In this study, we selectively probe electrocatalytic surface Pd-H* intermediates under operando conditions on single palladium nanocubes, using single-molecule reaction imaging. We choose palladium nanoparticles as electrocatalysts because of the central role of surface Pd-H* intermediates in a variety of electrochemical hydrogenation reactions5,8,29. We visualize the formation and consumption of Pd-H* during electrocatalytic HER under operando conditions on single palladium nanocubes and find that individual nanocubes show interparticle heterogeneity in the stability and reactivity of Pd-H* intermediates. H* atom spillover is clearly observed from the formed Pd-H* on individual palladium nanocubes to the surrounding substrate surface, extending hundreds of nanometres away. Moreover, analysis of single-particle behaviours unveils intraparticle heterogeneity that can be quantitatively modelled by Gaussian broadening among individual surface sites on the same nanocube, which further enables the uncovering of three subpopulations of palladium nanocubes with distinct stability and reactivity, attributable to different compositions in surface sites. We further show correlations between the reactivity, stability and transition-state properties of Pd-H* intermediates across the three subpopulations within the same sample where sample preparation variation can be minimized, which demonstrates the unique advantage of our operando single-molecule reaction imaging approach.
On palladium nanocatalysts in aqueous electrolytes, the electrocatalytic HER typically involves an initial generation of the Pd-H* intermediate from proton reduction (that is, the Volmer step; Fig. 1a, step 1) and a subsequent consumption of one Pd-H* together with another proton reduction to form H (that is, the Heyrovsky step; Fig. 1a, step 2). The Tafel step of combining two Pd-H* intermediates is also possible for the HER, but was shown to be a less probable pathway. To probe the key surface Pd-H* intermediate, our strategy is to use the non-fluorescent molecule resazurin (Rz) as a probe for reducing species to react with Pd-H* on the surface to generate the fluorescent resorufin (Rf) for single-molecule fluorescence imaging (Fig. 1a, step 3), provided that the direct electrochemical reduction of Rz does not interfere with imaging the reaction between Pd-H* and Rz. It is worth noting that by probing the Pd-H* intermediate, potential H bubble formation from the HER, which often prevents proton access to the catalyst surface and interrupts the reaction, would be irrelevant to our experiments because Pd-H* will not be further formed when H bubbles cover the palladium surface.
We chose palladium nanocubes of ∼70 nm in size, synthesized via seed-mediated growth, as model electrocatalysts (Supplementary Fig. 1). They present only one type of facet: (100). Their size is also large enough that edge and corner sites contribute less than ∼1.1% (for an idealized cubic particle; Methods), negligible relative to their total amount of surface sites, and that individual nanocubes can be readily observed under bright-field transmission optical microscopy (see below).
We first performed ensemble-level cyclic voltammetry (CV) to assess these palladium nanocubes deposited on a gold disk electrode for electrocatalytic HER (note that the gold electrode is not active for the HER in the potential range we study here). Survey CV with a wide potential window shows typical electrochemical features of palladium nanocatalysts in aqueous electrolyte (Fig. 1b, (1)), including the overlapping reduction features of H to form surface Pd-H* (that is, H) and bulk Pd-H (that is, H) and the subsequent H evolution in the potential range of ∼0.1 to -0.25 V (all potentials here are versus the reversible hydrogen electrode (RHE) unless otherwise specified), and the corresponding oxidative desorption of H and H in the potential range of ∼0 to 0.56 V. Subsequent CV with a narrower potential window resolves the reversible H and H peaks, and indicates that the redox features for the formation and reduction of surface PdO species are ∼1.0 and ∼0.74 V, respectively (Fig. 1b, (2)); the derived hydrogen adsorption/desorption formal potential E(Pd/Pd-H*, CV) is ∼0.28 V (or -0.14 V versus standard hydrogen electrode (SHE)).
