On the origin of the elusive first intermediate of CO₂ electroreduction

Significance

The understanding of a catalytic reaction starts with understanding its first elementary step. Surprisingly, despite the large number of studies, it is unclear whether one common or two different first intermediates control the selectivity of CO₂ electroreduction to formate and CO. We settle this controversy for Cu, which is best known for its unique capacity to synthesize C₁₊ products but is just emerging as a superior earth-abundant catalyst for CO and formate. We provide solid experimental and theoretical support of the one common first-intermediate (Hori’s) model, the first intermediate being carboxylate. This outcome is an essential milestone toward accurate specification of the reaction descriptors in the growing effort to accelerate the discovery of a viable CO₂ electroreduction catalyst.

Abstract

We resolve the long-standing controversy about the first step of the CO₂ electroreduction to fuels in aqueous electrolytes by providing direct spectroscopic evidence that the first intermediate of the CO₂ conversion to formate on copper is a carboxylate anion *CO₂⁻ coordinated to the surface through one of its C–O bonds. We identify this intermediate and gain insight into its formation, its chemical and electronic properties, as well as its dependence on the electrode potential by taking advantage of a cutting-edge methodology that includes operando surface-enhanced Raman scattering (SERS) empowered by isotope exchange and electrochemical Stark effects, reaction kinetics (Tafel) analysis, and density functional theory (DFT) simulations. The SERS spectra are measured on an operating Cu surface. These results advance the mechanistic understanding of CO₂ electroreduction and its selectivity to carbon monoxide and formate.

The electrocatalytic conversion of abundant CO₂ to fuels and other useful chemicals has attracted significant interest in the past decade as a promising complementary way to store excessive renewable energy (1, 2). However, commercialization of this green technology is hampered by the absence of a viable catalyst. Even though the catalyst discovery can significantly be accelerated using the emerging approach of computational catalyst design (3), its applicability to CO₂ electroreduction has been limited mostly to explaining experimental trends, rather than to predicting new materials (4–6). This situation roots from a poor mechanistic understanding of this reaction, which starts with a poor understanding of its very first step—activation of CO₂.

In fact, it is currently unclear whether CO₂ electroreduction on different materials starts with one common or several different first intermediates and what their structures are. Specifically, Hori et al. (7, 8) have postulated that, on metals such as Au, Cu, and Ag with medium to high overpotentials for hydrogen evolution reaction (HER), CO₂ is reduced through one common first intermediate, which is carboxylate *CO₂⁻ (* denotes the surface-coordinated state of the ligand):

$$\mathrm{C}{\mathrm{O}}_2+{\mathrm{e}}^{-}=*{{\mathrm{CO}}_2}^{-}.$$ [1]

This first step is supported by the reaction kinetics analysis (7, 9–15).

If so, it is still unclear what the structure of *CO₂⁻ is and how it controls the reaction pathway. As a ligand, CO₂ is very versatile in terms of its coordination to metal centers (16, 17). In the context of CO₂ electroreduction, the most often cited carboxylate structures include η²(O,O)-CO₂⁻ (O,O-down) (18–20), η¹(C)-CO₂⁻ (C-down) (21), and η²(C,O)-CO₂⁻ (C,O-down) (22–24) (Fig. 1). As proposed by Hori et al. (7, 8), strongly adsorbed η¹(C)-CO₂⁻ or η²(C,O)-CO₂⁻ can be converted to CO after they are protonated to carboxyl *COOH (Fig. 1):

$$*{{\mathrm{CO}}_2}^{-}+{\mathrm{H}}^{+}=*\mathrm{COOH},$$ [2]

which reductively dissociates to CO and H₂O upon proton coupled electron transfer (PCET):

$$*\mathrm{COOH}+{\mathrm{e}}^{-}+{\mathrm{H}}^{+}=\mathrm{CO}+{\mathrm{H}}_2\mathrm{O}.$$ [3]

In contrast, the reaction can be directed toward formate if *CO₂⁻ is adsorbed through oxygen(s) or weakly through carbon. This structure allows the carbon atom to be hydrogenated by either a direct Tafel-like reaction with surface hydride *H,

$$*{{\mathrm{CO}}_2}^{-}+*\mathrm{H}={\mathrm{HCOO}}^{-},$$ [4a]

or by PCET,

$$*{{\mathrm{CO}}_2}^{-}+{\mathrm{e}}^{-}+{\mathrm{H}}^{+}={\mathrm{HCOO}}^{-}.$$ [4b]

An alternative is a more contemporary model that has largely been drawn from density functional theory (DFT) simulations (5, 6, 22, 25, 26). It postulates that CO and formate have different precursors—carboxyl *COOH and O (or O,O)-coordinated formyloxyl *OCHO, respectively (Fig. 1). However, there is no consensus about how these two species are formed. According to Nørskov and coworkers (5, 26), carboxyl and formyloxyl are formed through PCET:

$$\mathrm{C}{\mathrm{O}}_2+{\mathrm{e}}^{-}+{\mathrm{H}}^{+}=*\mathrm{COOH}$$ [5]

and

$$\mathrm{C}{\mathrm{O}}_2+{\mathrm{e}}^{-}+{\mathrm{H}}^{+}=*\mathrm{OCHO},$$ [6]

respectively. In contrast, Goddard and coworkers (22, 25) adhere to the [1] and [2] pathway for carboxyl, and the hydride transfer (Eley–Rideal) mechanism for formyloxyl:

$$\mathrm{C}{\mathrm{O}}_2+*\mathrm{H}=*\mathrm{OCHO}\mathrm{.}$$ [7]

The third possibility can be the direct two-electron reduction of adsorbed bicarbonate HCO₃⁻ (27, 28).

Fig. 1.

Possible first intermediates of CO₂ electroreduction in aqueous electrolytes.

The aforementioned uncertainty is explained by the phenomenological complexity of CO₂ electroreduction, coupled with intrinsic limitations of the electrochemical and computational methods. Electrochemical surface-enhanced vibrational spectroscopy has not helped much in clarifying this issue either, even though this method has already established itself as a powerful tool to magnify our view of electrified interfaces (29). The only surface species unambiguously identified in aqueous electrolytes on Cu (30–35), Au (14), and Ag (36, 37) is carbon monoxide, while the structure of the other surface species remains highly controversial. This situation stems from the fact that the adsorbate structures have mostly been proposed based on the similarity of their one C–O stretching peak with a peak theoretically predicted or observed in other systems without corroborating by other spectral peaks and the ¹³C/¹²C and D₂O/H₂O isotope exchange effects. This approach is evidently reductionistic as it neglects the strong effects of the catalyst surface, interfacial electric field (including ionic pairing with a counterion), and hydration on the adsorbate spectrum, which can lead to great confusion (38, 39). Moreover, frequencies of the C–O stretching vibrations involve the same CO₂ moiety. Hence, their values are not characteristic of the species shown in Fig. 1. Another source of confusion can be uncontrollable organic contaminations, which can be spotted by the ¹³C/¹²C isotope exchange effect. Finally, to get surface enhancement, the earlier studies have been performed on model roughened surfaces, without validating their catalytic activity. As a result, even if the CO₂-derived adsorbates are identified correctly, it does not warrant that the reaction intermediate is among them because the latter can be formed only on an operating catalyst.

