{"id":"a8547b75-f07c-4b11-ab88-5b26360e9109","arxiv_id":"2411.16330","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A constant-potential neural-network molecular dynamics framework simulates Au-water interfaces at nanosecond scale, revealing that K+ cations promote CO2 activation by enhancing surface charge accumulation.","lead":"Researchers built a machine-learning simulation framework that keeps electrodes at a constant voltage while modeling reactions over nanoseconds, something standard quantum simulations cannot do. They used it to show how potassium ions help carbon dioxide bind to gold electrodes and suppress competing hydrogen production.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central cation-promotion mechanism rests on qualitative Bader charge differences that may be within the veNNP's 0.05 e prediction error, leaving the strongest claim quantitatively unsupported.","rationale":"The reader's formal weakest_assumption concerned the absolute potential scale (U_SHE and PZC deviations), but the reader's rationale also noted that the central mechanistic claim is inferred from Bader charge analysis without quantitative error assessment. I agree that the potential-scale uncertainty is a genuine issue, but it is not the most load-bearing for the paper's central claim: a uniform shift of the potential reference would not change the relative comparison between systems with and without cations at the same nominal potential, nor would it alter the observation that CO2 adsorption occurs only at the most negative potentials. The more fragile link is the mechanistic explanation itself, which is the paper's strongest and most novel conclusion. The free-energy data convincingly show that cations promote CO2 activation, but the proposed pathway—enhanced surface electron accumulation and dipole stabilization by the cation field—is supported only by qualitative Bader charge inspection. Bader charges are model-dependent, the veNNP's charge predictions have a documented 0.05 e RMSE, and no statistical comparison is given for the with- versus without-cation charge differences. If that difference is within noise, the mechanism fails even though the promotional effect stands. A quantitative check on the charge difference would settle this directly. Since the concern does not invalidate the framework or the observed promotional effect, the conditional verdict remains appropriate, pending the suggested verification.","tokens_in":39817,"tokens_out":4657,"duration_ms":47982,"concrete_test":"From the 1 ns constant-potential trajectories (or by re-running them if the data are released), compute the mean Bader charge on the Au surface (or on protruding sites) for Au(110) with and without 2 K+ ions at -0.6 V vs. SHE. Apply block averaging over 200 fs blocks to obtain standard errors on the mean difference. Then evaluate the veNNP's Au Bader-charge RMSE by comparing its predictions to DFT Bader charges on a held-out set of ~100 snapshots sampled from these trajectories. If the mean cation-induced charge difference is less than twice the standard error or comparable to the Au Bader-charge RMSE, the claim that cations promote surface charge accumulation is not statistically supported by the presented data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline mechanistic conclusion is that cations promote CO2 activation by 'facilitating surface charge accumulation' on the Au electrode, which stabilizes the bent CO2 dipole through an intensified electric field (Discussion). The free-energy results (Fig. 6a,b) do show that K+ lowers the CO2 adsorption barrier, but the proposed explanation depends on the claim that cations increase electron accumulation on the Au surface. This claim is supported only by qualitative inspection of Bader charge distributions (Fig. 5a, Supplementary Videos 15-18) and a statement that 'the presence of nearby cations enhances this localized accumulation of electrons at the interface' (Results, Atomic charge analysis). No quantitative difference, statistical uncertainty, or comparison to the model's charge-prediction error is provided. The veNNP predicts Bader charges with RMSE 0.05 |e| (Fig. 2c), which is large relative to the subtle, distributed charge changes expected from a K+ ion located >4 Å from the surface. Moreover, Bader charges are not uniquely defined physical observables, and the NNP's dedicated charge network is fitted to Bader targets, so systematic errors in the charge landscape are possible. If the cation-induced Au charge accumulation is comparable to or smaller than the prediction error or the temporal