{"id":"43d6325b-79db-4ec6-9692-622f36cd8daa","arxiv_id":"2509.00322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Nanosecond simulations with a neural network potential show that the B-terminated Co3O4(001) surface forms a more ordered, more hydroxylated interfacial water layer than the A-termination.","lead":"A machine-learned atomistic model trained on density functional theory data was used to simulate water on the two faces of a cobalt oxide catalyst for a nanosecond. It finds that the B-terminated surface orders water into a tightly bound quasi-epitaxial layer with more hydroxyl groups, while the A-terminated surface binds water more weakly and less regularly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HDNNP is spin-unaware despite spin-polarized DFT; spin-state averaging could bias the B-vs-A hydroxylation contrast.","rationale":"The reader's weakest assumption identified the DFT+U reference and the MLP trained on it as the critical link, and my concern is a specific, missing sub-requirement within that link: spin consistency. The paper itself states that spin-dependent properties are a known complication for magnetic oxides (Introduction, citing magnetite work), yet the HDNNP architecture has no spin channel and the authors never describe how the reference spin states were stabilized or checked. This is more concrete than the general 'DFT+U might be biased' concern because it pinpoints a specific internal mismatch between the electronic-structure data (spin-polarized) and the ML representation (spin-free). The central claim is a quantitative comparison of hydroxylation and surface–water interactions; if the potential averages over spin states, the relative well depths for proton transfer can be wrong, making the B/A contrast an artifact. I am not claiming this defect is proven—the force RMSE is low and prior AIMD is consistent—but it is the least-secure load-bearing condition and is testable with modest computational effort. The CONDITIONAL verdict already given remains appropriate; my concern reinforces the need for verification but does not change the verdict class. I partially agree with the reader's weakest assumption because it flagged the broad DFT/MLP accuracy but not the specific spin-state averaging mechanism.","tokens_in":23125,"tokens_out":11122,"duration_ms":148978,"concrete_test":"Extract the converged total magnetic moments from all 3,540 DFT reference interface structures (or a sizeable random subset) and group them by spin state. Compute HDNNP energy/force residuals per group. If residuals for minority-spin groups are significantly larger than the overall RMSE, the potential is averaging over spin states. Independently, select 20 hydroxylated configurations from the B-termination and 20 from the A-termination 300 K trajectories, recompute their DFT energies with two different initial magnetic moments (e.g., antiferromagnetic vs. ferromagnetic alignment of Co ions near the surface), and compare to HDNNP predictions. If the spin-state spread exceeds ~0.2 eV or the HDNNP error relative to the lower-energy spin state exceeds ~0.1 eV on these structures, the spin-averaging concern is confirmed and the B-vs-A hydroxylation contrast is not credible as a physical predi","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central B-versus-A contrast in hydroxylation and interfacial ordering is a direct output of the HDNNP. The potential is a standard second-generation HDNNP using only element-based ACSFs (Supplementary Tables I and II), with no spin input, yet the reference DFT is spin-polarized for Co3O4, a magnetic spinel containing Co2+ (high-spin tetrahedral) and Co3+ (low-spin octahedral). For magnetic oxides the PES is not single-valued in atomic coordinates alone: the energy depends on the magnetic state, and spin-polarized AIMD can sample a mixture of local magnetic states, especially when different initial magnetizations relax to different metastable configurations (cf. Refs. 54 and 58). A spin-free HDNNP trained on such a mixture learns an averaged potential that may not correspond to the ground-state Born-Oppenheimer surface governing proton transfer and water dissociation. The reported RMSEs (1.288 meV/atom, 0.1284 eV/Å) only quantify the fit to the average labels; they do not verify fidelity to any spin-resolved surface. Since the paper's headline result is a difference in hydroxylation (Os–H RDF peak at ~1 Å, Fig. 4b) and inner-sphere adsorption (Co*–Ow peak at ~2 vs ~3 Å, Fig. 4a), any systematic spin-averaging error can shift these populations and potentially erase or invert the claimed B/A ordering. The