In our study, we focus on Cu because it is one of the most effective pure metal catalysts of CO₂ electroreduction, along with Ag, Pd, and Au (18). At the same time, Cu stands out of this group due to its low cost and unique capacity to synthesize C₁₊ products. The interest in Cu has additionally been elevated by recent findings of its remarkably high efficiency in the synthesis of CO and formate (12, 40, 41)—two of the most feasible targets of the CO₂ conversion technology (2).

Paradoxically, among the large number of studies of CO₂ electroreduction on Cu, only three have attempted to gain insight into the first step of this reaction (12, 22, 26). In particular, Li and Kanan (12) have concluded on the basis of the experimental Tafel slope that the CO synthesis on Cu starts with the formation of carboxylate *CO₂⁻ [1]. This result is supported by a quantum chemical study of Goddard and coworkers (22), which predicts that CO₂ is converted to CO on Cu(100) through steps [1]–[3]. In contrast, according to Nørskov and coworkers (26), Cu(111) and Cu(100) prefer to synthesize formate because these surfaces provide a more favorable thermodynamics for the formation of formyloxyl [6] over carboxyl [5]. Thus, the chemical identity of the first intermediate of the CO₂ electroreduction on Cu and hence the nature of the selectivity of Cu to CO and formate remains elusive.

To bridge this gap, we take advantage of SERS. In contrast to its more often used rival, surface-enhanced infrared absorption spectroscopy (SEIRAS), SERS can be measured on an actual catalyst. In addition, SERS can access a much broader spectral range, which allows for more reliable interpretation of the adsorbate structure. Moreover, SERS spectra are absolute, while SEIRAS spectra are differential and therefore are distorted by the negative peaks of the species present on the electrode surface at the reference potential.

To verify our assignment of SERS peaks, we use the D₂O/H₂O and ¹³C/¹²C isotope exchange, as well as electrochemical Stark effects. The latter presents the shift of the vibrational frequency of the adsorbate with electrode potential. To the first order of the theory, the frequency shift $\mathrm{\Delta }\nu$ by electric field $\overrightarrow{F}$ can be written as $\mathrm{\Delta }\nu \approx -\left|\mathrm{\Delta }\overrightarrow{\mu }\right|\cdot \left|\mathrm{\Delta }\overrightarrow{F}\right|\cdot \cos \phi$, where $\overrightarrow{\mathrm{\Delta }F}$ is the change in the local electric field, $\mathrm{\Delta }\overrightarrow{\mu }$ is the change in the vibrating dipole moment between the ground and first excited vibrational states, and $\phi$ is the angle between the vectors of the dipole and the field (42). Hence, the sign of the Stark effect characterizes the orientation of the adsorbate with respect to the applied electric field $\overrightarrow{F}$. A change in the Stark tuning rate ($\mathrm{\Delta }\nu /\mathrm{\Delta }E$, where $\mathrm{\Delta }E$ is a change in the electrode potential) suggests that the adsorbate or its local environment is perturbed by a certain surface reaction, which can help identify peaks that characterize the same adsorbate. We further corroborate the spectroscopy-based conclusions with the Tafel analysis of the reaction kinetics and with DFT simulations.

Using this advanced approach, we validate Hori’s hypothesis that CO₂ electroreduction starts with a common first intermediate, which is carboxylate. We identify the structure of this intermediate as η²(C,O)-CO₂⁻, find that it can be formed at potentials significantly anodic of the onset of CO₂ electroreduction, and conclude that the electrocatalytic activity of the metal ties with strengths of the metal–C and metal–O bonds of η²(C,O)-CO₂⁻.

Materials and Methods

A roughened Cu electrode was prepared from a 0.5-mm-thick copper foil (99.9985% metals basis) by one oxidation–reduction cycle as reported in SI Appendix, section S1. The surface morphology was assessed using SEM (SI Appendix, Fig. S5). The roughness factor of 17 ± 3 of the surface was estimated using the dependence of the electric double-layer capacitance on the scan rate in cyclic voltammograms (CVs) (SI Appendix, Fig. S6). The surface crystallographic composition, as inferred from the hydroxyl adsorption peaks in CVs, is dominated by the (100), (110), and (111) planes (SI Appendix, Fig. S7A). The electrocatalytic activity of the Cu surface was characterized using CVs (SI Appendix, Fig. S7B) and preparative electrolysis (SI Appendix, Fig. S8). The electrolysis products were measured using NMR and gas chromatography. SERS spectra were acquired in a laboratory-made one-compartment three-electrode spectroelectrochemical cell using a LABRAM-ARAMIS confocal Raman microscope with a HeNe (633-nm) laser. DFT calculations were performed using a plane wave DFT code (The Vienna Ab initio Simulation Package, VASP-5.3.3). Details of the experimental and theoretical methods are provided in SI Appendix, section S1.

Results and Discussion

Catalytic Activity of the Roughened Cu Surface.

All potentials are reported in the Ag/AgCl reference scale. The CO₂ reduction activity of the Cu surface was characterized using preparative electrolysis in CO₂-saturated 0.1 M NaHCO₃ (pH 6.8) at potentials from −1.05 to −1.27 V, where we detected synthesis of formate (43),

$$\mathrm{C}{\mathrm{O}}_2+2{\mathrm{e}}^{-}+{\mathrm{H}}^{+}={\mathrm{HCOO}}^{-},{\mathrm{E}}_0=-0.63\mathrm{V}\left(\mathrm{Ag}/\mathrm{AgCl}\right),$$ [8]

at a Faradaic efficiency from 1 to 1.7% (SI Appendix, Fig. S8A), although traces of formate were found at −0.95 V. At the same time, only H₂ (no CO) was detected in the gas phase. The Tafel slope of reaction [8] is 130 ± 10 mV/decade (SI Appendix, Fig. S8B), which suggests that its rate-determining step (RDS) is step [1] (SI Appendix in ref. 15), as in the case of the CO synthesis (12). CVs show that the electrocatalytic current, which is dominated by HER, starts at approximately −0.8 V (SI Appendix, Fig. S7B).

We deliberately selected this particular Cu surface and electrolyte for our study because more catalytically active surfaces and conditions coproduce more coadsorbed CO and other species, which complicate interpretation of the spectra. Moreover, the pathway of CO₂ to formate on Cu is much more uncertain than to CO (Introduction).

Operando SERS.

To validate reproducibility, we measure the dependence of SERS spectra on potential four times, each time on a freshly prepared Cu surface. We measure the spectra on an actual catalyst and report their replicates.