We then evaluated the electrochemical reductions of our probe molecule Rz on palladium surfaces, using a palladium disk electrode. Palladium nanocubes dispersed on a gold disk electrode are unsuitable here because Rz can also be electrochemically reduced by gold electrodes. Ensemble-level CV shows the expected irreversible reduction peak of Rz to form Rf, with a cathodic peak potential E(Rz/Rf) at ∼0.47 V, and the reversible redox peaks between Rf and dihydroresorufin (RfH) centred at ∼0.38 V (Fig. 1b, (3)), as previously reported. The E(Rz/Rf) at ∼0.47 V is much more positive than E(Pd/Pd-H*) at ∼0.28 V, confirming that Pd-H* has sufficient thermodynamic capability to reduce Rz to Rf. Moreover, the peak current density of Rz reduction scales linearly with the square root of the scan rate (Fig. 1b, (3) inset, and Supplementary Fig. 2), indicating that the direct electrochemical reduction of Rz on Pd is diffusion-controlled without specific molecular adsorption of Rz or Rf onto the palladium surface. The peak current-scan rate relationship also gives the diffusion coefficient of Rz (and thus the structurally similar product molecule Rf) in the solution as ∼1.1 × 10 cm s (Supplementary Note 1), consistent with the literature. Because of motion blurring, such diffusion makes it impossible to image the direct electrochemical reduction of Rz to Rf under our typical imaging frame time of 100 ms, thus avoiding interference of this reaction with our targeted reduction of Rz by the surface Pd-H* intermediate.
We therefore proceeded with single-molecule imaging to probe the electrocatalytic Pd-H* species on single palladium nanocubes using our design (Fig. 1a). We dispersed these nanocubes on a transparent indium tin oxide (ITO) working electrode inside a microfluidic electrochemical cell (Supplementary Fig. 3). Under optical transmission imaging at diffraction-limited resolution (∼300 nm), single palladium nanocubes and nanocube-clusters, verified by correlated ex situ scanning electron microscopy (SEM), show distinct darkness (Fig. 1c, (1) and (2)). The single nanocubes had darknesses of <7,000 (a.u.), clearly differentiable from multinanocube clusters (Fig. 1d and Supplementary Note 2).
Upon introducing an aqueous electrolyte containing Rz and at an applied electrochemical potential (E = -0.11 V) at which Pd-H* is produced, we observed stochastic fluorescence bursts of Rf formation on the surface of single palladium nanocubes (Supplementary Fig. 6); each fluorescence burst can be localized to ∼22-nm precision (Supplementary Fig. 4) in correlation with optical transmission and SEM imaging (Fig. 1e-g and Supplementary Note 2). The activity of single palladium nanocubes in this fluorogenic reaction is highly heterogeneous (Fig. 1e). Note that the heterogeneity of activity among different nanocubes possibly contains many contributions, including both the reaction between Pd-H* and Rz (single-molecule fluorescence microscopy) and the formation of Pd-H* (single-particle electrocatalysis). Therefore, to ensure statistical significance, below we only analyse single nanocubes with 5 to ∼950 reaction events during our imaging time, corresponding to the higher activity tail (∼25%) of the total nanocube population (Supplementary Fig. 7). More importantly, these results constitute a first-of-its-kind demonstration of probing surface M-H* intermediates with single-particle and single-molecule resolution.
Interestingly, around each palladium nanocube, the spatial spread (up to hundreds of nanometres away) of the individual reaction events substantially exceeds the ∼70-nm size of the nanocubes (Fig. 1f,g, left). We attribute this spatial spread to H* spillover from the palladium nanocubes onto the ITO support, where it can also react with Rz (Fig. 1h). We used the two-dimensional (2D) standard deviation, s.d., to quantify this spatial spread of reaction events around each nanocube (Fig. 1i, inset, and Supplementary Note 2). We further titrated the electrochemical potential in the cathodic direction from 0.52 to -0.38 V, covering the range of Pd-H* formation and subsequent H evolution (Fig. 1b, (2)), and imaged the fluorogenic probe reaction on many single palladium nanocubes. By averaging the results over many nanocubes (Fig. 1i, black squares), we found that s.d. becomes statistically significant starting at ∼0.22 V, consistent with the E(Pd/Pd-H*, CV) for Pd-H* formation (Fig. 1b, (2)).