fluctuations, the central mechanism is not established, even though the promotional effect on the barrier remains real. The potential-scale issue raised by the reader is real but secondary: an overall shift in U_SHE would not eliminate the with/without-cation comparison at matched nominal potentials. The charge-analysis gap strikes directly at the paper's most novel and emphasized conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'constant potential reactor' framework that combines a variable-electron-count neural network potential (veNNP) with a modified Nosé-Hoover constant-potential molecular dynamics scheme, an active learning protocol, and slow-growth free energy calculations. The method is applied to Au(111)/Au(110)-water interfaces with CO2 and K+ ions, producing nanosecond-scale trajectories and free-energy barriers for the Volmer step and CO2 adsorption. The central conclusions are that cations promote CO2 activation by enhancing electron accumulation on the Au surface and by stabilizing the bent CO2 dipole through the cation electric field, and that surface reconstruction together with cation-induced disruption of the hydrogen-bond network affects CO2RR/HER selectivity. The paper also reports validation accuracies of the veNNP (energy 0.7 meV/atom, forces 14 meV/Å, Fermi level 4 meV, Bader charges 0.05 e) and describes the methodology in detail.","tokens_in":40064,"tokens_out":10477,"duration_ms":97491,"significance":"If the framework performs as claimed, it represents a useful advance in simulating electrified solid-liquid interfaces at near-DFT accuracy over nanosecond timescales, a regime previously inaccessible to constant-potential AIMD. The inclusion of multi-objective learning with variable electron counts, active learning, and constant-potential dynamics is a valuable methodological combination, and the paper gives quantitative validation of the machine-learned model. The proposed cation-promotion mechanism is plausible and consistent with several experimental observations, but some of the supporting evidence remains qualitative. The paper's primary value is as a methodological demonstration; the mechanistic conclusions require additional quantitative support before they can be considered firmly established.","major_comments":[{"comment":"The absolute electrode potential is obtained from Eq. (1) with a fixed USHE value of 4.44 V, yet the paper reports PZC values for Au(111) and Au(110) that deviate from experiment and notes that a fitted USHE of 4.2 V (Yu et al., ref. 28) brings them closer. Because the central potential-dependent comparison is between -0.4 V and -0.6 V where CO2 adsorption is observed, a plausible ~0.24 V offset in the reference could shift the adsorption threshold and might blur the distinction between the two conditions. I request that the authors demonstrate that their qualitative conclusions (e.g., CO2 adsorption appears only at -0.4 V and -0.6 V; the Volmer barrier decreases with decreasing potential) are robust to the choice of USHE, for example by recalibrating the potential axis with a 4.2 V reference or by reporting the results as a function of the measured Fermi level with both reference values.","section":"Results, 'Nanosecond-scale constant potential MD simulations'; Eq. (1)"},{"comment":"The extended Nosé-Hoover Lagrangian (Eq. 8) requires that the electrode potential Φ entering the equation of motion for the electron number (Eq. 12) equal the derivative of the potential energy with respect to the electron number for the dynamics to conserve the underlying Hamiltonian. The veNNP predicts the energy (Eq. 18) and the corrected Fermi level (Eq. 27) as separate outputs of independent dual-layer networks, with a loss function (Eq. 28) that does not couple these two quantities. Unless ∂E_NN/∂ne is explicitly trained to match the predicted Fermi level, the fictitious potentiostat can exchange energy with the system in a non-Hamiltonian way, potentially biasing the canonical sampling and the reported free-energy differences. I ask the authors to verify on the test set whether the analytical derivative of the NN energy with respect to electron number reproduces the predicted Fermi level, and to discuss the magnitude of any mismatch and its effect on the MD trajectories.","section":"Methods, 'Molecular dynamics at a constant electrode potential' and 'A versatile machine learning framework' (Eqs."