Comparison of the spectra measured in CO₂-saturated 0.1 M NaHCO₃ (Fig. 2 and SI Appendix, Figs. S11, S13, and S15) with the reference spectra measured in Ar-saturated 0.1 M NaClO₄ (SI Appendix, Fig. S9) reveals that the CO₂-saturated conditions give birth to a multifold of new peaks. In the following, we assign the main peaks to *CO, *CO₃²⁻, and *CO₂⁻, specify structures of these species, and show that the effect of potential on *CO₂⁻ agrees with the status of this species as the first intermediate.

Fig. 2.

Operando SERS of rough Cu surface in CO₂-saturated in 0.1 M NaHCO₃ (pH 6.8). Spectra were measured from −0.1 V toward the cathodic direction. Panels on the Right show the enlarged and scaled by intensity peaks of adsorbed *CO₃²⁻ and *CO₂⁻. The enlarged peak of *CO is shown in SI Appendix, Fig. S10. The color code in the Right panels is the same as in the main figure. The intensity of the spectrum measured at −1.2 V is underestimated because the SERS signal dropped due to HER. For the sake of simplicity, we label the vibrations of adsorbed ligands as the parent vibrations of the corresponding free ligands.

Carbon monoxide.

Among the adsorbates formed on Cu under the CO₂-saturated conditions, the easiest to identify is *CO. Its weak C≡O stretching peak, νCO, is observed at 2,050–2,070 cm⁻¹ starting from −0.9 V (Fig. 2 and SI Appendix, Fig. S10), that is, only in the electrocatalytic region. The peak position corresponds to the CO molecules linearly coordinated to atop Cu sites (30, 31, 44, 45). This peak is negligibly perturbed by the D₂O/H₂O isotope exchange (SI Appendix, Fig. S16) and hence is not confused with the νCu–H peak of adsorbed hydrogen reported in this region (31). We are unable to distinguish the corresponding Cu–CO frustrated rotation and Cu–CO translation expected at ∼280 and 360 cm⁻¹, respectively (ref. 46 and SI Appendix in ref. 47). The 360-cm⁻¹ peak is not observed even when the relative νCO intensity is higher by a factor of 3 (SI Appendix, Fig. S16). The absence of the low-frequency Cu–CO peaks of *CO in some cases has been noted earlier (SI Appendix in ref. 47) but is not understood as of yet.

The νCO peak of *CO is subject to the electrochemical Stark effect with a Stark tuning rate of 40 ± 5 cm⁻¹/V (Fig. 3A), typical of CO chemisorbed on atop sites (44, 45). The positive sign of the rate (red shift) reflects that the C≡O dipole is stretched out by the negative electric field (directed toward the surface) (scheme in Fig. 3A).

Fig. 3.

Effect of electrode potential on the frequencies and intensities of (A) νCO of *CO, (B) ν₁CO₃²⁻ of *CO₃²⁻ (C) ν_asCO₂⁻ of *CO₂⁻ and (D) νCu–C of *CO₂⁻ in the spectra shown in Fig. 2. The black arrows in the schemes of the corresponding vibrations show the directions of their dipole moments $\overrightarrow{\mu }$. $\overrightarrow{F}$ is the local (negative) electric field at a negatively charged electrode. The shadowed areas show the region where cathodic current is observed (SI Appendix, Fig. S7B). Oxygen is red. Carbon is gray. Copper is blue. The potential is additionally reported vs. the reversible hydrogen electrode (RHE) scale, which is calculated as E(RHE) = E(Ag/AgCl) + 0.209 + 0.059pH.

To conclude, even though the CO level in the gas phase is below the detection limit, CO is still formed in a small amount on the Cu surface.

Carbonate.

At −0.3 V and more negative potentials, the SERS spectra are dominated by a strong peak at ∼1,065 cm⁻¹ (Fig. 2). This peak has earlier been detected at 1,050–1,056 cm⁻¹ on Cu in gas phase (46) and under electrochemical conditions (33–35), where it has been assigned to adsorbed carbonate. Below, we refine the peak assignment to the totally symmetric C–O stretching vibration, ν₁CO₃²⁻, of a weakly adsorbed monodentate carbonate η¹-CO₃²⁻ (scheme in Fig. 3B).

To start with, the 1,065-cm⁻¹ peak is in the characteristic frequency range of the ν₁CO₃²⁻ vibration of carbonate (38, 48, 49). This vibration presents centrosymmetric movements of the CO₃²⁻ oxygen atoms with respect to the central carbon atom (scheme in Fig. 3B). In agreement with this assignment, the 1,065-cm⁻¹ peak is negligibly affected by the ¹³C/¹²C and D₂O/H₂O isotope exchange (Fig. 4 and SI Appendix, Figs. S16 and S17).

Fig. 4.

Effect of the isotope exchange on SERS spectra of species adsorbed on the roughened Cu electrode at −0.5 V (Ag/AgCl) in CO₂-saturated 0.1 M NaHCO₃: (A) ¹³C/¹²C exchange and (B) D₂O/H₂O exchange. Results at −0.3 and −1.1 V are shown in SI Appendix, Fig. S17.

Next, the dependence of the ν₁CO₃²⁻ peak on potential indicates that the carbonate in question is chemisorbed. First, even though the peak intensity decreases, it remains prominent at potentials cathodic of −0.9 V (Fig. 3B). At these potentials, the electrocatalytic current is observed (SI Appendix, Fig. S7B) while the Cu surface is expected to be negatively charged (50). Moreover, at potentials from −0.3 to −0.9 V where the Cu surface becomes less positively charged, the peak intensity increases. Another indicator of chemisorption is the susceptibility of the ν₁CO₃²⁻ peak to the electrochemical Stark effect (Fig. 3B). Specifically, this peak red shifts at an initial rate of 5 cm⁻¹/V, which increases stepwise up to 12 cm⁻¹/V at −0.9 V. Both the shift and the break are reproduced qualitatively in SI Appendix, Figs. S12B and S14B. The Stark shift, which reflects a change in the local electric field of the vibrating dipole, is typically considered as an indicator of the chemisorbed state of an interfacial species (51). The reason is that such a species is exposed to a stronger interfacial electric field than the physisorbed one, which is screened from the electric field by the hydration shell. Chemisorption can also explain why the adsorbed species is CO₃²⁻ while the dominant carbonaceous species in the electrolyte is HCO₃⁻.

Now we look at the position of the ν₁CO₃²⁻ peak of the chemisorbed carbonate, which is at 1,060–1,068 cm⁻¹ (Fig. 3B). Its proximity to ν₁CO₃²⁻ of dissolved carbonate (1,063 cm⁻¹) (SI Appendix, Fig. S18) indicates that the adsorption strength of this carbonate is relatively weak. For comparison, strongly coordinated carbonates with bridging/bidentate and monodentate structures are characterized by ν₁CO₃²⁻ below 1,040–1,050 cm⁻¹ (38, 48, 49). A similarly weak perturbation of the ν₁CO₃²⁻ peak is observed for a carbonate weakly chemisorbed at the Fe(III)(hydr)oxide–water interface (39). This species has a η¹-CO₃²⁻ structure in which the noncoordinated oxygen(s) atom is involved in strong noncovalent interactions (ionic paring) with a coadsorbed hydrated alkali metal cation or hydronium (39). Hence, it is plausible that the weakly chemisorbed carbonate on Cu has a similar structure.

