{"id":"7bb8a371-bad4-456f-8685-058c5c437312","arxiv_id":"2506.10548","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A machine-learned force field plus a learned charge response accelerates finite-field simulations of the Au(100)/NaCl(aq) interface and predicts a voltage-driven reorientation of interfacial water at the anode.","lead":"The authors built two machine learning models, one for atomic forces and one for electron charge response, and used them to run fast molecular dynamics simulations of a gold/water-salt interface at fixed voltage. The method reproduces ab initio results and reveals how water molecules at the positive electrode flip their orientation at high voltages.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The extrapolation claim is backed by the model's own smooth response rather than by DFT references outside 0-2 V, and the MLEDR charge model used for capacitance has no reported accuracy; the 2.5 V turnover and 20.8 uF/cm2 capacitance rest on this gap.","rationale":"The paper is a well-executed proof of concept: it combines two previously developed models, provides force RMSE, validates structural distributions at training potentials, and releases the PES code and dataset. I find no internal algebraic inconsistency or evidence of a fatal flaw. However, the strongest claims - continuous potential extrapolation to +/-4 V, the 20.8 uF/cm2 capacitance, and the water reorientation turnover at ~2.5 V - depend on extrapolation beyond the {0,1,2} V training set and on the MLEDR charge response, whose accuracy is not reported. The Figure 2 polarization ramp is a necessary but not sufficient check: a smoothly responding ML model could be smoothly wrong, and the AIMD comparison is explicitly limited by short timescales. The 8x8 large-cell results similarly compare MLMD against smaller-cell MLMD, not against DFT. These are not accusations of error; they are missing validation steps. Because the concerns are empirically addressable and the method is promising, the appropriate verdict remains CONDITIONAL, not ACCEPT (pending validation) and not REJECT.","tokens_in":11124,"tokens_out":7666,"duration_ms":95395,"concrete_test":"Perform fresh DFT single-point calculations (same CP2k/PBE-D3/DZVP protocol as Section 2.2) on ~50-100 configurations drawn from the 5 ns MLMD trajectories at U = 3 V, U = 4 V, and U = -3 V for the 4x4 cell, and compute (1) FIREANN force RMSE against DFT forces at these extrapolated potentials; (2) MLEDR predicted integrated surface charge vs. DFT charge-density integration. If the force RMSE stays near the 43 meV/A training value and the surface charge error is below 10% of the reported sigma_m(U) at those potentials, the extrapolation and the ~2.5 V turnover are credible; if errors grow substantially, the headline extrapolation claim must be weakened. A complementary check is to run 50+ ps AIMD at U = 3 V starting from the equilibrated MLMD configuration and compare the water orientation distribution's angular peak shift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The single most load-bearing gap is that the paper's 'successful extrapolation' to +/-4 V is validated only by internal consistency, not by comparison with DFT at extrapolated potentials. Section 3.1 demonstrates that when the applied potential is ramped over [-4,4] V, the FIREANN-driven electrolyte polarization follows the ramp and returns on reversal. Because the trajectory is generated by the same FIREANN potential being tested, this shows self-consistency but not accuracy. The AIMD comparison in Fig. 2c is weakened by the paper's own statement that 10-100 ps AIMD is insufficient for electrolyte relaxation, so it cannot serve as a strong reference for the slow ionic polarization that dominates the response at these concentrations. The high-potential structural results - the anodic water reorientation turnover near 2.5 V (Section 3.3, Fig. S3) and the interpretation that the anode outcompetes adsorbed Cl- - lie entirely outside the 0-2 V training window. Second, the MLEDR model that supplies the charge response used to compute the Helmholtz capacitance (Section 3.2, Eq. 7) has no error metric in the main text; the paper reports only the FIREANN force RMSE (Table S1, ~43 meV/A), and the electron response dataset is 'available upon request,' not released. The capacitance value 20.8 uF/cm2 and the claimed charge-inversion mechanism therefore cannot currently be independently audited. Because both headline numbers depend on unvalidated extrapolation of one model and unquantified accuracy of another, the central claim is conditional on tests that the paper does not report.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a machine-learning-accelerated finite-field method for electrochemical interfaces. It combines two neural-network models: FIREANN, a field-dependent machine-learned potential trained on DFT forces, and MLEDR, a machine-learned electron-density-response model trained on DFT charge densities. The method is demonstrated on the Au(100)/NaCl(aq) interface, with training data from active-learning AIMD at 0, 1, and 2 V cell potentials. The authors report nanosecond-scale MLMD trajectories, a Helmholtz capacitance of about 20.8 uF/cm2 at 0 V, and a potential-dependent reorientation of interfacial water at the anode starting near 2.5 V, which they attribute to competition between the charged anode and adsorbed chloride ions. They further claim successful extrapolation of the ML models to cell potentials in the [-4, 4] V range and scalability to an 8x8 supercell with 2880 atoms.","tokens_in":11443,"tokens_out":3487,"duration_ms":43791,"significance":"If