},{"comment":"The mechanistic claim that cations promote CO2 activation by enhancing electron accumulation on the Au surface is supported only by qualitative inspection of Bader charge distributions (Fig. 5a, Supplementary Videos 15-18) and the statement that 'the presence of nearby cations enhances this localized accumulation of electrons at the interface.' No quantitative measure of the cation-induced charge difference is reported. Given the veNNP's Bader charge RMSE of 0.05 |e| (Fig. 2c), the subtle, distributed charge changes expected from a K+ ion located more than 4 Å from the surface may be within the model's prediction error. The authors should provide numerical averages with statistical uncertainties (e.g., region-resolved Au charges with and without cations, trajectory-block averages) and demonstrate that the differences exceed both the model RMSE and the temporal fluctuations. Without such quantitative support, the central mechanistic conclusion is not yet established.","section":"Results, 'Atomic charge analysis' and Discussion (Fig. 5a-c)"},{"comment":"The applied potential is reported to fluctuate by up to ±0.3 V around the target, while the target potentials in the series (0.2, 0, -0.2, -0.4, -0.6 V) are separated by only 0.2 V. This implies that the instantaneous potential distributions for adjacent conditions overlap substantially, which could compromise the potential dependence of the free-energy barriers and adsorption onsets. The authors should report the mean and standard deviation of the actual applied potential for each production and slow-growth simulation, and should ideally tighten the potential control (e.g., by adjusting the fictitious masses M_ne and M_s) so that adjacent target potentials are statistically distinguishable in the sampled distributions.","section":"Results, 'Nanosecond-scale constant potential MD simulations' (Fig. 3, Supplementary Figs. 19-20)"}],"minor_comments":[{"comment":"The network name is spelled inconsistently: 'SpookeyNet' appears in the Methods section (e.g., 'we used features of SpookeyNet as inputs') while 'SpookyNet' is used elsewhere; please ensure consistent spelling throughout.","section":"Throughout; Methods, 'A versatile machine learning framework'"},{"comment":"The loss-function weights are denoted a_E, a_F, a_Fermi, a_q in the text but appear as a_S, a_F, a_S_Fermi, a_q in Eq. (28) and the surrounding text; please align the notation.","section":"Methods, 'Hyperparameters and Training'"},{"comment":"The sentence 'CO2 adsorption occurred spontaneously at -0.4 V and -0.6 V2' contains a misplaced citation superscript '2' that appears to be a reference to Monteiro et al.; please correct the citation placement.","section":"Results, 'Volmer Step and CO2 Adsorption'"},{"comment":"The statement that datasets and source code are available 'upon reasonable request' is restrictive for a methods-oriented paper; I recommend depositing the training data, trajectories, and code in a public repository to facilitate reproducibility and adoption by the community.","section":"Data and Code Availability"},{"comment":"The sentence 'It was not possible to study with previous purely explicit models' is vague about which specific finding is meant; please clarify the statement to specify the observation that could not be captured by purely explicit models.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The potential inconsistency between the NN-predicted energy and the NN-predicted Fermi level (major comment 2) is a serious technical point that should be addressed before publication, as it concerns the correctness of the constant-potential dynamics themselves. The Bader-charge evidence (major comment 3) is also important because the headline mechanistic claim depends on it. If the authors can provide the requested quantitative analysis and demonstrate robustness of the potential-dependent conclusions to the USHE uncertainty, the manuscript would be suitable for publication. The paper fits the journal's scope as a computational methods contribution. I would also encourage the authors to consider making the code publicly available, as the source-code availability statement currently limits the broader impact of the framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper actually delivers a working constant-potential reactor that couples a variable-charge neural network (veNNP) with Otani-style constant-potential MD, active learning, and slow-growth free energies. That integration is real and is not present in the cited literature. The veNNP validation is solid on energy, forces, and Fermi level, and the ns-scale trajectories show the potentiostat holding potential within +/-0.3 V. The observation that CO2 adsorbs spontaneously only at more negative potentials, and that K+ lowers the adsorption