{"id":"bbe15f20-4fb6-41ce-a7e2-2ee4cfdf4583","arxiv_id":"2505.19745","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"RAZOR machine-learns the work function and Born charges of electrified interfaces, enabling bias-dependent molecular dynamics that predicts a pH-driven OH adsorption site switch on Cu(100).","lead":"A new machine-learning framework (RAZOR) extends interatomic potentials to predict how charged metal surfaces and adsorbates respond to electrode bias, then uses it to simulate hydroxide on copper. The model reproduces the experimentally seen pH-dependent switch of OH adsorption sites at Cu(100).","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Second-order truncation of E(q) is not independently validated: the AITD comparison relies on the same quadratic form, so the finite-bias site switch could be an artifact of the assumed parabola.","rationale":"The paper presents a clever and potentially useful extension of MLIPs to electrified interfaces by learning the work function and Born effective charges, and the MD application to OH on Cu(100) is physically plausible. The main weakness is real and load-bearing: the second-order truncation of E(q) is built into the RAZOR architecture, and the AITD comparison cited for validation may share the same quadratic assumption, making it a consistency check rather than an independent test. This matters because the headline site switch and its pH dependence occur over a narrow charge window, where even modest cubic terms or a site-dependent capacitance could move the crossing. The proposed direct finite-q DFT calculation for the two reference sites would settle the issue concretely without requiring new methodology. The reader's CONDITIONAL verdict is appropriate: the method could be correct, but the key application-level claim currently rests on an assumption whose magnitude has not been measured. No adjustment to the verdict is needed, though the paper should either provide the finite-q validation or clearly label the truncation as an approximation with estimated error bounds.","tokens_in":10023,"tokens_out":5242,"duration_ms":59519,"concrete_test":"Compute constant-charge DFT energies with the same ENVIRON implicit-solvent settings for the static c(2x2)-OH_bridge and c(2x2)-OH_hollow cells at q = 0, +/-0.02, +/-0.05, and +/-0.10 e per Cu surface atom. Fit these energies to E(q)=E0+phi0 q+q^2/(2C_el) for each site separately and compare to the RAZOR-MLIP parabola. If the DFT points deviate from the quadratic by more than ~20 meV per OH in the switch range, or if the fitted C_el differs appreciably between bridge and hollow, then the neglected site-dependent quadratic term shifts the crossing charge and the site-switch conclusion is not secure. If the data fall on the parabola with site-independent C_el, the truncation is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that RAZOR-MLIP MD reveals a charge-induced bridge-to-hollow site switch, and thereby a pH-dependent adsorption site, rests on the energy model E(q)=E0+phi0 q+q^2/(2C_el,0), with C_el,0 treated as configuration-independent. Because of this construction, the relative stability of bridge and hollow at finite q is determined wholly by E0 and phi0; any site dependence of the second-order response or higher-order terms is excluded by fiat. The main-text validation against AITD is not an independent test if AITD invokes the same second-order expansion and the same implicit-solvent capacitance: agreement then only demonstrates internal consistency, not that the truncation is accurate over the applied q range. The paper reports no direct finite-q DFT energies for the relevant structures, so the magnitude of neglected cubic or site-dependent quadratic terms is unknown over the q range where the switch occurs (0.02 to 0.05 e per Cu atom). Given that the experimental pH-dependence is the paper's headline application, this unvalidated assumption is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces RAZOR, a framework that augments standard machine-learning interatomic potentials with the response of an electrode's total energy to excess charge q. It learns the work function phi0 as the first-order derivative of energy with respect to q, stabilizes this via Born effective charges, and treats the second-order response through a fixed, configuration-independent capacitance derived from an implicit solvent model. The resulting surrogate is used for constant-charge MD of 0.5 ML OH on Cu(100), yielding a charge-induced site switch from bridge at negative q to hollow at positive q. After a Legendre transform and CHE referencing, the authors obtain a pH-dependent preferred adsorption site that they argue matches experimental reports.","tokens_in":10171,"tokens_out":5928,"duration_ms":68510,"significance":"If the second-order truncation is validated, RAZOR is an elegant and practical extension of MLIPs to biased electrochemical interfaces. The learning targets are clearly defined from DFT, the additional data cost is modest (less than a doubling, by the authors' estimate), and the approach is architecture-agnostic. The MD result is a concrete, falsifiable prediction (site switch in a narrow charge window) that matches experiments, and the agreement with AITD provides a useful internal consistency check between the learned MLIP and static DFT inputs. The main reservation is that the truncation and the configuration-independence of the capacitance are not tested against explicit finite-q DFT data, so the quantitative site-switch prediction rests on an assumption that is load-bearing for the headline application.","major_comments":[{"comment":"The central validation of the second-order truncation is not independent. The comparison to AITD in Fig. 2 (bottom) uses the same quadratic expansion of Eq. (1) and the same configuration-independent C_el,0; agreement between RAZOR-MLIP and AITD therefore establishes internal consistency