the central claims hold, this is a valuable methodological contribution: it replaces classical electrode/electrolyte descriptions in finite-field simulations with fully ML-learned DFT-level components, enables nanosecond sampling at first-principles accuracy, and provides a route to potential-dependent interfacial structure and capacitance. The paper's strengths include the use of active learning for dataset construction, the release of the FIREANN code and the PES dataset, and the explicit recognition that AIMD timescales are insufficient for electrolyte relaxation. The method's headline quantitative outputs (the capacitance and the high-potential water-orientation turnover), however, rest on two models whose accuracy outside the training window is not directly established, so the current evidence is suggestive rather than conclusive.","major_comments":[{"comment":"The claimed extrapolation to the full [-4, 4] V range is supported only by the model's own smooth response, not by comparison with DFT at potentials outside the 0-2 V training window. The ramp test in Fig. 2 shows that the FIREANN-generated electrolyte polarization follows the applied potential and returns on reversal, but because the trajectory is generated by the very potential being tested, this demonstrates internal consistency rather than accuracy. Furthermore, the AIMD comparison in Fig. 2c, limited to 30 ps, cannot serve as a stringent reference for the slow ionic polarization that dominates the response at this concentration, as the paper itself notes that 10-100 ps AIMD is insufficient for electrolyte relaxation. I recommend adding direct DFT checks, e.g., forces, energies, or charge densities computed for configurations sampled at 3 V and 4 V, or alternatively reframing the high-potential results as unvalidated predictions rather than validated extrapolations.","section":"Sec. 3.1, Fig. 2"},{"comment":"The MLEDR model that provides the surface charge density sigma_m used in the Helmholtz capacitance calculation has no reported accuracy metric in the main text or the SI. The only error reported is the FIREANN force RMSE of about 43 meV/A (Table S1); there is no RMSE or other validation for the predicted electron density response. Since the capacitance value of about 20.8 uF/cm2 and the charge-transfer/charge-inversion analysis in Fig. 3 depend entirely on MLEDR, the absence of any error metric prevents independent audit of these headline results. Please report MLEDR training and validation errors, include a direct comparison of predicted versus DFT charge-density differences for representative configurations, and make the electron-response dataset publicly available rather than 'available upon request.'","section":"Sec. 3.2, Eq. (7)"},{"comment":"The anodic water-reorientation turnover starting at about 2.5 V lies entirely outside the 0-2 V training range, so its reliability depends on the unvalidated extrapolation discussed above. In addition, the large-cell (8x8, 2880-atom) simulations in Fig. S5 are claimed to 'compare well' with the smaller training cell, but the comparison appears to be only visual and no quantitative error metric is provided for the large-cell results. To support the structural conclusions, please provide a quantitative comparison of ion concentration and water orientation profiles between the 4x4 and 8x8 cells, and include a DFT spot-check of forces and/or charge response at high cell potentials and in the larger cell.","section":"Sec. 3.3, Fig. S3 and Fig. S5"}],"minor_comments":[{"comment":"The text says 'the charge response defined in Eq. 5' but Eq. (5) is the message-passing iteration; the MLEDR target is defined in Eq. (6). Please correct the cross-reference.","section":"Sec. 2.1"},{"comment":"The sentence 'which can be used to generate new FI-EAD features via Eq. (3)' appears to refer to Eq. (4), which is the squared linear combination that forms the FI-EAD feature. Please check the equation numbering and the intended reference.","section":"Sec. 2.1"},{"comment":"The caption states 'The grey area in all panel indicates the Au electrode'; 'all panel' should be 'all panels'.","section":"Fig. 3 caption"},{"comment":"The Table of Contents entry contains the typo 'finite-filed simulations'; it should be 'finite-field simulations'.","section":"Table of Contents entry"},{"comment":"Reference [49] is listed as '2025' with no journal, volume, or DOI; please provide complete citation information.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I support considering this manuscript for publication after a major revision. The central methodology is promising and the code release is a positive step, but the two headline numbers (capacitance and high-potential water turnover) currently rest on unvalidated extrapolation and an unquantified MLEDR model. I would encourage the editor to require the MLEDR dataset to be released and to ask for direct DFT validation at potentials outside the training window as a condition of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth taking seriously. It demonstrates that two of the group's own ML models, FIREANN and MLEDR, can be combined to run constant-potential finite-field MD on an Au(100)/NaCl(aq) interface at nanosecond timescales and at system sizes far beyond AIMD, with no classical model of electrode or electrolyte. That integration is new, and the validation on electrolyte polarization and interfacial structure against 30 ps AIMD is honest and reasonably convincing. The force RMSE (~43 meV/A) is in the normal range, and the authors are upfront that AIMD is too short to relax the electrolyte fully, which argues in their favor.