barrier, is a plausible and potentially important result. The strongest soft spot is the mechanistic claim. The paper argues that K+ promotes CO2 activation by facilitating surface charge accumulation on Au, which stabilizes the bent CO2 dipole via the cation electric field. The barrier data support the promotional effect, but the proposed explanation rests on qualitative inspection of Bader charge distributions from the ML charge model. The veNNP's Bader charge RMSE is 0.05 e, and the cation is >4 Å from CO2. The charge differences said to show cation-enhanced electron accumulation are not quantified against either that error or temporal fluctuations. Bader charges are also not uniquely defined observables. So the mechanism is plausible but not established by the evidence presented. This is not a fatal flaw for the framework, but it is a load-bearing weakness for the paper's most emphasized conclusion. The potential-scale issue is real but secondary. A global U_SHE shift would relabel absolute potentials, but the with/without-cation comparison at matched nominal potentials survives. The +/-0.3 V potential fluctuation is comparable to the 0.2 V intervals, so the potential-dependent adsorption onset should be treated with some caution, but it does not undermine the cation comparison. The paper is also honest about its limitations: it flags the U_SHE uncertainty, discusses anomalous Volmer trajectories, and notes the chosen reaction coordinate may not be the true pathway in some cases. That integrity counts. The lack of public code and data is a real hinderance for reproducibility, though the authors offer materials on request. This is a paper for computational electrochemistry folks who want a working constant-potential ML pipeline. It deserves a serious referee. The framework contribution is strong enough to justify revision even if the mechanistic interpretation needs to be softened or strengthened with quantitative charge analysis.","headline":"A useful constant-potential ML framework, but the headline cation mechanism leans on charge analysis that is not yet quantitatively supported.","tokens_in":635,"tokens_out":878,"would_cite":true,"duration_ms":20479,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A constant-potential reactor framework couples variable-charge neural network potentials to constant-potential molecular dynamics, enabling nanosecond-scale simulations of electrified gold–water interfaces that reveal how K+ cations…","keywords":["constant potential molecular dynamics","variable-charge neural network potential","electrified solid-liquid interface","CO2 electroreduction","alkali metal cation effect","hydrogen evolution reaction","electrical double layer","surface reconstruction"],"falsifier":"Redo the same constant-potential slow-growth simulations using a potential reference fitted to experimental PZCs (for example, U_SHE = 4.2 V) or the cSHE method: if spontaneous CO2 adsorption no longer appears at -0.4 V versus -0.6 V, or if K+ no longer lowers the adsorption barrier, the potential-dependent mechanism is an artifact of the 4.44 V scale. Alternatively, run the K+-containing CO2 adsorption trajectory with a more concentrated electrolyte or a different cation and look for K–O distances below 4 Å; finding direct coordination would overturn the paper's statement that cations act only through the electric field.","tokens_in":107,"feed_emoji":"⚡","tokens_out":7347,"duration_ms":172491,"temperature":0.7,"pith_summary":"This paper claims that a new computational framework, a constant potential reactor, can simulate electrochemical reactions at electrified solid–liquid interfaces with density functional theory accuracy over nanoseconds, long enough to watch interfacial structure evolve in real time. The paper applies it to Au(111) and Au(110) surfaces in water with CO2 and K+ ions, and reports two mechanistic findings: potassium cations promote CO2 activation and suppress the hydrogen evolution reaction, and surface protrusions created by reconstruction act as electron-rich active sites. A sympathetic reader would care because the method addresses a long-standing gap: previous ab initio simulations could only afford picosecond trajectories, usually at fixed charge, while real electrochemical interfaces operate at constant applied potential. If the claims hold, the framework offers a general route to studying electrode–electrolyte phenomena that were previously inaccessible to first-principles simulation.","feed_headline":"K+ ions help CO2 reduction by charging the gold surface","feed_subtitle":"Constant-potential simulations