{"id":"4fe4304a-fb26-4d60-80d9-e4f7b34d38aa","arxiv_id":"2502.08348","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CO2 hydration at the air-water interface proceeds via an 'In and Out' mechanism with bulk-like free energy profiles.","lead":"This paper uses machine-learned simulations to trace how CO2 reacts with water at an air-water interface. It finds that CO2 dives into the first water layer, reacts, and the product floats back out, with reaction energies nearly identical to bulk water.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The strongest evidence for the 'In and Out' mechanism rests on a position-averaged free energy surface and a narrowly conditioned dint average; without position-resolved free energies the interfacial mechanism may be an artifact.","rationale":"The paper has substantial strengths: the revPBE-D3 MACE model is validated against DFT forces, water RDFs, density, interfacial tension, and CCSD(T) cluster energies; the gas-phase and bulk barriers compare well with prior high-level calculations; and the 'In and Out' trend appears in restrained-MD tests for all three MLPs, including the RPA model. These give me confidence that the work is serious and largely reproducible in spirit. However, the central claim that the interfacial reaction is surface-mediated and bulk-like in free energy requires that the data isolate the interfacial region. The current interfacial FEP is a marginal over all depths, and the key Figure 3a average is conditioned on a narrow interfacial window that excludes the equilibrium positions of the reactant and product states. This is directly testable with the existing trajectories, making it more load-bearing than the reader's primary concern about the MACE cutoff, though that cutoff and the RPA model transfer are genuine caveats. The reader's conditional verdict is appropriate; the position-resolved reanalysis should be a condition of acceptance rather than a reason to reject, because the mechanism may well survive the test.","tokens_in":25408,"tokens_out":7556,"duration_ms":88282,"concrete_test":"From the existing metadynamics trajectories, recompute (i) the unrestricted conditional average <dint|sCO> over all configurations, and (ii) the depth-restricted free energy F(sCO, s(OH) | -2 <= dint <= 2 Å) by reweighting with the metadynamics bias. If the unrestricted average no longer dips near sCO ~ 2.9, or if the interfacial-conditioned barrier differs from the bulk barrier by more than about 2 kcal/mol, then the 'In and Out' mechanism and the near-identical bulk/interfacial profile claim are not supported by the current data. A complementary direct check is a new metadynamics run that includes dint as a third collective variable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is not the MACE cutoff but the statistical definition of the 'interfacial' free energy and the depth-reaction correlation. The interfacial free energy profile (Figure 2) is obtained from metadynamics in which only sCO and s(OH) are biased; dint is a passive observable. The slab includes a bulk-like central region, so F(sCO, s(OH)) is a marginal free energy averaged over all depths. Equality of this marginal profile with the bulk profile therefore does not establish that the surface layer has bulk-like reactivity; it only shows that the whole slab is bulk-like on average. Separately, the dashed average dint curve in Figure 3a is computed only over configurations with -1.5 <= dint <= +1.5 Å. Figure 4a shows that CO2 peaks at about +1.6 Å and bicarbonate at about -4.5 Å, so this window clips the dominant reactant population and excludes the dominant product population. The reported dip-and-rise in average dint(sCO) may reflect selection of the interfacial tail rather than a true reaction-coordinate-dependent translation. The representative trajectory (Figure 3b) and restrained-MD tests (SI Section 3, Figure S13) are suggestive, but they are not sufficient to quantify how general or how robust the pathway is. The MACE 5 Å cutoff and the transfer of the RPA model from bulk/gas training data to interfaces are genuine secondary concerns, but they become decisive only after the position-averaging issue is resolved, because the central mechanistic and bulk-vs-interface claims depend on separating surface-region reactivity from bulk-region reactivity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains MACE machine-learned potentials for the CO2 + H2O hydration reaction at revPBE-D3, BLYP-D3, and RPA levels and uses well-tempered metadynamics in gas-phase, bulk, and a 25 Å water-slab systems. It reports near-identical bulk and interfacial free energy surfaces, with a computed pKa of 3.9 and an H3O+ adsorption free energy of -1.3 kcal/mol that match experiment, and it proposes a surface-mediated 'In and Out' mechanism in which CO2 adsorbs at the air-water interface, dives into the first water layer to react, and the carbonic acid product is subsequently expelled. The central claim is that solvation conditions near the interface are