Hydrogen spillover has been extensively studied at the ensemble level. Using single-molecule imaging of fluorophore adsorption on H bubbles, Zhang et al. have imaged electrocatalytic H gas formation from single gold nanoplates, where the bubble showed larger spatial spread than the nanoplate size, which they attributed to spillover. Our single-molecule reaction imaging approach here provides the first in situ spatial visualization of migrated hydrogen atoms, utilizing the chemical reactivity of this intermediate species.
A range of additional evidence supports the success of our single-molecule reaction imaging in probing the electrocatalytic surface Pd-H* species and the consequent migrated H*. At a fixed applied electrochemical potential (for example, E = -0.18 V) where Pd-H* forms, the single-particle fluorogenic reaction rate (v) is independent of the solution flow rate in the electrochemical cell (Supplementary Fig. 8), indicating that the reaction kinetics is not limited by reactant supply and corroborating that the reaction is a specific probe of surface H* instead of the diffusion-controlled direct electrochemical reduction of Rz. The single-particle and particle-averaged reaction rates also scale linearly with Rz concentration (Fig. 2a); therefore, in this concentration range (0-100 nM), this fluorogenic reaction is in the linear regime of Langmuir kinetics of surface reactions, consistent with the design in probing surface species. Moreover, the residence time (τ) of the product Rf on single palladium nanocubes is largely independent of E (Fig. 2b and Supplementary Note 2), supporting that the binding of Rf, and thus the structurally analogous reactant Rz, to the palladium nanocube (or ITO) surface are potential independent under our experimental conditions.
Moreover, for our electrochemical potential titration over the range of 0.52 to -0.38 V covering Pd-H* formation and subsequent H evolution, control experiments showed that the activity (Supplementary Fig. 9) and morphology (as synthesized, Supplementary Fig. 1; after imaging experiments, Fig. 1f,g) of palladium nanocubes are stable over the 6- to 12-h period of our imaging experiments. Averaged over tens of nanocubes, the probe reaction rate v toward the cathodic direction shows an onset potential of ∼0.32 V (Fig. 2c). This onset potential correlates closely with E(Pd/Pd-H*, CV) (Fig. 1b, (2)), further supporting that Rz probes the formation of surface Pd-H* (and the migrated H*), whereas Rz's direct electrochemical reduction, which occurs near 0.5 V and generates freely diffusing Rf, is not imaged (Fig. 1b, (3)).
Interestingly, while the particle-averaged v shows a rise and decay behaviour with increasingly negative E, individual particles show diverse v versus E behaviours (Fig. 2c), of approximately three types: rise and decay (type 1, Fig. 2d), rise and plateau (type 2, Fig. 2e), or rise monotonically (type 3, Fig. 2f), with more negative onset potential and higher maximum v from type 1 to type 3 in general, with the type 1 behaviour dominating (∼60% particles). As Rz samples Pd-H* (or the migrated H*), the overall rise and decay behaviour suggests that the original Pd-H* coverage (θ) on a palladium nanocube should in general increase and then decrease with more negative E.
Taking into account the above observations, we formulated a minimal mechanism for the electrocatalytic fluorogenic reaction (Supplementary Note 4), including: an initial Volmer step with a formal reduction potential E that produces more Pd-H* with more negative E (Fig. 1a, step 1); a subsequent Heyrovsky step of Pd-H* consumption to form H with a more negative formal reduction potential E (note E + E = 0 V versus RHE), which should decrease θ at further negative E (Fig. 1a, step 2); and Rz samples Pd-H* to form the fluorescent Rf following a Langmuir-type surface reaction (Fig. 1a, step 3). The migrated H* is approximated as being equivalent to the original Pd-H* as spillover does not change the amount of surface H*. This minimal model gives the (specific) reaction rate v as (equation (1)):
Here k is the (specific) rate constant of the fluorogenic reaction, F is the Faraday constant, R is the gas constant, T is the temperature, Δβ = β - β, where β and β are the symmetry factors of the Volmer and Heyrovsky steps, respectively, and A is a complex term that contains multiple kinetic and thermodynamic parameters of the Volmer and Heyrovsky steps (Supplementary Equations (9) and (10)). At constant pH and [Rz], equation (1) predicts qualitatively the rise and decay behaviour of v with increasingly negative E (Supplementary Fig. 11): the second term in the denominator gives the rise part with only one parameter E; the third term in the dominator gives the decay part, where the steepness of the decay is determined by Δβ and its starting potential is given by the combination of Δβ and A; and k in the numerator determines the absolute magnitude of v.