introduction explicitly flags this issue for magnetic oxides, but the Methods and SI never state how spin states were selected, verified, or made consistent in the reference data. This is a missing justification for a key step, making the central claim contingent on an unverified feature of the model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a high-dimensional neural network potential (HDNNP) for the Co3O4(001)/water interface using spin-polarized DFT reference data (optPBE-vdW+U) and an active-learning scheme, then performs ~1 ns molecular dynamics simulations of a ~7,900-atom cell with both A- and B-terminated surfaces in contact with liquid water. The central claim is that the B-termination stabilizes a compact, quasi-epitaxial hydration layer with strong surface templating, enhanced hydroxylation, and a well-organized hydrogen-bond network, whereas the A-termination forms a diffuse, weakly ordered contact layer. The paper also reports rare proton-transfer events, including transient hydronium-like configurations and intermittent protonation of surface hydroxyls, which are illustrated through four selected trajectories.","tokens_in":23463,"tokens_out":4620,"duration_ms":60559,"significance":"If the central contrast is reliable, this is a valuable contribution: it extends prior AIMD studies by orders of magnitude in time and system size, uses a standard active-learning protocol, and provides a concrete termination-dependent picture of interfacial water at a catalytically relevant spinel oxide. The paper is honest about the complexity of magnetic oxides and cites known issues with HDNNPs for magnetite/water (Ref. 54). However, the main physical conclusions are direct outputs of a spin-unaware potential trained on spin-polarized DFT, and the manuscript does not document how spin states were controlled in the reference data. This makes the B-vs-A hydroxylation contrast a correctness-risk point that needs explicit verification. The lack of statistical error bars for RDFs/density profiles and the single-trajectory basis of the proton-transfer analysis also limit the strength of the claims.","major_comments":[{"comment":"The reference DFT is spin-polarized (optPBE-vdW+U), yet the HDNNP uses only element-based ACSFs (SI Tables I and II) with no spin descriptor. The manuscript never states how magnetic moments were initialized or whether spin states were consistent across the bulk, water, and interface reference sets. For a mixed-valence magnetic spinel, DFT+U is known to have multiple metastable magnetic solutions (Ref. 58), and the Introduction itself flags this as a complication for magnetite/water HDNNPs (Ref. 54). If the training labels mix different local magnetic states, the reported RMSEs (1.288 meV/atom, 0.1284 eV/Å) only measure the fit to averaged labels, not fidelity to the ground-state Born-Oppenheimer surface governing water dissociation. Since the B-vs-A hydroxylation contrast is a direct MD output, this is a load-bearing unverified assumption. Please report the spin initialization/control p","section":"Section III A/III C"},{"comment":"The RDFs (Co*–Ow and Os–H) and the density profiles (Fig. 5) are presented without error bars or block averages. The central quantitative claim is that the B-termination shows a sharper ~2 Å peak and a stronger ~1 Å Os–H peak than the A-termination. A single 1 ns trajectory per termination does not by itself establish that these peak-height differences are statistically significant. Please add uncertainty estimates (e.g., standard errors from 5–10 trajectory blocks) or state the statistical precision explicitly. Without this, the 'strong templating' conclusion rests on visual inspection of single-trajectory histograms.","section":"Section IV B, Fig. 4"},{"comment":"The proton-transfer analysis is based on four selected tagged trajectories. Statements such as 'residence time is on the order of ~50 ps' (for one event) and 'rare hydronium-like configurations' are not supported by an event statistics. Please either provide statistical sampling (number of events, lifetimes, per-molecule rates) or present these as purely illustrative case studies and remove quantitative phrases. This does not undermine the structural B/A contrast, but it currently overreaches what can be concluded from selected trajectories.","section":"Section IV A, Fig. 2"}],"minor_comments":[{"comment":"SI Table III lists 'test fraction 0.05', while the main text states that 90% of data were used for training and 10% for testing. Please resolve this inconsistency.","section":"SI Table III vs. main text III C"},{"comment":"The color code states that interfacial oxygen species are blue and lattice oxygens red, but in some panels it is hard to distinguish blue oxygens from red lattice oxygens. Larger symbols or a clearly labeled inset would help.","section":"Fig. 3 caption"},{"comment":"The data availability statement says 'available from the corresponding authors upon request'. For a paper whose main contribution is a trained ML potential, depositing the potential and reference data in a public repository would significantly improve reproducibility.","section":"Data availability"},{"comment":"The notation 'N^α_atoms atoms' is awkward; consider simplifying to 'N_α' or 'N_atoms(α)'.