Apart from the ν₁CO₃²⁻ peak, η¹-CO₃²⁻ is expected to generate two ν₃C–O peaks that originate from the doubly degenerate C–O stretching vibration ν₃CO₃²⁻ at ∼1,400 cm⁻¹ of free CO₃²⁻ (D_3h) (SI Appendix, Fig. S18). We associate these peaks with weak features at 1,350 and 1,430–1,450 cm⁻¹, which are resolved in the spectra where the masking peaks of the coadsorbed carboxylate are suppressed (SI Appendix, section S4 and Figs. S19C, S20, and S21). Noteworthy is that earlier SEIRAS studies have assigned the high-frequency ν₃C–O vibration of adsorbed carbonate to a peak at ∼1,540 cm⁻¹, which is most pronounced at potentials anodic of the electrocatalytic region (30–32, 35). This peak is also observed by SERS (Fig. 2 and SI Appendix, Figs. S11, S13, and S15). However, as discussed in detail in SI Appendix, section S4, we rule out this assignment because the 1,540-cm⁻¹ peak is suppressed under certain conditions while the accompanying ν₁CO₃²⁻ peak at 1,065 cm⁻¹ remains prominent (SI Appendix, Figs. S19–S21). In addition, the carbonate origin of the 1,540-cm⁻¹ peak is in apparent conflict with the ¹³C shift and the spectral properties of the ν₁CO₃²⁻ peak. We assign the 1,540-cm⁻¹ peak to the coadsorbed carboxylate in Carboxylate.

As of now, carbonate adsorption on the metals that are active in CO₂ reduction is poorly understood. Since carbonate is a hard base, it is anticipated to form mostly ionic (electrostatic) bonds with the metal. The most straightforward pathway to form such a bond is through ligand exchange of HCO₃⁻ with adsorbed water/hydroxyl followed by deprotonation (38, 52). Alternatively, η¹-CO₃²⁻ can be formed by activation of CO₂ on the surface oxygen/OH⁻ sites or *CO₂⁻ (the Freund–Messmer mechanism) (46, 48, 49, 53), suggesting that η¹-CO₃²⁻ and *CO₂⁻ establish surface equilibrium. Carbonate adsorption can be stabilized by surface defects (53). The break in the Stark tuning rate of the carbonate peak in the electrocatalytic region can be attributed to a decrease in the local surface concentration of η¹-CO₃²⁻ by negative surface charge and hydrogen adsorption. A decrease in the local surface concentration of randomly adsorbed dipoles pulls their collective vibrational frequency down because it makes their dipole–dipole coupling weaker (54).

Obviously, further experimental and DFT studies are required to shed light on the carbonate adsorption on CO₂ electroreduction catalysts.

Carboxylate.

In the context of the CO₂ electroreduction, of particular interest are peaks at ∼1,540 and 350 cm⁻¹, which are best resolved at potentials from −0.3 to −0.9 V (Fig. 2 and SI Appendix, Figs. S11, S13, and S15). These peaks also persist at potentials cathodic of approximately −0.9 V where the Cu surface synthesizes formate (SI Appendix, Fig. S8), albeit the 350-cm⁻¹ peak has a very low signal-to-noise ratio of 2–3 (Fig. 2, Insets and SI Appendix, Fig. S11, Insets). The 1,540-cm⁻¹ peak has earlier been assigned to adsorbed carbonate (30–32, 35). We confidently discard this assignment in view of the arguments presented in SI Appendix, section S4. In the following, we prove that the 1,540 and 350-cm⁻¹ peaks are vibrational fingerprints of η²(C,O)-CO₂⁻ (Fig. 1).

Before we proceed further, we should mention that the 1,540-cm⁻¹ peak falls into the characteristic frequency range not only of the high-frequency ν₃C–O vibration of carbonate complexes, but also the asymmetric C–O stretching vibration, ν_asCO₂⁻, of carboxylates (17, 48, 49, 55), while the 350-cm⁻¹ peak is in the frequency range of the adsorbate–surface vibrations (17, 48).

Our first important finding is that the 1,540-cm⁻¹ peak correlates with the 350-cm⁻¹ peak in all of the spectral sets acquired (Fig. 3D and SI Appendix, Figs. S12D, S14D, S15, S19C, S20, and S21). Moreover, intensities and positions of these two peaks always change with potential in concert. Of particular importance is that breaks in their dependences are always observed at the same potential. For example, in Fig. 3 C and D, intensities of both the peaks increase as potential is scanned from −0.3 to −0.7 V, decreasing at more cathodic potentials. Similarly, the peak positions change marginally at potentials from −0.3 to −0.5 V and red shift at potentials from −0.5 to −0.9 V. A concerted response to a stimulus (in our case, electrode potential) is the main criterion used to discriminate peaks of the same species from a crowded spectrum (56). On this ground, we conclude that the 1,540 and 350-cm⁻¹ peaks characterize the same species.

The second important result is that both these peaks are significantly suppressed in Ar-saturated 0.05 M Na₂CO₃ (pH 11.5) (SI Appendix, Fig. S21). Under these conditions, the theoretical concentration of dissolved CO₂ is as low as 0.02 μM, much lower than 33 mM in CO₂-saturated 0.1 M NaHCO₃. This result suggests that the 1,540 and 350-cm⁻¹ peaks characterize activated CO₂ rather than adsorbed carbonate.

In the next step, we exclude all hypothesized protonated species from the picture (Fig. 1). The rationale is the negligible impact of the D₂O/H₂O isotope exchange on the whole spectra, including the 1,540-cm⁻¹ peak (Fig. 4B and SI Appendix, Fig. S16). Notable is the 1,200- to 1,400-cm⁻¹ region characteristic of the C–OH stretching and HO–C bending peaks of *COOH (37, 57). Hence, the available options narrow down to carboxylates.

Another key result is that both the 1,540- and 350-cm⁻¹ peaks red shift in ¹³CO₂-saturated NaH¹³CO₃ by a factor of 1.026 ± 0.002 (Fig. 4A and SI Appendix, Fig. S17A). Hence, these two peaks are not due to any contamination. The shift of the 1,540-cm⁻¹ peak is more compatible with ν_asCO₂⁻ of carboxylate than ν₃C–O of carbonate. In fact, the ¹³C shift factor of 1.026 ± 0.002 is comparable to 1.029 and 1.027 observed for ν_asCO₂⁻ of sodium acetate (58) and CO₂^•− in a Ne matrix (59), respectively, but lower than 1.033 observed for the high-frequency ν₃CO₃⁻ vibration of HCO₃⁻ in water (60). This relationship is in line with a notion that the C–O stretching vibration that involves a carbon atom connected to two oxygen atoms is typically less affected by the ¹³C/¹²C exchange than that when a carbon atom connected to three oxygen atoms (60). The ¹³C effect on the 350-cm⁻¹ peak indicates that the carboxylate is attached to the surface through the C atom. Hence, the ¹³C/¹²C isotope exchange leaves only the η¹(C)-CO₂⁻ and η²(C,O)-CO₂⁻ carboxylates for further consideration (Fig. 1).