Fitting the v versus E data yields four parameters (E, Δβ, A, k) for each nanocube and their averages (Fig. 2c-f, red lines). The fitted E, the formal reduction potential to produce Pd-H*, is heterogeneous among individual nanocubes, following an approximate Gaussian dispersion (Fig. 2g, top); large heterogeneity is also observed for the fitted k (Supplementary Fig. 10). The heterogeneity in these parameters is expected given the diverse behaviours of individual nanocubes. More importantly, the fitted curves, although qualitatively producing the v versus E trends, do not reproduce sufficiently the rising part, which is determined by the second term in the denominator of equation (1). This second term contains only one parameter, E, the Pd-H* formal reduction potential, which is not affected by H* spillover.
Considering that each nanocube could present various types of surface sites, we introduced Gaussian broadening of reduction potentials among different surface sites to account for intraparticle heterogeneity in determining v (Supplementary Equation (27) and Supplementary Note 4):
Here e is the formal reduction potential of the Volmer step at individual surface sites, and g is a Gaussian-broadening function centred at E' with a width w (that is, the standard deviation) that quantifies intraparticle site-to-site heterogeneity. This intraparticle Gaussian broadening can effectively change the slope of the rising part in the ν versus E curve (Supplementary Fig. 11). We note that an alternative model incorporating the Tafel reaction can also satisfactorily fit the data (Supplementary Note 4) -- however, previous research showed that the Tafel step does not play a dominant role in the HER on palladium; the alternative model is more complex; and the heterogeneity of the fitted E remains regardless of the model due to the diverse behaviours of individual nanocubes. We therefore settled on the Gaussian-broadening model for its simplicity for further analysis.
The incorporation of Gaussian broadening yielded substantially improved fitting results for both individual nanocubes and their average (Fig. 2c-f, blue lines, and Supplementary Fig. 13). The three types of v versus E behaviours of individual nanocubes are all sufficiently described (Fig. 2d-f, blue lines), supporting the effectiveness of the model and the importance of intraparticle heterogeneity reflected by broadening width w. Interparticle heterogeneity in E' and k is still apparent, as expected (Fig. 2g, right, and Supplementary Fig. 14). Moreover, the extracted E of individual nanocubes show a direct correlation with the original E (Fig. 2g), indicating that this Gaussian-broadening model does not alter the overall trends of the deduced formal reduction potential of Pd-H* formation. More importantly, E' for almost all nanocubes are more negative than E (Fig. 2g), indicating that the conventional model without intraparticle heterogeneity overestimates the stability of the electrocatalytic intermediate Pd-H*. This observation also highlights the substantial influence on the apparent formal reduction potentials (that is, the stability of Pd-H*) of intraparticle heterogeneity, which is now quantified by the Gaussian distribution width w in equation (2). Relatedly, ensemble-averaged measurements (for example, CV) that do not resolve inter- or intraparticle heterogeneity also overestimate the reduction potential, apparent in the notably more positive E(Pd/Pd-H*, CV) (∼0.28 V; Fig. 1b, (2)) compared with the mean E' (∼0.08 V) of the individual nanocubes (Fig. 2g, right). This overestimation stems from the fact that some nanocubes have more positive reduction potentials, leading to an earlier rise of current with increasingly negative E in the ensemble-level CV. Altogether, the analysis here demonstrates that using single-particle measurements, together with the Gaussian-broadening model, we can evaluate both inter- and intraparticle heterogeneity, and obtain more reliable thermodynamic potential of the Volmer reaction. The results here may represent the first application of Gaussian broadening to the microkinetic analysis of electrocatalysis.