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The spin-state inconsistency is the main technical risk. The authors are experienced in this area and likely have information about spin initialization that is simply missing from the manuscript; however, as written, the central B-vs-A contrast is not fully supported. I would encourage the editor to request the spin-state verification and minor statistical additions described above, rather than rejecting the manuscript, since the methodology is standard and the physical scenario is plausible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-built HDNNP study and the extended sampling is a genuine step beyond the ~20 ps AIMD work. The B-termination result—stronger templating, more hydroxylation, more ordered contact layer—is not wildly new qualitatively, since Kox et al. saw the same direction, but the nanosecond trajectories and the rare proton transfer events are new. The training looks careful: standard active learning, good train/test RMSEs (1.288 meV/atom, 0.128 eV/Å), reasonable ACSF grids, and they openly discuss previous failures on magnetite.\n\nThe weakest point is one the stress-test flags: the HDNNP is spin-unaware. Reference DFT is spin-polarized, and for a magnetic spinel with Co2+/Co3+ that matters. If the AIMD sampling or single-point calculations include a mixture of local magnetic states, a spin-averaged potential may not represent any physical Born-Oppenheimer surface, and the B/A hydroxylation ordering could shift. The paper never states how spin states were initialized, checked for consistency, or verified to remain in the intended magnetic ground state. This is a missing justification, not a demonstrated error. The authors cite the magnetite difficulties, so they know the risk; they just don't tell us how they handled it for Co3O4. A referee should ask for magnetic moment distributions in the reference set and MD snapshots.\n\nOther soft spots are secondary. The RDFs and density profiles come with no error bars; there is presumably only one trajectory per termination (or two, at 300 and 400 K). The proton transfer analysis is four selected tagged molecules, which is fine for illustrating mechanisms but not for statistics. The data and potential are 'available upon request,' which limits reproducibility. All of these are fixable in revision.\n\nMy own verdict: the central claim holds up as far as I can tell. The B/A contrast is large, consistent with prior AIMD, and physically reasonable given the exposed Co3+ sites. The spin concern is real but not confirmed. I'd send it to peer review. The right referee will push on the spin-state consistency and ask for block-averaged errors, but the core methodology is sound and the extended timescale genuinely advances the subfield.","headline":"A solid nanosecond-scale MLP study of the Co3O4(001)-water interface; the B-vs-A contrast is plausible and consistent with prior AIMD, though the spin-unaware potential and missing error bars keep the verdict conditional.","tokens_in":23984,"tokens_out":2081,"would_cite":true,"duration_ms":28533,"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":"The B-termination of Co3O4(001), exposing Co3+ sites, templates water into a compact quasi-epitaxial layer with enhanced hydroxylation, while the A-termination leaves water diffuse and disordered.","keywords":["high-dimensional neural network potential","Co3O4(001)","water interface","surface termination","hydroxylation","proton transfer","molecular dynamics","spinel oxide"],"falsifier":"Replace the DFT+U reference with a hybrid functional (for example HSE06 or PBE0) on small A- and B-terminated Co3O4(001) model systems with explicit water at the same coverage; if the B-termination no longer shows greater hydroxylation and stronger inner-sphere Co–Ow binding than the A-termination, the central claim collapses.","tokens_in":22998,"feed_emoji":"💧","tokens_out":8773,"duration_ms":93081,"temperature":0.7,"pith_summary":"This