Finally, the ν_asCO₂⁻ character of the 1,540-cm⁻¹ peak rules out the η¹(C)-CO₂⁻ structure (C_2v symmetry). Its ν_asCO₂⁻ vibrational dipole is parallel to the surface. According to the surface selection rules, such a dipole would be strongly suppressed in SERS and silent in SEIRAS, in apparent contradiction with what is observed in Fig. 2 and SI Appendix, Figs. S11, S13, and S15 and refs. 30–32, 35. The same token supports the above rejection of η²(O,O)-CO₂⁻.

Thus, the only species left from the hypothesized set (Fig. 1) is η²(C,O)-CO₂⁻. On this basis, we assign the 1,540- and 350-cm⁻¹ peaks to the ν_asCO₂⁻ and νCu–C vibrations of η²(C,O)-CO₂⁻, respectively. The ν_asCO₂⁻ vibration presents mostly stretching of the C–O bond that points out of the surface, while νCu–C is mostly stretching of the Cu–C bond (SI Appendix, Fig. S22). This peak assignment is corroborated by the dependence of these peaks intensities on potential (see below and Effect of cathodic potential), as well as by the DFT simulations [DFT Model of CO₂ Activation on Cu(111)]. The ν_asCO₂⁻ frequency of 1,515–1,545 cm⁻¹ of η²(C,O)-CO₂⁻ is significantly lower than 1,658 cm⁻¹ of CO₂^•− radical in a Ne matrix (59). Given that the red shift of the ν_asCO₂⁻ vibrational mode correlates with the activation and charge density on a CO₂ molecule (61), this fact suggests that the carboxylate is adsorbed strongly and bears a substantial negative charge, which also agrees with the DFT data.

The Stark tuning rate of the 1,540- and 350-cm⁻¹ peaks is +25–65 and +25 ± 2 cm⁻¹/V, respectively (Fig. 3 C and D and SI Appendix, Figs. S12 C and D, S14 C and D, and S15). Its positive sign indicates that the dipole moments of the corresponding vibrations are parallel to the negative electric field. This alignment is evident for the noncoordinated C–O bond and hence ν_asCO₂⁻ of η²(C,O)-CO₂⁻ (scheme in Fig. 3C). For the Cu–C bond, the parallel alignment is predicted by DFT in the next section (scheme in Fig. 3D). The positive Stark tuning rate also excludes carbon monoxide *CO as a possible origin of the 350-cm⁻¹ peak because the rate of the metal–C vibration of *CO is always negative (from −6 to −20 cm⁻¹/V) (45, 62). Additional evidence of a non-CO origin of the 350-cm⁻¹ peak is provided by the absence of the νCO peak of *CO at potentials from −0.3 to −0.7 V where the 350-cm⁻¹ peak is most pronounced (Fig. 2 and SI Appendix, Figs. S11 and S13).

The symmetric stretching ν_sCO₂⁻ and in-plane bending δCO₂⁻ vibrations of η²(C,O)-CO₂⁻ can be associated with weak peaks at ∼1,330 and 700 cm⁻¹, respectively (Fig. 2). The main argument is that their intensities correlate with intensities of the carboxylate peaks at 1,540 and 350 cm⁻¹ rather than the carbonate peak at 1,065 cm⁻¹ (SI Appendix, Fig. S15). The ν_sCO₂⁻ vibration presents mostly stretching of the C–O bond coordinated to the surface, while δCO₂⁻ presents an umbrella-like movement of the oxygen atoms (SI Appendix, Fig. S22). Hence, dipoles of these two vibrations are almost parallel to the surface, which can explain the low intensity of their SERS peaks. Accordingly, the ¹³C/¹²C isotope exchange red shifts the 1,330-cm⁻¹ peak by a factor of 1.013 ± 0.003, while the 700-cm⁻¹ peak remains intact (Fig. 4A). It should be noted that these two peaks overlap with the ν₃C–O and δCO₂⁻ peaks of coadsorbed η¹-CO₃²⁻, respectively (SI Appendix, Fig. S19C).

The proposed assignment of the carboxylate peaks is in good agreement with literature data for alkali metal-CO₂⁻ and macrocycle-CO₂⁻ adducts, as well as for *CO₂⁻ formed on metals and metal oxides from gas phase (16, 17, 48, 49, 63–66). In particular, the ν_asCO₂⁻ and ν_sCO₂⁻ peaks of *CO₂⁻ on transition metals have been reported at 1,530–1,660 and 1,130–1,350 cm⁻¹, respectively, while its δCO₂⁻ peak, at 650–820 cm⁻¹ (16, 17, 48). The *CO₂⁻ formed from gas phase on CuO nanoparticles and supported Cu has been associated with peaks at ∼1,535 and 1,210 cm⁻¹ (64, 65). Peaks at 1,529–1,520 and 1,347–1,310 cm⁻¹ characterize CO₂ activated on a Co(I) macrocycle in wet acetonitrile (66). In addition, there is a good match between the ν_asCO₂⁻, δCO₂⁻, and νCu–C frequencies of η²(C,O)-CO₂⁻ and their values predicted by DFT [DFT Model of CO₂ Activation on Cu(111) and SI Appendix, section S6].

Finally, we should mention an array of small peaks, which evolves in the electrocatalytic region of potentials (cathodic of −0.8 V) (Fig. 2). These peaks are practically immune to the D₂O/H₂O isotope exchange (Fig. 4B and SI Appendix, Fig. S16). At the same time, some of them (e.g., a peak at 1,600–1,690 cm⁻¹) red shift by the ¹³C/¹²C isotope exchange (Fig. 4A and SI Appendix, Fig. S17), indicating that the corresponding vibrations involve carbon. Accurate assignment of these peaks calls for a separate detailed study.

The proposed assignment of the main peaks is summarized in Table 1.

Table 1.

Assignment of vibrational peaks of adsorbates formed on the Cu electrode in a CO₂-saturated bicarbonate electrolyte

Frequency, cm−1	Assignment
340–350	νCu–C of η2(C,O)-CO2−
610 and 530–535	νCu–O of Cu(I)/Cu(II) oxides
703	In-plane δCO2− of η2(C,O)-CO2−
1,070–1,060	ν1CO32− of η1-CO32−
1,330	νsCO2− of η2(C,O)-CO2−
1,350–1,430	ν3C–O of η1-CO32−
1,540–1,520	νasCO2− of η2(C,O)-CO2−
2,074–2,060	νCO of *CO on atop sites

Effect of cathodic potential.