bulk-like to within about 1 Å, making the interfacial reaction as feasible as the bulk reaction.","tokens_in":25754,"tokens_out":7436,"duration_ms":82830,"significance":"If the claims hold, the work would be a valuable contribution to interfacial reactivity, with implications for ocean acidification and for general 'on-water' chemistry. Strengths include the use of multiple electronic-structure references, validation against DLPNO-CCSD(T)-F12 for gas-phase and cluster models, free-energy convergence checks in the SI, and reproduction of experimental pKa and H3O+ surface affinity. The central mechanistic conclusion, however, rests on a depth-averaged free energy surface and a conditional dint average; these statistical issues need to be resolved before the paper's strongest claims can be fully accepted.","major_comments":[{"comment":"The interfacial free energy surface in Fig. 2 is a marginal free energy over all depths because the metadynamics biases only sCO and s(OH)aq and the reacting species is free to move anywhere in the 25 Å slab. Since the slab contains a bulk-like central region (and two air-water interfaces), equality of this marginal FES with the bulk FES does not by itself establish that the surface layer has bulk-like reactivity; the profile could be dominated by bulk-like configurations. Please provide a depth-resolved analysis, such as F(sCO, dint) or free energy profiles conditioned on a surface-localized window, to support the statement that the surface layer itself has bulk-like free energies.","section":"Results, 'The interfacial reaction of CO2 is bulk-like'; Methods, 'Metadynamics'"},{"comment":"The dashed average dint(sCO) curve is computed only over configurations with -1.5 ≤ dint ≤ +1.5 Å, as stated in the caption, but the main text describes it without this conditioning. Figure 4a shows that CO2 peaks at dint ≈ +1.6 Å and bicarbonate at dint ≈ -4.5 Å, so this window excludes the dominant reactant and product populations. The observed dip-and-rise in the conditional average may therefore reflect selection of the interfacial tail rather than a true reaction-coordinate-dependent translation of the molecule. Please report the unconditional average dint(sCO) or the full joint distribution P(sCO, dint), and discuss how the conditional window affects the inferred mechanism. The representative trajectory and restrained-MD tests are suggestive but do not replace this statistical quantification.","section":"Figure 3a and associated main-text paragraph"},{"comment":"The RPA model is trained on 500 bulk and 200 gas structures and then used to validate the interfacial 'In-and-Out' mechanism through interfacial density profiles and restrained MD (Figures S12 and S13). No interfacial structures appear to be in the RPA training set, and no interfacial force or energy validation for the RPA model is shown. Because cross-model agreement is invoked as evidence that the mechanism is robust to the electronic-structure treatment, the RPA model's interfacial transferability should be demonstrated, for example by adding interfacial reference structures or reporting interfacial prediction errors. The 5 Å cutoff (10 Å receptive field) is also a concern given the known importance of long-range electrostatics at liquid-vapor interfaces; a quantitative cutoff sensitivity test would strengthen the claim.","section":"SI Section 3, 'BLYP-D3 and RPA Models'; Methods, 'Training'"}],"minor_comments":[{"comment":"The description 'interfacial region within 10 Å of each interface' and the slab length of 25 Å are ambiguous because the two 10 Å regions overlap; please define the effective bulk-like volume and the distance convention used.","section":"Methods, 'System Setup'"},{"comment":"Please specify the normalization and units of the frequency colorbar; 'capped at 50' is not self-explanatory.","section":"Figure 3a caption"},{"comment":"The main text states a cumulative simulation time of 50 ns for each system, while SI Figure S9 reports over 60 ns for the gas-phase reaction; please reconcile these numbers.","section":"Results, 'Metadynamics'"},{"comment":"Reference 29 (Devlin et al.) has the title '1,'; the reference title appears incomplete and should be corrected.","section":"References"},{"comment":"The caption states that the most convincing evidence is in the main text's statistical analysis; given the conditional-window issue in Figure 3a, this sentence should be revised once the statistical analysis is clarified.","section":"SI Figure S13 caption"}],"recommendation":"major_revision","confidential_remarks":"This is a strong computational study, but the current presentation overstates the conclusiveness of the interfacial mechanism because of the depth-averaging and conditional-averaging issues described in the major comments. The required additional analyses are within the scope of the manuscript and should be feasible with the existing simulation data, so I do not see grounds for rejection. The paper is well matched to the journal's scope in physics and physical chemistry."