\n\nThe soft spots are the ones you'd expect. The extrapolation claim to ±4 V is supported only by internal consistency of the model's own response, not by DFT at potentials outside the 0-2 V training window. The high-voltage water reorientation turnover near 2.5 V sits entirely in that extrapolated regime. And the MLEDR charge response model, which feeds the headline capacitance of ~20.8 uF/cm2, has no error metric in the main text; the electron response dataset is only 'available upon request.' Those are real gaps, but they are addressable, not structural. The method does not fall apart on internal contradiction; it simply needs stronger reference data at extrapolated potentials and a reported accuracy for the charge model.\n\nThe paper is for people who work on ML for electrochemical interfaces. It is a legitimate proof-of-concept, and it deserves a serious referee. I would send it to review, with the expectation that the authors be asked to add DFT spot checks at 3-4 V and report MLEDR errors and dataset release.","headline":"A credible ML finite-field workflow with real validation, but its headline capacitance and high-voltage turnover need stronger extrapolation checks.","tokens_in":12028,"tokens_out":1529,"would_cite":true,"duration_ms":17028,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two neural networks trained on DFT data replace ab initio molecular dynamics for the Au(100)/NaCl(aq) interface, enabling nanosecond trajectories and extrapolation to cell potentials beyond the training range.","keywords":["electrochemistry","electric double layer","finite-field simulation","machine learning potential","neural network force field","electron density response","Helmholtz capacitance","Au(100)/NaCl(aq) interface"],"falsifier":"Run fresh DFT reference calculations at cell potentials inside the extrapolated region, for example 3.5 V and -3.5 V, and in the 2880-atom cell, then compare atomic forces and field-induced electron densities with FIREANN and MLEDR predictions; if the force error grows far beyond the reported 43 meV/Å or the charge response shifts the Helmholtz capacitance by more than a few microfarads per square centimeter, the extrapolation claim would be overturned.","tokens_in":10872,"feed_emoji":"⚡","tokens_out":9046,"duration_ms":92482,"temperature":0.7,"pith_summary":"Electrochemical interfaces are hard to simulate because the electrolyte needs nanoseconds to equilibrate, while finite-field ab initio molecular dynamics is limited to tens of picoseconds. This paper argues that two neural networks trained on first-principles data remove that bottleneck: a field-dependent machine learning potential predicts atomic forces under an applied electric field, and a machine-learned electron density response model predicts how charge rearranges at the electrode. Tested on a Au(100)/NaCl(aq) full cell, the models produce multi-nanosecond trajectories, a Helmholtz capacitance of about $20.8\\ \\mu\\mathrm{F}/\\mathrm{cm}^2$ at 0 V, and electrolyte polarization that tracks cell potentials from -4 V to 4 V even though training used only 0, 1, and 2 V. The longer trajectories also reveal a structural turnover: near the anode, interfacial water reorients above about 2.5 V as the positively charged electrode begins to outcompete adsorbed chloride ions for water molecules.","feed_headline":"Machine-learned force fields give nanosecond charged-interface runs","feed_subtitle":"Two neural networks trained on DFT replace ab initio MD and predict capacitance at a gold/NaCl interface.","key_machinery":"The argument rides on two paired neural network models plus a specific cell geometry. The FIREANN potential uses the field-induced embedded atom density (FI-EAD) descriptor, whose field-dependent orbital makes the network explicitly aware of the applied field's direction and magnitude, with message-passing iterations adding nonlocal electrostatic information; this model supplies forces at every configuration. The MLEDR model places ghost atoms at electron-density grid points and learns the difference between the electron density with and without the applied field, giving the charge response used for capacitance. The finite-field cell setup (reference [39]) imposes a voltage across a complete cell without a vacuum gap, so electroneutrality and constant-potential conditions emerge naturally from the metal's free-electron response.","core_discovery":"On the authors' terms, the central discovery is that fully first-principles finite-field molecular dynamics can be replaced by a two-model machine learning pipeline with no classical approximation for either electrode or electrolyte. The field-dependent potential learns the potential energy surface from DFT forces, while the electron density response model learns the field-induced density difference, so the surface charge response and differential Helmholtz capacitance follow from machine learning predictions rather than additional DFT post-processing. The specific physical finding is a potential-driven turnover in anodic water orientation: at low potentials, adsorbed chloride ions orient water with an O-H bond toward the electrode, but above about 2.5 V the concentrated positive charge on the anode attracts the oxygen end of water more strongly, shifting the dominant angular population from roughly 135 degrees to smaller angles. The authors attribute this discovery to nanosecond-scale sampling, which removes relaxation artifacts that make short AIMD concentration profiles unreliable.","pith_inferences":["If this sparse-training recipe generalizes, active learning at a few cell potentials could become a standard protocol for mapping the