At −0.1 V, the Cu surface is almost completely passivated by Cu(I)/Cu(II) oxides, which are characterized by a typical doublet at ∼535 and 608 cm⁻¹ (33, 67) (Fig. 2). At −0.3 V, these oxides are partially reduced, as follows from a decrease in the doublet intensity. Concomitantly, distinct peaks of η²(C,O)-CO₂⁻ at 1,540 and 350 cm⁻¹ and of η¹-CO₃²⁻ at 1,065 cm⁻¹ pop up. As potential is biased further, intensity of the η²(C,O)-CO₂⁻ peaks increases and reaches maximum at approximately −0.7 V (Fig. 3 C and D). This point coincides with the potential where intensity of the oxide doublet ceases to decrease further (Fig. 2), which is reproduced in the other sets (SI Appendix, Figs. S11, S13, and S15). Hence, the surface coverage of η²(C,O)-CO₂⁻ increases mostly due to the generation of the adsorption sites by reduction of the blocking oxides.

The found formation of η²(C,O)-CO₂⁻ at potentials much more anodic than the onset of the CO₂ electroreduction current is a very interesting phenomenon. Given that our Cu surface is highly defective and partially oxidized, this finding strongly supports the emerging understanding that the CO₂ activation can be enabled by cooperative effects from the surface defects and/or residual oxygen (68–71), as well as coadsorbed electron donors such as water (69) or hydroxyls. It is also in line with the well-established fact that *CO₂⁻ can be formed on partially reduced metal oxides from gas phase (49, 65) and on heteroaromatic macrocycles in solution (20, 66). This reaction is commonly attributed to the CO₂ activation at a Lewis pair where a partially reduced metal cation serves as an electron donor, while the Lewis acid stabilizes the oxygen atom (72) [DFT Model of CO₂ Activation on Cu(111)]. The η²(C,O)-CO₂⁻ formation at anodic potentials is also consistent with its spontaneous formation predicted by DFT on free and oxide-supported Cu⁰ clusters (72, 73), as well as its large binding energy of −2.16 eV on Cu(111) in the presence of coadsorbed cation [DFT Model of CO₂ Activation on Cu(111)].

As potential is scanned from −0.7 to approximately −1.0 V, intensities of the η²(C,O)-CO₂⁻ peaks go down (Fig. 3 C and D), which can be attributed to a slow takeoff of the η²(C,O)-CO₂⁻ conversion to formate. This interpretation is supported by a small reduction peak observed in this potential range in the first CV (SI Appendix, Fig. S7B) and the traces of formate detected in the electrolyte by NMR at −0.95 V. The decrease in the peak intensities in this potential range is qualitatively consistent with the reaction rate being controlled by the second step as proposed recently on the basis of Tafel slopes for the CO synthesis on Au and Ag (74). However, this conclusion needs to be tested by a rigorous study of the reaction kinetics on the Cu surfaces that are sufficiently catalytically active in this potential range.

In the potential range from −0.3 to −1.0/−0.9 V, the C–O and Cu–C bonds of η²(C,O)-CO₂⁻ are gradually activated (become weaker), as concluded from the red shifts of its ν_asCO₂⁻ and νCu–C peaks (SI Appendix, section S5). The activation can be accounted for by the electrostatic and chemical effects induced by a negative surface charge [DFT Model of CO₂ Activation on Cu(111)].

The pattern changes qualitatively at −1.1 V in Fig. 3 C and D and at −0.9 V in SI Appendix, Fig. S14 C and D. Here, intensities of the carboxylate peaks drop dramatically and afterward slowly decrease (SI Appendix, Fig. S14 C and D). This new pattern perfectly complies with a CO₂ reduction reaction where the RDS is the η²(C,O)-CO₂⁻ formation [1] as suggested by the Tafel analysis (SI Appendix, Fig. S8B). Under these conditions, η²(C,O)-CO₂⁻ is expected to occupy only a minute fraction of surface sites, which explains the low intensities of the carboxylate peaks. Even though a further cathodic bias increases the formation rate of η²(C,O)-CO₂⁻, its surface concentration remains low due to its fast consumption in the second step [4a] and [4b].

Thus, the effect of potential is fully consistent with the status of η²(C,O)-CO₂⁻ as the first intermediate. The lag between the CO₂ activation and the activation followed by the second reaction step provides an exciting opportunity to study spectroscopically factors that promote the first and second step of the CO₂ reduction to formate and CO, as well as to uncover the nature of the second step of these two reactions.

DFT Model of CO₂ Activation on Cu(111).

To verify the SERS results and assist in their interpretation, we model the CO₂ activation on Cu(111) using a plane wave DFT method (see SI Appendix for detail). In agreement with prior theoretical (21, 69, 75) and experimental (46, 76) works, we find that clean Cu(111) does not activate CO₂ (Fig. 5A). This can be explained by the relatively high work function of Cu and the low position of its d-band with respect to the Fermi level.

Fig. 5.

Top and side views of the optimized structures of CO₂ on Cu(111) (A) linear physisorbed CO₂ and (B) η²(C,O)-CO₂⁻ stabilized by Na⁺ hydrated with eight water molecules, (CO₂+Na+8H₂O)/Cu(111). Bond lengths and angles are in angstroms and degrees, respectively. Bader charges of atoms are in |e| units. Atom color codes: dark blue, Cu; red, oxygen; brown, carbon; yellow, sodium. The subsurface Cu atoms are shown in the top views as transparent blue spheres. Hydrated Na⁺ is removed from the top view (B) to show the coordinating Cu atoms.

In the next computational experiment, we place a relaxed Na⁺ cation hydrated by eight water molecules on the top of the linear CO₂ molecule relaxed over Cu(111) (Fig. 5A). This system is dubbed as (CO₂+Na+8H₂O)/Cu(111) hereafter. In this case, CO₂ is activated (Fig. 5B), adopting the η²(C,O)-CO₂⁻ geometry, in unison with the SERS results. The CO₂ activation makes both the C–O bonds longer than in linear CO₂, the effect being much more pronounced for the surface coordinated C–O bond, while the O–C–O angle decreases to 119°. This species has a relatively large binding energy of −2.16 eV, which indicates that in the presence of the cation CO₂ interacts with the Cu surface strongly. The promoting effect of alkali metals on the CO₂ activation on metal surfaces has earlier been attributed to a decrease in the metal work function (46, 75–78) and electric polarization of adsorbed carboxylate by the cation (24).