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper gives a new 'In and Out' route for CO2 hydration at the air-water interface and argues the surface layer is bulk-like. The mechanism is probably real, but the main statistical support has a conditioning problem that needs to be fixed before the claim is air-tight.\n\nWhat's good: The model development is thorough. Three MACE potentials trained at revPBE-D3, BLYP-D3, and RPA levels; gas-phase and cluster validation against DLPNO-CCSD(T); pKa of 3.9 for H2CO3/HCO3- and -1.3 kcal/mol hydronium adsorption energy matching experiment. The observation that transition-state structures are found in the first water layer, and that solute-solvent coordination and H-bond counts there are bulk-like, is a solid piece of interfacial science. The representative trajectory and the restrained MD tests across nine runs at three levels of theory all show the same qualitative in-and-out motion. That is real evidence.\n\nSoft spots: The biggest problem is in Figure 3a. The free energy surface from the unconstrained metadynamics is a marginal over all depths. The slab has a large bulk region, so equality of the marginal profile with the bulk profile does not establish that the surface layer alone has bulk-like reactivity. The dashed line in 3a is computed only for configurations with dint between -1.5 and +1.5, while CO2 peaks at +1.6 and bicarbonate at -4.5 (Figure 4a). Conditioning on that window clips the reactant peak and excludes the product well, so the dip-and-rise could be a selection artifact. The representative trajectory and restrained MD rescue the qualitative claim, but they are not statistical proof. I would want to see either a position-resolved free energy surface (FES as a function of sCO and dint) or a reweighted analysis that isolates the surface region. The lack of error bars on the main profiles is a minor issue; the MACE 5 Å cutoff and the RPA model trained only on bulk/gas structures are secondary concerns, since the cross-checks at BLYP-D3 and RPA give the same qualitative picture.\n\nWho it's for: computational chemists interested in aqueous interfaces, CO2 chemistry, or MLP-based enhanced sampling. It deserves a serious referee; the claims are important enough and the work is careful enough. I'd send it to review but ask for a revision that addresses the conditioning issue and, ideally, position-resolved free energies. It would be a strong paper if that is done.\n\nRecommendation: engage with it, but require a revised statistical analysis.","headline":"Solid MLP study with a genuinely interesting mechanism, but the main statistical evidence is weakened by a conditioning window and a marginal free energy surface; the claim is probably right but not yet proven.","tokens_in":26279,"tokens_out":6481,"would_cite":true,"duration_ms":72036,"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":"CO2 hydration at the air-water interface follows an 'In and Out' mechanism — the molecule dives into the surface layer, reacts, and returns — with free energies and barriers nearly identical to bulk.","keywords":["CO2 hydration","air-water interface","machine-learned potentials","well-tempered metadynamics","free energy surface","interfacial reactivity","ocean acidification","carbonic acid"],"falsifier":"Run the same gas-bulk-interface comparison with a potential that includes explicit long-range electrostatics (or with system sizes large enough to converge the interface energetics) and check whether the interfacial barrier still matches the bulk to within ~2 kcal/mol and whether the reacting carbon still dips below the instantaneous interface at the transition state; alternatively, probe the top few angstroms of a CO2-exposed water surface with surface-specific sum-frequency generation and look for a carbonic-acid/bicarbonate signal appearing on surface-reaction timescales, as the authors themselves propose.","tokens_in":25181,"feed_emoji":"🌊","tokens_out":14648,"duration_ms":130544,"temperature":0.7,"pith_summary":"This paper claims that the CO2 hydration reaction — $\\mathrm{CO_2} + \\mathrm{H_2O} \\to \\mathrm{H_2CO_3}$, the process behind ocean acidification — runs at the air-water interface by a surface-mediated 'In and Out' mechanism: CO2 sitting on the surface dives into the top molecular layer of water, reacts there to form carbonic acid (and bicarbonate), and is then pushed back out toward the surface. Using machine-learned interatomic potentials trained at revPBE-D3, BLYP-D3, and random-phase-approximation levels, with 50 nanoseconds of cumulative well-tempered metadynamics per environment, the authors find the interfacial free-energy profile is nearly identical to the bulk profile, with barriers agreeing within 1–2 kcal/mol at a bulk-like $\\Delta F^\\ddagger \\sim 20$ kcal/mol. The reason, they argue, is that solvation near the interface is bulk-like to within about an angstrom: coordination numbers and hydrogen-bond counts in the first water layer match bulk values, so the surface supports the