Fitting the v versus E data of each nanocube with the Gaussian-broadening model not only provided the five parameters (E', w, Δβ, A, k) but also their uncertainties. Figure 3a shows the 2D correlation between the error of extracted E' and that in k among the individual nanocubes -- note that a larger error here reflects that the respective parameter is less determined by the experimental data, even though the fitting goodness is generally high (Supplementary Fig. 13). This 2D scatter plot and the corresponding 1D projections immediately resolve three subpopulations (S1, S2 and S3) among the individual nanocubes with increasingly larger uncertainties in these two and other fitted parameters (Fig. 3a and Supplementary Fig. 16). Averaging within each subpopulation gives three types of behaviour of v versus increasingly negative E (Fig. 3b): rise and decay (S1), rise and plateau (or gentle decay) (S2), and rise monotonically (S3), all of which were identified qualitatively in individual nanocubes earlier (Fig. 2d-f) but which are now quantitatively resolved into subpopulations.
Fitting the average ν versus E behaviour of each subpopulation also rationalizes their being resolved in the uncertainties of their fitted parameters (Supplementary Table 2). Within the experimental potential window, which was chosen to avoid electrochemical corrosion of the ITO substrate electrode, the S1-type nanocubes, which have the most positive E', show the full rise-and-decay behaviour of v with increasingly negative E (Fig. 3b, red), described by equation (2), according to our electrocatalytic mechanism; overall, this full behaviour leads to these nanocubes having the smallest uncertainties in their fitted parameters. For the S2-type nanocubes, the decay part of their v versus E behaviour is incomplete (Fig. 3b, blue), leading to them having larger parameter uncertainties. The S3-type nanocubes have the most negative E' and only show the rising part (Fig. 3b, green), leading to them having the largest parameter uncertainties.
The different electrocatalytic properties of the three types of nanocubes can perhaps be attributed to the different surface site compositions. The surfaces of these 70-nm palladium nanocubes have predominantly (100) facets and contain three types of hydrogen-adsorption sites, hollow, bridge and atop, with a population ratio of 1:2:1 (ref. ) (Supplementary Fig. 18). The stability of the Pd-H* intermediates at these sites, quantified by the hydrogen-adsorption energy, is known to follow the order hollow > bridge > atop. The more stable sites should have more positive reduction potential in forming Pd-H*. The reduction potentials E' of the three types of nanocubes follow the order S1 > S2 > S3, with S1 having the most positive potential. We therefore propose that these three types of nanocubes possess different compositions of contributing adsorption sites: S1 nanocube surfaces are dominated by hollow sites, whereas S2 and S3 nanocube surfaces contain increasing contributions from bridge and atop sites. Note that such attribution does not mean the three subpopulations have one-to-one correspondence to the three types of surface sites. Several factors might play a role in the variation of surface site composition among individual nanocubes. First, there could be different amounts of ligand residuals on individual palladium nanocubes, even though electrochemical ligand stripping was used to remove most of the surface ligands (Methods). Second, the annealing process of palladium nanocubes on ITO to strengthen their attachment could cause different extents of palladium surface reconstruction due to different amounts of surface ligand being present. Nanocube-ITO contact variations may also contribute here but this is less likely to be a dominant reason because such contact variation would lead to larger interparticle heterogeneity compared with intraparticle heterogeneity, which was not observed in our study (see below). In addition, different catalytic activities of edges, corners, facets and surface defects can also contribute to the differences among the three subpopulations.
The large dataset of the dominant S1 subpopulation (∼60% nanocubes) also enabled us to analyse in detail the intra- and interparticle heterogeneity of the nanocubes. Within S1, the standard deviations w of the Gaussian-broadening function for individual nanocubes are negatively correlated with their E' (Fig. 3c), suggesting that the intraparticle surface site heterogeneity is smaller for nanocubes with more stable Pd-H*, probably due to the increased dominance by hollow sites. More importantly, the E' of individual nanocubes follows a Gaussian distribution (Fig. 3d), whose standard deviation (0.13 V), which quantifies the interparticle heterogeneity, is smaller than the average of intraparticle heterogeneity, 〈w〉 (0.25 ± 0.04 V), where 〈 〉 denotes averaging (Supplementary Table 2). The smaller interparticle heterogeneity probably results from the averaging effect of single-particle data over all surface sites on the same particle. Consistently, w fitted from the averaged data of S1, which also tend to reflect the interparticle heterogeneity, is 0.17 V (Fig. 3b, red curve), very close to the standard deviation of single-particle E'. These results help justify our use of a Gaussian-broadening model and corroborate that the w of a single nanocube provides a good estimate of the dispersion of single-site e values, without needing to resolve individual sites experimentally.