at DFT accuracy trace the cation effect to surface electron buildup, not direct bonding.","key_machinery":"The central object is the variable electronic neural network potential (veNNP): a graph neural network with four dual-layer output heads that predict, alongside local energies and forces, the atomic charges and the corrected Fermi level, with the total system electron count fed in as a feature. Coupled to a modified Nosé–Hoover Lagrangian, of the form $\\mathcal{L} = \\sum_i (\\tfrac{1}{2} m_i s^2 \\dot{\\mathbf{r}}_i^2 - \\mathcal{E}) - \\tfrac{1}{2} M_s \\dot{s}^2 - g k_B T \\ln s + \\tfrac{1}{2} M_{n_e} \\dot{n}_e^2 - \\Phi n_e$, the electron number $n_e$ becomes a dynamical variable that flows in and out of the system to hold the applied potential fixed. The potential itself is assigned from the predicted Fermi level via $U_{\\mathrm{SHE}} = -E_{\\mathrm{Fermi}}/e - 4.44$ V. The veNNP is trained iteratively through an active-learning loop in which a committee of four models flags high-uncertainty configurations for fresh DFT calculations, combining enhanced sampling with targeted first-principles refinement.","core_discovery":"The central claim is that nanosecond-scale, DFT-accurate molecular dynamics of electrified interfaces under constant electrode potential is achievable by training variable-charge neural network potentials (veNNPs) that simultaneously predict energy, forces, atomic charges, and the corrected Fermi level, and by driving them with a Nosé-type Lagrangian that treats the system's electron count as a dynamical variable coupled to a fictitious potentiostat. Using this reactor on gold–water interfaces, the paper finds that CO2 adsorption onto Au(110) occurs spontaneously only at -0.4 V and -0.6 V versus SHE in the presence of K+, and that cations stabilize the activated, bent CO2 by enhancing electron accumulation on the gold surface; the resulting CO2δ– dipole is stabilized by the cation's electric field. K+–O distances remain above 4 Å, ruling out direct coordination. The paper further claims that cations disrupt the interfacial hydrogen-bond network, impeding H+ transfer and thereby suppressing HER, and that reconstructed protruding Au sites accumulate electrons and activate CO2 at potentials where flat sites do not.","pith_inferences":["The framework's charge–structure–potential coupling is not specific to gold or CO2; it should transfer to other electrocatalytic interfaces, and the same active-learning loop could be used to build veNNPs for other metals, alloys, or single-atom catalysts with little modification.","A direct testable corollary of the field-mediated mechanism is that different alkali cations (Li+, Na+, Cs+) should order CO2 activation barriers by their surface-enhancing or field-stabilizing propensity, and that ordering could be measured in the same slow-growth setup.","Because the paper notes the 4.44 V SHE reference may be off by about 0.2–0.3 V, the exact potential window for CO2 activation (-0.4 V versus -0.6 V) is the least secure quantitative output; re-referencing with a fitted U_SHE or the cSHE method could sharpen or shift the reported thresholds.","The slow-growth barriers carry systematic uncertainty from anomalous desorption trajectories and non-minimum pathways; replacing them with a more converged free-energy estimator such as OPES metadynamics would test whether the cation promotion and the protruding-site promotion survive."],"forward_implications":["If the framework delivers DFT accuracy at nanosecond timescales, constant-potential simulations can replace constant-charge AIMD for reactions where interfacial restructuring, ion adsorption, or proton transfer matter, since the reactor freely exchanges electrons with a fictitious potentiostat.","The cation mechanism implies that alkali cations improve CO2 reduction on gold without binding to CO2; the promotion should scale with the cation's ability to stabilize a dipole, not with its coordination strength.","The hydrogen-bond-network disruption by cations provides a concrete selectivity origin: CO2 activation is enhanced while HER is suppressed through hindered H+ transport, matching experiments that show cations raise CO2RR selectivity.","Surface reconstruction becomes a first-order descriptor: reconstructed protruding sites on Au(110) activate CO2 at -0.6 V even without cations, offering a possible explanation for contradictory experimental reports on whether cations are required."],"supporting_citations":[{"why":"Supplies the constant-electrode-potential molecular dynamics formalism, a modified Nosé–Hoover Lagrangian with fluctuating