same chemistry as the interior. If this is right, the roughly 350 million square kilometres of ocean surface offers an additional, bulk-equivalent channel for carbonic-acid formation, and surface-adsorbed CO2 deserves a larger role in models of acidification.","feed_headline":"CO2 reacts at the air-water surface by diving in and out","feed_subtitle":"Surface CO2 forms carbonic acid at bulk-like speed, adding an extra channel to ocean acidification.","key_machinery":"The argument is carried by three interacting pieces. First, a MACE potential — a machine-learned interatomic potential built from the atomic cluster expansion and message passing — trained on roughly 8000 structures labelled at the revPBE-D3 level, with companion models at BLYP-D3 and RPA levels, supplies ab initio-quality forces at the multi-nanosecond timescales needed to capture full reactive events. Second, well-tempered metadynamics with two collective variables, the carbon–oxygen coordination number $s_{\\mathrm{CO}}$ (tracking the attack of water on carbon) and a protonation-state coordinate $s(\\mathrm{OH})$ (distinguishing CO2, bicarbonate, and carbonic acid), maps the free-energy surface while leaving the molecule's position relative to the interface completely free. Third, the Willard–Chandler instantaneous-interface construction assigns every configuration a signed depth $d_{\\mathrm{int}}$, and the joint distribution of $s_{\\mathrm{CO}}$ and $d_{\\mathrm{int}}$ is what exposes the 'In and Out' motion: the reaction coordinate and the surface position move together, with the transition state sitting deepest inside the first water layer.","core_discovery":"The central discovery is that the reaction site at the air-water interface is dynamic and coupled to the extent of reaction. In the joint statistics of the carbon–oxygen coordination number $s_{\\mathrm{CO}}$ and the signed depth $d_{\\mathrm{int}}$ relative to the Willard–Chandler instantaneous interface, the reacting carbon starts on the air side ($d_{\\mathrm{int}} > 0$), sinks into the first water layer as water attacks, reaches its deepest point ($d_{\\mathrm{int}} \\sim -2$ Å) at $s_{\\mathrm{CO}} \\approx 2.9$ — roughly the transition state — and then rises again toward the surface as carbonic acid forms. The authors name this the 'In and Out' mechanism and verify that it recurs in potentials trained at three different levels of electronic-structure theory, including beyond-DFT random phase approximation. They also find that the different species self-segregate by charge: neutral CO2 and carbonic acid sit on top of the interface, while charged bicarbonate and the zwitterionic transition state sit inside the water, and that these populations coexist with bulk-like solvation numbers within ~1 Å of the surface. From this they conclude that the interface is not a distorted version of bulk chemistry but essentially bulk chemistry in a thin surface layer, making the reaction equally feasible in both environments and placing a heightened importance on surface-adsorbed CO2 in ocean acidification.","pith_inferences":["A step the authors leave implicit: the depth of the reacting species relative to the interface could itself be promoted from an observable to a collective variable in future enhanced-sampling studies, since this work shows the molecular position and the reaction coordinate move together rather than independently.","The near-identical barriers mean the interfacial contribution to acidification is best described as an added reaction volume at equal rate, not a catalytic speed-up; any 'surface enhancement' language should therefore be read as an extra channel, not an acceleration.","A concrete test of the mechanism's generality would be to apply the same protocol to SO2 hydration at the air-water interface, where earlier work reported genuine surface stabilization of sulfur species; an 'In and Out' path there would support the emerging-charge picture, while a different path would define its limits.","Because ocean water contains salts, an immediate extension is to check whether dissolved NaCl shifts the 'In and Out' trajectory at the interface; the authors flag ionic species and multi-carbon clustering as future work, and ion effects on interfacial solvation could either sharpen or blur the bulk-like picture."],"forward_implications":["Surface-adsorbed CO2 hydrates at essentially the bulk rate, so the ocean surface adds a genuine acidification channel: carbonic acid and bicarbonate form within a couple of angstroms of the surface rather than only after CO2 dissolves into the bulk.","Because the interfacial and bulk free-energy surfaces agree to within 1–2 kcal/mol, the acid–base equilibrium between carbonic acid and bicarbonate (a pKa near 3.9 in the computed profiles) is effectively unchanged at the surface, so no separate surface pKa correction is needed for this step.","For reactions in which charge emerges at the transition state, the reaction is predicted to occur in the first molecular water layer — the balance point