paper asks what liquid water actually does on the two inequivalent (001) faces of the Co3O4 spinel, a catalyst for oxidations carried out in water. By training a high-dimensional neural network potential on density functional theory and validating it against quantum reference data, the authors run nanosecond molecular dynamics on a roughly 8,000-atom slab with more than 20 monolayers of water, far beyond the reach of ab initio molecular dynamics. They find that the cobalt-rich B-termination forces the first water layer into a compact, quasi-epitaxial, laterally ordered adlayer with clear registry to the lattice, extensive surface hydroxylation, and a well-organized hydrogen-bond network, while the A-termination produces a diffuse, weakly templated, poorly hydroxylated contact layer. The long trajectories also expose rare proton-transfer events, including transient hydronium-like species and intermittent protonation and deprotonation of surface hydroxyls, events that would be missed in few-picosecond AIMD runs. If the central contrast is right, it identifies the B-termination as the water-handling face that matters for aqueous oxidation and helps explain why dopants that occupy Co3+ sites suppress catalytic activity.","feed_headline":"Co3O4's Co3+ face locks water into a crystal-like layer","feed_subtitle":"The Co3+ face hydroxylates and orders water far more than the Co2+ face.","key_machinery":"The load-bearing instrument is a high-dimensional neural network potential (HDNNP): a supervised regression model that maps atom-centered symmetry functions—local structural fingerprints within a 6.35 Å cutoff—through element-specific feedforward neural networks to atomic energies, whose sum gives the total energy and whose derivatives give forces. Trained on 15,118 DFT reference structures computed with optPBE-vdW plus Hubbard U = 2.43 eV, spanning bulk Co3O4, bulk water, ice, and interfaces, with active learning to select uncertain configurations, the potential enables 1 ns simulations of a 7,936-atom four-layer slab in contact with roughly 2,000 water molecules. That scale is what allows","core_discovery":"On its own terms, the central discovery is a termination-controlled dichotomy in interfacial water at Co3O4(001). The B-terminated surface, exposing octahedral Co3+ ions, stabilizes a first hydration layer whose oxygens sit at inner-sphere Co–O distances near 2.0 Å and second-shell positions near 3.3–3.6 Å, giving a sharply peaked Co–Ow radial distribution function and a strong Os–H peak at 1 Å associated with protonated surface oxygens. The A-termination, exposing tetrahedral Co2+, shows only a broad feature near 3 Å, weaker protonation, and a disordered first layer with no registry to the lattice. The authors interpret the B-side ordering as quasi-epitaxial templating of water by the surfa","pith_inferences":["Beyond the paper: the quasi-epitaxial layer implies a lattice match between the water oxygen sublattice and the Co3+ surface sites; a testable prediction is that the B-termination should show a distinct low-frequency vibrational signature in surface-specific sum-frequency generation experiments.","Beyond the paper: if the B > A hydroxylation ordering is generic, pH-dependent activity of Co3O4 should shift with the relative abundance of the two terminations, so particle morphology engineering that exposes more B-type facets could enhance aqueous oxidation rates.","Beyond the paper: the hydronium-like events seen only late in 1 ns runs raise the question of whether even longer trajectories would reveal additional rare chemistry, such as transient cobalt dissolution or oxygen exchange between water and the oxide lattice, which the current potential may or may not capture.","Beyond the paper: because the neural network potential inherits the energetics of its DFT reference, the termination contrast should be re-tested with a hybrid functional before being used for quantitative catalyst design."],"forward_implications":["If the B-termination truly templates water quasi-epitaxially and hydroxylates more strongly, the active surface for aqueous oxidation on Co3O4(001) is likely the Co3+-rich termination, consistent with the experimental observation that substituting Fe3+ into Co3+ sites poisons the catalyst.","Interfacial water on the B face is not passive solvent: it acts as a proton relay through transient OH–, H3O+, and surface hydroxyls that store and shuttle protons during oxidation chemistry.","Nanosecond sampling is necessary to capture rare protonation events; conclusions drawn from ~20 ps AIMD simulations may miss hydronium-like