The predicted vibrational frequencies of η²(C,O)-CO₂⁻ are in good agreement with the experimental values (SI Appendix, Table S2). In particular, the theoretical ν_asCO₂⁻ and νCu–C frequencies of 1,503 and 356 cm⁻¹ are close to the experimental values of 1,515–1,525 and 345 cm⁻¹ in the electrocatalytic region, respectively (Fig. 3 C and D and SI Appendix, Figs. S12 C and D, S14 C and D, and S15). The only exception is the ν_sCO₂⁻ vibration, which is predicted to be at ∼970 cm⁻¹ vs. 1,330 cm⁻¹ observed experimentally at anodic potentials. This discrepancy can tentatively be attributed to limitations of DFT in general and our DFT methodology in particular in modeling stretching vibrations of a molecular bond that is strongly coordinated parallel to the surface (see SI Appendix for more detail).

The η²(C,O)-CO₂⁻ configuration is most popular in molecular complexes of CO₂ (16). It also has earlier been predicted for CO₂ activated on low-index Cu surfaces either by a bias force or alkali metal placed directly on the surface, on-side of CO₂ (24, 75). Our result is that η²(C,O)-CO₂⁻ forms an ionic pair with a hydrated alkali metal cation located above the adsorbed anion. In addition to the electrolyte cations coadsorbed on side, a cation coordinated on top is likely to constitute the local environment of the anion adsorbed at a negatively charged electrode surface. The position of the cation is important for a better understanding of the CO₂ activation process, especially the promoting effects of cations and the role of hydrating water molecules (H-bonding) (79–82).

The Bader charge analysis (Fig. 5B and SI Appendix, Table S1) shows that activated CO₂ gains a net charge of −1.12|e|, that is, close to one electron, while the C atom gains −1.65|e|. The difference is explained by the depletion of the negative charge on the O atoms, especially the O atom coordinated to the surface (it loses −0.38|e|). This effect is caused by donation of π electrons of CO₂ to the surface (see below). In parallel, the Cu atoms in the Cu–C and Cu–O bonds acquire positive charges of +0.11|e| and +0.15|e|, respectively, which additionally stabilize the adsorbed anion electrostatically.

To interpret the Bader charges, we compare η²(C,O)-CO₂⁻ with free CO₂^•− anion radical, which is simulated as a CO₂+Na pair (without surface). In contrast to η²(C,O)-CO₂⁻, all of the atoms in CO₂^•−gain negative charge (SI Appendix, Table S1). It follows that the O atoms in η²(C,O)-CO₂⁻ are less basic than in CO₂^•−, which makes them less susceptible to the electrophilic attack including protonation to *COOH. Hence, reaction [2] is very unlikely on Cu at neutral pH, given that COOH^• (protonated CO₂^•−) is rather acidic [its experimental pK_a values are as low as 1.4–3.4 (83, 84)]. This result can explain the absence of carboxyls *COOH on the Cu surface in a SERS-detectable amount.

The lower basicity of η²(C,O)-CO₂⁻ can also imply that, instead of previously proposed reaction [3], CO is produced on Cu through the direct dissociation of the carboxylate:

$$*{{\mathrm{CO}}_2}^{-}=*\mathrm{CO}+*{\mathrm{O}}^{-},$$ [9]

as observed in gas phase (46, 48) and proposed for CO₂ electroreduction by Ikeda et al. (85). Of particular interest is that pathway [9] dynamically supplies adsorbed oxygen, which is known to promote both the CO₂ activation and reduction (69).

In general, activation of CO₂ on transition metals is described as an acid–base process where the molecule accepts electron density on its lowest unoccupied molecular orbital (LUMO) and simultaneously donates electron from its highest occupied molecular orbital (HOMO) to an empty d orbital of the coordinating metal (16, 17, 48). The particular the d states involved depend on the metal and structure of the coordinating site(s). The frontier orbitals of bent CO₂, which are represented by the corresponding molecular orbitals of CO₂^•−, are shown in Fig. 6A. The LUMO presents the antibonding 6a₁ (in-plane) orbital, which descends from the degenerate anti-bonding 2π_u* LUMO of linear CO₂. It is heavily centered on the carbon atom. According to our density of states calculations, it presents a mix of 33% C 2p and 9% C 2s with 26% O 2p and 2% O 2s from each O atom. The HOMO is a nonbonding in-plane 4b₂ orbital, which descends from the degenerate nonbonding O lone pair 1π_g. This orbital is 100% centered on the O atoms.

Fig. 6.

(A) Isosurfaces of constant electron density of LUMO and HOMO of bent CO₂, which are the 6a₁ and 4b₂ molecular orbitals, respectively. (B and C) The electron density difference for formation of (B) CO₂^•−…Na⁺, Eq. 12, and (C) η²(C,O)-CO₂⁻ on Cu(111) simulated as (CO₂+Na+8H₂O)/Cu(111), Eq. 10. (D) Redistribution of electron density in CO₂/Cu(111) system by Na hydrated by 8H₂O, Eq. 11. Yellow isosurfaces show the electron density gain, while cyan corresponds to the electron density loss. Atom colors: dark blue, Cu; red, oxygen; brown, carbon; yellow, sodium.

Even though η²(C,O)-CO₂⁻ has been predicted for CO₂ activation on several metals (21, 22, 24, 75, 78, 86, 87), its electronic structure has not been studied in detail. Therefore, we calculate spatial redistribution of the electronic charge density, $\mathrm{\Delta }\rho$, using two formulas. In the first one, we subtract the charge density of hydrated Na, ${\rho }_{\mathrm{Na}+8\mathrm{H}2\mathrm{O}}$, CO₂, and Cu(111) from that of the full system, ${\rho }_{\mathrm{Cu}+\mathrm{C}{\mathrm{O}}_2+(\mathrm{Na}+8\mathrm{H}2\mathrm{O})}$:

$$\mathrm{\Delta }\rho ={\rho }_{\mathrm{Cu}+\mathrm{C}{\mathrm{O}}_2+(\mathrm{Na}+8\mathrm{H}2\mathrm{O})}-{\rho }_{\mathrm{C}{\mathrm{O}}_2}-{\rho }_{\mathrm{Cu}}-{\rho }_{\mathrm{Na}+8\mathrm{H}2\mathrm{O}}.$$ [10]

This $\mathrm{\Delta }\rho$ visualizes the charge redistribution upon the formation of (CO₂+Na+8H₂O)/Cu(111). In the second formula, we subtract the charge density of hydrated Na, ${\rho }_{\mathrm{Na}+8\mathrm{H}2\mathrm{O}}$, and the CO₂/Cu(111) system, ${\rho }_{C{O}_2+\mathrm{Cu}}$, from that of the full system, ${\rho }_{\mathrm{Cu}+\mathrm{C}{\mathrm{O}}_2+(\mathrm{Na}+8\mathrm{H}2\mathrm{O})}$:

$$\mathrm{\Delta }\rho ={\rho }_{\mathrm{Cu}+\mathrm{C}{\mathrm{O}}_2+(\mathrm{Na}+8\mathrm{H}2\mathrm{O})}-{\rho }_{\mathrm{C}{\mathrm{O}}_2+\mathrm{Cu}}-{\rho }_{\mathrm{Na}+8\mathrm{H}2\mathrm{O}}.$$ [11]

The resulting plot visualizes the charge rearrangement in the CO₂/Cu(111) system upon its interaction with the hydrated Na (upon accepting its electron). Hence, a difference between $\mathrm{\Delta }\rho$ plots [10] and [11] can shed light onto the effect of negative potential on η²(C,O)-CO₂⁻.