between stabilizing developing charge and not disturbing the hydrogen-bond network; the paper suggests phase-transfer catalysis and 'on-water' reactions as places this should be tested.","The same 'In and Out' behavior appears with revPBE-D3, BLYP-D3, and RPA reference data, supporting the authors' claim that the mechanism is a general property of the interface rather than an artifact of one electronic-structure choice."],"supporting_citations":[{"why":"Supplies the gas-phase microsolvation benchmark (barrier ~50 kcal/mol with one water, nearly halved by a second), which the paper reproduces to within 1 kcal/mol.","marker":"[14]"},{"why":"Shows H2CO3 forms via HCO3− in bulk water, giving the stepwise bulk mechanism that the interfacial profiles are compared against.","marker":"[17]"},{"why":"Provides the revPBE-D3 metadynamics study of CO2 reactivity and speciation whose bulk free energies and barrier the present work matches.","marker":"[20]"},{"why":"Documents a sulfur species genuinely stabilized at the air-water interface, the contrasting case that makes the bulk-like interfacial profile for CO2 meaningful.","marker":"[33]"},{"why":"Defines the instantaneous interface construction that yields the depth coordinate dint on which the 'In and Out' mechanism is detected.","marker":"[46]"},{"why":"Earlier work establishing that CO2 prefers to sit on top of the air-water interface, the reactant starting point of the mechanism.","marker":"[47]"},{"why":"Provides the experimental sum-frequency-generation value of −1.3 kcal/mol for hydronium at the interface, which the model matches exactly and uses to validate interfacial energetics.","marker":"[54]"},{"why":"Makes the case that long-range interactions matter at liquid-vapor interfaces, the acknowledged limitation that the 5 Å cutoff potential must overcome.","marker":"[70]"}],"fun_headline_variants":["CO2 dives in, reacts, and carbonic acid exits water","In-and-out surface mechanism drives CO2 hydration","Surface CO2 dips in, reacts, and then pops out as acid","CO2's quick in-out at water surface mirrors bulk kinetics","Air-water CO2 reacts via 'in and out' to make acid"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the machine-learned potential, built from a 5 Å cutoff and without explicit long-range interactions, faithfully reproduces the interfacial free-energy surface, so that if long-range forces materially change the relative stability of CO2, the transition state, or carbonic acid near the surface, both the 'In and Out' mechanism and the near-equality of bulk and interfacial barriers would be called into question.","fun_headline_variants_meta":{"raw":{"variants":["CO2 dives in, reacts, and carbonic acid exits water","In-and-out surface mechanism drives CO2 hydration","Surface CO2 dips in, reacts, and then pops out as acid","CO2's quick in-out at water surface mirrors bulk kinetics","Air-water CO2 reacts via 'in and out' to make acid"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001264,"raw_usage":{"total_tokens":5245,"prompt_tokens":1083,"completion_tokens":4162,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":4073}},"tokens_in":699,"tokens_out":4162,"duration_ms":30843,"temperature":1.0,"reasoning_tokens":4073,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T05:27:22.854462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same gas-bulk-interface comparison with a potential that includes explicit long-range electrostatics (or with system sizes large enough to converge the interface energetics) and check whether the interfacial barrier still matches the bulk to within ~2 kcal/mol and whether the reacting carbon still dips below the instantaneous interface at the transition state; alternatively, probe the top few angstroms of a CO2-exposed water surface with surface-specific sum-frequency generation and look for a carbonic-acid/bicarbonate signal appearing on surface-reaction timescales, as the authors themselves propose.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the gas-phase microsolvation benchmark (barrier ~50 kcal/mol with one water, nearly halved by a second), which the paper reproduces to within 1 kcal/mol."},{"cited_title":"Stirling \\ and\\ author I","cited_arxiv_id":null,"evidence_quote":"Shows H2CO3 forms via HCO3− in bulk water, giving the stepwise bulk mechanism that the interfacial profiles are compared against."},{"cited_title":"Polino , author E","cited_arxiv_id":null,"evidence_quote":"Provides the revPBE-D3 metadynamics study of CO2 reactivity and speciation whose bulk free energies and barrier the present work matches."},{"cited_title":"Buttersack , author I","cited_arxiv_id":null,"evidence_quote":"Documents a sulfur species genuinely stabilized at the air-water interface, the contrasting case that makes the bulk-like interfacial profile for CO2 meaningful."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier work establishing that CO2 prefers to sit on top of the air-water interface, the reactant starting point of the mechanism."}],"review_version":1}