electron number, that the reactor adapts.","marker":"[44]"},{"why":"Supplies the graph-neural-network architecture and four dual-layer output networks that the veNNP extends to predict charges and Fermi level.","marker":"[74]"},{"why":"Implicit solvation model providing the corrected Fermi level from which the applied electrode potential is computed.","marker":"[69]"},{"why":"Experimental report that CO2 electroreduction on gold, copper, and silver is absent without metal cations in solution; the paper's cation mechanism is designed to explain this observation.","marker":"[2]"},{"why":"Introduces the cSHE method for computing potentials of zero charge, used here as a more accurate potential reference than the fixed 4.44 V value.","marker":"[24]"},{"why":"Supplies the fitted 4.2 V SHE reference that brings computed PZCs closer to experiment, exposing the uncertainty in the potential scale.","marker":"[28]"},{"why":"Earlier AIMD study of CO2 activation at Au(110)-water interfaces whose conclusions this work extends to constant potential and nanosecond timescales.","marker":"[51]"}],"fun_headline_variants":["K+ ions charge gold to spur CO2 reduction","Cations boost CO2 reduction by charging gold, not binding","DFT-accurate reactor shows K+ charging gold aids CO2","Nanosecond DFT simulation reveals K+ effect on CO2","Constant-potential MD captures cation charge effect on CO2"],"cache_read_input_tokens":42624,"weakest_assumption_plain":"The computed electrode potentials hinge on converting the implicit-solvent corrected Fermi level to the standard hydrogen electrode scale with the fixed 4.44 V reference, and the paper's own PZC comparison suggests the true reference may be about 0.2–0.3 V lower; if the reference shifts or is system-dependent, the potentials at which CO2 adsorption appears could be mislabeled.","fun_headline_variants_meta":{"raw":{"variants":["K+ ions charge gold to spur CO2 reduction","Cations boost CO2 reduction by charging gold, not binding","DFT-accurate reactor shows K+ charging gold aids CO2","Nanosecond DFT simulation reveals K+ effect on CO2","Constant-potential MD captures cation charge effect on CO2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1367,"prompt_tokens":928,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":544,"tokens_out":439,"duration_ms":5102,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:14:35.640478+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Redo the same constant-potential slow-growth simulations using a potential reference fitted to experimental PZCs (for example, U_SHE = 4.2 V) or the cSHE method: if spontaneous CO2 adsorption no longer appears at -0.4 V versus -0.6 V, or if K+ no longer lowers the adsorption barrier, the potential-dependent mechanism is an artifact of the 4.44 V scale. Alternatively, run the K+-containing CO2 adsorption trajectory with a more concentrated electrolyte or a different cation and look for K–O distances below 4 Å; finding direct coordination would overturn the paper's statement that cations act only through the electric field.","supporting_citations":[{"cited_title":"& Otani M","cited_arxiv_id":null,"evidence_quote":"Supplies the constant-electrode-potential molecular dynamics formalism, a modified Nosé–Hoover Lagrangian with fluctuating electron number, that the reactor adapts."},{"cited_title":"T., Chmiela S., Gastegger M., Schutt K","cited_arxiv_id":null,"evidence_quote":"Supplies the graph-neural-network architecture and four dual-layer output networks that the veNNP extends to predict charges and Fermi level."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Implicit solvation model providing the corrected Fermi level from which the applied electrode potential is computed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Experimental report that CO2 electroreduction on gold, copper, and silver is absent without metal cations in solution; the paper's cation mechanism is designed to explain this observation."},{"cited_title":"& Cheng J","cited_arxiv_id":null,"evidence_quote":"Introduces the cSHE method for computing potentials of zero charge, used here as a more accurate potential reference than the fixed 4.44 V value."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fitted 4.2 V SHE reference that brings computed PZCs closer to experiment, exposing the uncertainty in the potential scale."},{"cited_title":"& Hansen H","cited_arxiv_id":null,"evidence_quote":"Earlier AIMD study of CO2 activation at Au(110)-water interfaces whose conclusions this work extends to constant potential and nanosecond timescales."}],"review_version":1}