zn foil) are shown for comparison. Figure S18. The evolution of the relative concentrations of oxide, fcc-like and hcp-like phases, as obtained by the NN from operando Zn K-edge EXAFS data for bimetallic catalysts with a Zn loading of (a,d) 15%, (b,e) 7% and (c,f) 4% during CO 2 RR at -10 mA/cm 2 (a-c) and -500 mA/cm 2 (d-f) current density.          In panel a, the sample was accidentally exposed to higher potentials, where the Zn(II) species are unstable. Therefore, we suspect that the amount of Zn present during the reduction cycle might be less than 4% in this sample. This, however, should not affect the main CV features, corresponding to the oxidation of copper as well as the reduction of surface Cu species.