Moreover, even within the single subpopulation S1, we also find that the ensemble-averaged result overestimates the reduction potential: E' (0.2 ± 0.1 V) fitted from the averaged data of S1 is more positive than 〈E'〉 (0.09 ± 0.02 V) of the same population. This overestimate is consistent with that in the comparison between E(Pd/Pd-H*, CV) and 〈E'〉 of the entire population (S1 + S2 + S3) as discussed earlier.
Taken together, these results reveal three subpopulations with distinct stabilities of their Pd-H* intermediates, which are attributable to the different compositions of their surface sites. By analysing the most dominant subpopulation (S1) and thus removing part of the interparticle heterogeneity, we reaffirm the importance of deconvoluting both inter- and intraparticle heterogeneity to obtain a more reliable determination of the reduction potential in forming Pd-H*.
The resolution of three subpopulations among the palladium nanocubes further enabled us to examine the subpopulations' differences and relations in their electrocatalytic properties. Notably, on average, across the three subpopulations there exists a negative correlation between k, the rate constant for the electrocatalytic intermediate Pd-H* (and the migrated H*) to reduce/hydrogenate Rz, and E', the Gaussian-broadening-corrected reduction potential of Pd-H* (Fig. 4a, solid squares), suggesting that more stable Pd-H* intermediates tend to possess lower reactivity, or produce lower-reactivity migrated H*, toward the following electrochemical hydrogenation reaction. However, the k of S2 deviates notably from the fitted line, probably due to contamination by the migrated H* on ITO from H* spillover. The extent of such contamination is reflected by s.d. (Fig. 1i), which quantifies the spatial spread of the reaction events from the nanocube, and which is the largest, and thus suggests the highest contribution of ITO-H*, in S2 (Fig. 4a, inset). To account for this contamination, we further normalized k by the median value of s.d., med(s.d.), for each subpopulation (note that k/med(s.d.) does not mean k per unit density, but rather gives a means to normalize the potential contribution of spilled-over hydrogen to the observed k), which yielded a clear negative correlation between k/med(s.d.) and E'. Using the mean value of s.d. gave similar results (Supplementary Fig. 19). This negative correlation can be rationalized by the potential energy surfaces in the Marcus model of two sequential reactions (Fig. 4c, reaction coordinate ①): an initial electrochemical generation of the Pd-H* intermediate (I) on the palladium surface, followed by the formation of the product Rf (P). A more stable Pd-H* intermediate (dashed blue curve, Fig. 4c, bottom left) would entail a higher activation barrier and thus a smaller rate constant for the second reaction toward the product state. Consistently, the same negative correlation is observed between the k and E' of individual nanocubes (Supplementary Fig. 14).
Moreover, on average, Δβ (≡ β - β), the difference between the symmetry factors of the Volmer and Heyrovsky steps of the HER, and E' are also negatively correlated across the three subpopulations (Fig. 4b). Such behaviour can also be rationalized by the Marcus model of two sequential reactions (Fig. 4c, reaction coordinate ②). β, with values ranging between 0 and 1, is a measure of the symmetry of the energy barrier for an electron-transfer reaction. A larger β corresponds to a transition state closer to the product state. With a more stable Pd-H* intermediate (that is, more positive E'; Fig. 4c, (3), blue dashed line), the transition state (TS1) for forming Pd-H* moves closer to the reactant state, decreasing β, while the transition state (TS2) leading Pd-H* toward HER moves closer to the product state, increasing β, which together should result in smaller Δβ as observed experimentally.
Traditionally, to tune and study the stability and reactivity of Pd-H*, one prepares different palladium nanoparticles with different surface structures, for example, using palladium nanoparticles of varying shapes, facets and dopants, which inevitably faces the materials science challenge of precise sample control. In contrast, through single-molecule reaction imaging, here we leverage the natural heterogeneity among individual particles of a single batch of sample. The results provide important insights into the correlation between the reactivity, stability and transition states of two parallel electrochemical hydrogenation reactions, both involving Pd-H* intermediates, across different surface sites (Fig. 4c).