To facilitate interpretation of the $\mathrm{\Delta }\rho$ plots, we also calculate $\mathrm{\Delta }\rho$ induced by Na in the CO₂+Na pair (the CO₂^•− surrogate) as follows:

$$\mathrm{\Delta }\rho ={\rho }_{\mathrm{C}{\mathrm{O}}_2+\mathrm{Na}}-{\rho }_{\mathrm{C}{\mathrm{O}}_2}-{\rho }_{\mathrm{Na}}.$$ [12]

Fig. 6B shows that the charge redistribution upon the CO₂^•− formation corresponds to filling of the LUMO of CO₂. Specifically, $\mathrm{\Delta }\rho$ features depletion of the π electron density between C and O atoms, consistent with the antibonding character of the 6a₁ orbital. The density loss indicates that the C–O bonds become more ionic. Concomitantly, the electron density around the C atom is strongly polarized. The C atom loses electron density from its sides coordinated to the O atoms while accumulating it on the lobe, which creates a local dipole moment on the C atom.

The $\mathrm{\Delta }\rho$ plot of η²(C,O)-CO₂⁻ on Cu(111), which is calculated using Eq. 10 (Fig. 6C), reproduces within the CO₂ moiety the main features associated with the CO₂^•− formation (Fig. 6B). In addition, the plot shows hybridization of the 6a₁ orbital with the Cu d_z2 and d_π (d_xz and d_yz) states, which reflects the formation of the covalent Cu–C and Cu–O bonds. The bond covalency (electron pairing) follows from the accumulation of electron density around the bond centers. The more diffuse volume of the accumulated electron density within the Cu–C bond is consistent with the shorter length of this bond compared to the Cu–O bond (Fig. 5). Because the electron density in the Cu–C bond is closer to the C atom (Fig. 6C), the Cu–C bond dipole is parallel to the negative electric field, which can explain the positive sign of the Stark tuning rate of the νCu–C peak of η²(C,O)-CO₂⁻ (Fig. 3D).

The plot in Fig. 6C also demonstrates that the d electron density on the coordinating Cu atoms is redistributed: Their d_z2 electron density is depleted due to the back-donation to the LUMO, while d_π states acquire electron density from the HOMO. A similar charge redistribution has been obtained for η¹(C)-CO₂⁻ and η²(C,O)-CO₂⁻ on Pt(110) (87). Along with the ionic pairing with alkali cation and the electrostatic stabilization by the positively charged Cu atoms, the covalency of the Cu–C and Cu–O bonds can explain the η²(C,O)-CO₂⁻ formation in the broad range of potentials.

Compared with $\mathrm{\Delta }\rho$ calculated using Eq. 10 (Fig. 6C), $\mathrm{\Delta }\rho$ calculated using Eq. 11 (Fig. 6D) shows a more massive electron depletion from the C–O bonds, suggesting their stronger activation. This result agrees with the red shift of the ν_asCO₂⁻ peak with a cathodic bias of electrode potential (Fig. 3C). Noteworthy is the electron distribution within the carboxylate-surface bonds. Even though electron density is accumulated between the Cu and C atoms of η²(C,O)-CO₂⁻, the negative charge lobe on the C atom is contracted toward the C atom, while the coordinating Cu atom loses d_π electron density. Simultaneously, the d_z2 density on this Cu atom is polarized away from the Cu–C bond. This charge redistribution indicates repulsive interactions, which make the Cu–C bond more ionic. A similar conclusion has been derived for CO₂⁻/Pt(110) based on the direct effect of the electric field (87). The increased ionic character of the Cu–C bond of η²(C,O)-CO₂⁻ at a more negatively charged surface suggests that this bond is activated. This effect can make feasible a nucleophilic attack at the C atom by hydrogen.

In contrast, the Cu atom of the Cu–O bond accumulates d_z2 electron density symmetrically, with the lobe expanding toward the O atom (Fig. 6D). Hence, the Cu–O bond can be stabilized by a negative potential, which can drive the dissociation of η²(C,O)-CO₂⁻ [9]. Given that the Cu–C bond can be concomitantly destabilized, this feature makes the metal–oxygen bond essential for the stabilization and function of η²(C,O)-CO₂⁻ as the first intermediate.

Our SERS and DFT results on CO₂ activation are summarized in Fig. 7.

Fig. 7.

Proposed model of CO₂ activation on a Cu electrode in aqueous electrolyte. η²(C,O)-CO₂⁻ is formed at potentials anodic of the electrocatalytic current. Cathodic polarization gradually activates this species (elongates the C–O bonds and decreases the O–C–O angle), which is accompanied by weakening of its Cu–C bond and stabilization of the Cu–O bond. The electrocatalytic reaction starts when η²(C,O)-CO₂⁻ is able to participate in the second electron transfer, for example, to react with *H to produce formate. Atom colors: red, oxygen; gray, carbon.

Conclusions

By applying a comprehensive suite of experimental and theoretical methods, we establish that the CO₂ conversion to formate on Cu starts with the formation of a carboxylate intermediate with the η²(C,O)-CO₂⁻ structure. This result can be generalized toward at least coinage metals (Cu, Au, and Ag), given their similarity in terms of the reaction kinetics and the electronic properties. At the same time, the formation of *COOH is unlikely on Cu at circumneutral and basic pH, which implies that η²(C,O)-CO₂⁻ is converted to CO through direct dissociation [9] as in gas phase.

η²(C,O)-CO₂⁻ is stabilized on a Cu surface due to the strong covalency of its two bonds with the surface, charge polarization in the system, as well as by the electrostatic interactions with a hydrated Na cation in the on-top position and the positively charged coordinating Cu atoms. The observed formation of η²(C,O)-CO₂⁻ at potentials much more anodic of the onset of the CO₂ electroreduction suggests the critical role of additional cooperative effects such as surface defects, residual surface oxide, and coadsorbed electron donors, which are yet to be understood. A negative potential activates the Cu–C and C–O bonds of η²(C,O)-CO₂⁻ en route to HCOO⁻ (Fig. 7), which is explained by the electrostatic and chemical effects. In contrast, the Cu–O bond can be stabilized, underscoring its important role in the selectivity of η²(C,O)-CO₂⁻ to CO and HCOO⁻.

These results indicate that the descriptors of the CO₂ conversion to HCOO⁻ and CO are the chemical and structural properties of η²(C,O)-CO₂⁻, raising intriguing questions about the exact relationship of these properties to the selectivity, activity, and energy efficiency of the catalyst.