configurations and water exchange with the epitaxial layer.","Raising the temperature to 400 K weakens but does not erase the B > A contrast in interfacial ordering and hydroxylation, so termination effects should persist under mild operating-temperature variation.","The successful transfer of one HDNNP across bulk oxide, bulk water, ice, and the oxide–water interface demonstrates that machine-learned potentials can handle mixed-valence spinels in aqueous environments."],"supporting_citations":[{"why":"Prior AIMD study of Co3O4(001)/water that defines the A and B terminations and the coverage-dependent behavior that this work extends to bulk water and nanosecond scales.","marker":"[38]"},{"why":"Introduces the high-dimensional neural network potential architecture used to build the machine-learned potential.","marker":"[56]"},{"why":"Supplies the bulk Co3O4 reference dataset used as part of the training set.","marker":"[76]"},{"why":"Supplies the curated bulk water reference dataset used for training the potential.","marker":"[75]"},{"why":"Provides the active-learning strategy used to grow the interface reference dataset by selecting high-uncertainty configurations.","marker":"[78]"},{"why":"Experimental study identifying Co3+ sites as the catalytically relevant sites in 2-propanol oxidation, motivating the focus on the B-termination.","marker":"[10]"}],"fun_headline_variants":["Co3+ face orders water into quasi-epitaxial layer at Co3O4(001)","Water templating on Co3O4 depends on surface termination","Co3+ termination creates crystal-like water layer, Co2+ does not","Neural network MD reveals Co3+ face orders water strongly","Hydration layer ordering on Co3O4(001) is termination-controlled"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central B-versus-A contrast rests on one load-bearing premise: the reference electronic structure method (optPBE-vdW with Hubbard U = 2.43 eV) describes water dissociation and proton transfer on cobalt sites accurately enough that the ranking of hydroxylation between the two terminations is correct, and the neural network potential inherits that ranking.","fun_headline_variants_meta":{"raw":{"variants":["Co3+ face orders water into quasi-epitaxial layer at Co3O4(001)","Water templating on Co3O4 depends on surface termination","Co3+ termination creates crystal-like water layer, Co2+ does not","Neural network MD reveals Co3+ face orders water strongly","Hydration layer ordering on Co3O4(001) is termination-controlled"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000755,"raw_usage":{"total_tokens":3227,"prompt_tokens":811,"completion_tokens":2416,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2317}},"tokens_in":555,"tokens_out":2416,"duration_ms":20597,"temperature":1.0,"reasoning_tokens":2317,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:42:08.081623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the DFT+U reference with a hybrid functional (for example HSE06 or PBE0) on small A- and B-terminated Co3O4(001) model systems with explicit water at the same coverage; if the B-termination no longer shows greater hydroxylation and stronger inner-sphere Co–Ow binding than the A-termination, the central claim collapses.","supporting_citations":[{"cited_title":"Impact of solvation on the structure and reactivity of the Co _3 O _4 (001)/H _2 O interface: Insights from molecular dynamics simulations","cited_arxiv_id":null,"evidence_quote":"Prior AIMD study of Co3O4(001)/water that defines the A and B terminations and the coverage-dependent behavior that this work extends to bulk water and nanosecond scales."},{"cited_title":"Generalized neural-network representation of high-dimensional potential-energy surfaces","cited_arxiv_id":null,"evidence_quote":"Introduces the high-dimensional neural network potential architecture used to build the machine-learned potential."},{"cited_title":"A high-dimensional neural network potential for Co _3 O _4","cited_arxiv_id":null,"evidence_quote":"Supplies the bulk Co3O4 reference dataset used as part of the training set."},{"cited_title":"Committee neural network potentials control generalization errors and enable active learning","cited_arxiv_id":null,"evidence_quote":"Supplies the curated bulk water reference dataset used for training the potential."},{"cited_title":"High-dimensional neural network potentials for magnetic systems using spin-dependent atom-centered symmetry functions","cited_arxiv_id":null,"evidence_quote":"Provides the active-learning strategy used to grow the interface reference dataset by selecting high-uncertainty configurations."}],"review_version":1}