full potential window of other metal/electrolyte interfaces, making finite-field machine learning molecular dynamics a screening tool for electrode materials.","The predicted anodic water reorientation near 2.5 V is a testable signature: in situ vibrational sum-frequency or surface-enhanced infrared spectroscopy of interfacial water should show the dominant water orientation shifting as the applied potential crosses that threshold.","Because the potential energy surface is trained on forces only, total energies are not directly available, so computing free energies and reaction barriers would require an additional energy model or thermodynamic integration.","The electron density response model predicts a density difference, not the total density; properties such as absolute work functions or site-resolved charge transfer would need a supplementary model for the field-free electron density."],"forward_implications":["Constant-potential machine learning molecular dynamics can run for nanoseconds on metal/electrolyte cells, so ion concentration profiles and potentials of mean force can be statistically converged instead of being read from 10-100 ps ab initio runs.","Differential Helmholtz capacitance can be obtained directly from the learned charge response, without extra DFT single-point calculations for every sampled configuration.","Cell potentials can be varied continuously beyond the training window, making it possible to map potential-driven structural transitions such as the water-orientation turnover near 2.5 V without retraining.","The same trained models transfer to cells larger than the training cell: the 2880-atom Au(100)/NaCl(aq) system reproduces the small-cell ion and water distributions, supporting size scalability.","AIMD concentration profiles contain artificial oscillations from incomplete relaxation, so the longer MLMD trajectories give more reliable interfacial structure."],"supporting_citations":[{"why":"Supplies the FIREANN field-dependent neural network potential and FI-EAD descriptor used to predict forces under applied fields.","marker":"[36]"},{"why":"Supplies the MLEDR electron density response model that learns field-induced density differences for charge response.","marker":"[37]"},{"why":"Provides the finite-field cell setup that imposes a voltage across a full cell without vacuum and underlies the simulations.","marker":"[39]"},{"why":"The finite-field AIMD baseline whose 10-100 ps timescale and capacitance results this work compares against.","marker":"[16]"},{"why":"Active learning workflow used to progressively sample the 12,139 training configurations.","marker":"[41]"},{"why":"Comparable machine-learned counter-ions study of Au(100) reporting 17-19 μF/cm² capacitance used as consistency reference.","marker":"[49]"}],"fun_headline_variants":["ML-accelerated finite-field MD flips water orientation at anode","Two neural nets replace DFT for charged interface simulations","Nanosecond ML runs reveal water turnover at Au electrode","Machine-learned fields predict capacitance and water flip at electrodes","DFT-free finite-field MD captures water flip at gold anode"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a force field and a charge-response model trained on 12,139 configurations at just 0, 1, and 2 V stay accurate from -4 V to 4 V and in a 2880-atom cell that was never part of the training set.","fun_headline_variants_meta":{"raw":{"variants":["ML-accelerated finite-field MD flips water orientation at anode","Two neural nets replace DFT for charged interface simulations","Nanosecond ML runs reveal water turnover at Au electrode","Machine-learned fields predict capacitance and water flip at electrodes","DFT-free finite-field MD captures water flip at gold anode"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1599,"prompt_tokens":933,"completion_tokens":666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":585}},"tokens_in":549,"tokens_out":666,"duration_ms":7413,"temperature":1.0,"reasoning_tokens":585,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:23:20.438102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run fresh DFT reference calculations at cell potentials inside the extrapolated region, for example 3.5 V and -3.5 V, and in the 2880-atom cell, then compare atomic forces and field-induced electron densities with FIREANN and MLEDR predictions; if the force error grows far beyond the reported 43 meV/Å or the charge response shifts the Helmholtz capacitance by more than a few microfarads per square centimeter, the extrapolation claim would be overturned.","supporting_citations":[{"cited_title":"Zhang, B","cited_arxiv_id":null,"evidence_quote":"Supplies the FIREANN field-dependent neural network potential and FI-EAD descriptor used to predict forces under applied fields."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MLEDR electron density response model that learns field-induced density differences for charge response."},{"cited_title":"Dufils, G","cited_arxiv_id":null,"evidence_quote":"Provides the finite-field cell setup that imposes a voltage across a full cell without vacuum and underlies the simulations."},{"cited_title":"Andersson, M","cited_arxiv_id":null,"evidence_quote":"The finite-field AIMD baseline whose 10-100 ps timescale and capacitance results this work compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Active learning workflow used to progressively sample the 12,139 training configurations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Comparable machine-learned counter-ions study of Au(100) reporting 17-19 μF/cm² capacitance used as consistency reference."}],"review_version":1}