between the learned MLIP and the static DFT inputs, but it does not test whether cubic terms or site-dependent second-order terms are negligible over the charge range where the site switch occurs (0.02 to 0.05 e per Cu atom). The manuscript reports no direct finite-q DFT energies for bridge and hollow reference structures. Since the additional DFT calculations at ±q used to construct Z* are already available, the authors should use them (or new calculations at representative q values) to compare the predicted E(q) of Eq. (1) with explicit DFT energies, and promote this comparison to the main text.","section":"Eq. (1), Fig. 2 bottom"},{"comment":"By construction, RAZOR's relative stability of bridge and hollow at fixed q is governed solely by E0 and phi0, because C_el,0 is taken to be configuration-independent and the quadratic term therefore cancels in energy differences between configurations of the same composition. This assumption is plausible for a metal electrode but is not tested. If the true second-order response (or the DL capacitance) differs between adsorption sites, the predicted switch could be an artifact. The authors should quantify the site dependence of C_el,0 (e.g. from finite-q DFT calculations for the two ordered c(2x2) structures) or otherwise bound the error this assumption introduces in the switch charge range.","section":"Paragraph beginning 'This leaves the electronic capacitance...'"}],"minor_comments":[{"comment":"The definition of the OH-bond angle phi and the fractional coverage shown in the top panel should be stated explicitly in the caption; the green line appears to be a smoothed or running quantity.","section":"Fig. 2 caption"},{"comment":"The exchange P_z <-> eps0 A Delta phi0 would be clearer if it stated explicitly that P_z is an extensive quantity for the simulation cell and A is the cell surface area, to avoid confusion with intensive potential drops.","section":"After Eq. (8)"},{"comment":"It would be helpful to state explicitly that Eq. (11) is exact under the RAZOR model's assumptions rather than an additional approximation, since this is the key step that again relies on the quadratic truncation.","section":"Paragraph preceding Eq. (11)"},{"comment":"The term 'non-Nernstian' is used without definition; a sentence explaining that the site preference changes with phi_E at fixed pH, or that the adsorption onset shifts nonlinearly with pH, would clarify the claim.","section":"Final paragraph before 'In summary'"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a condensed-matter/materials journal and the central idea is sufficiently distinct from prior ML electric-field response work. The main revision should focus on independent validation of the second-order expansion; the existing AITD comparison is not sufficient for that purpose. I do not see a fundamental flaw in the method itself, so with added finite-q DFT validation of the truncation and the site-independence of the capacitance, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The core idea is genuinely neat: instead of trying to learn a charged interface wholesale, RAZOR learns the work function and Born effective charges as linear-response corrections to a standard field-free MLIP, then uses the quadratic Taylor expansion to get finite-bias energies. That is a real extension of the response-learning MLIP literature (Gastegger, Gigli, Falletta, etc.) to extended metal/electrolyte interfaces, and the paper states the formalism cleanly. The Helmholtz relation (Eqs. 6-8) linking the work function difference to the dipole and the atomic Z* is used sensibly, and the MD application shows a sharp bridge-to-hollow switch for OH on Cu(100) over a narrow charge range, with the resulting pH-dependent adsorption-site preference matching the experimental papers they cite.\n\nSo the paper deserves credit for a promising method and a well-chosen demonstration.\n\nThe soft spot is exactly the one the stress-test note identifies. The central truncation E(q)=E0 + phi0 q + q^2/(2C_el,0) is what the architecture enforces, because C_el,0 is treated as configuration-independent and only E0 and phi0 are learned. The paper claims agreement with AITD 'supports the validity' of the second-order expansion, but the AITD reference uses the same quadratic energy model and the same implicit-solvent capacitance. That agreement is an internal-consistency check, not an independent test of the truncation. There are no direct finite-q DFT energies for the ordered bridge and hollow structures, so we do not know whether a cubic term or a site-dependent quadratic term is negligible over the 0.02-0.05 e/Cu range where the switch happens. Given that the site switch is the headline application, this is a load-bearing assumption.\n\nTwo smaller issues: the free-energy curves come without error bars, and there is no code or data release. Both are addressable, but they make it harder to assess the reliability of the MD result.\n\nOverall, this is a solid methods paper with one questionable validation loop. It deserves a serious referee — the idea is important and the authors know the literature — but I would send it back with a request for direct finite-q validation, error bars, and data/code before the quantitative claims are taken at face value.","headline":"RAZOR is a smart, clean way to add finite-bias response to MLIPs, but the headline site-switch result rests on a quadratic truncation that the paper's own AITD comparison cannot independently validate.","tokens_in":10755,"tokens_out":2334,"would_cite":true,"duration_ms":26381,"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":"RAZOR machine-learns the work function and Born effective charges of an electrified interface, extending machine-learned interatomic potentials to finite electrode bias and explaining the pH-dependent adsorption site of OH on Cu(100).","keywords":["machine-learned interatomic potentials","work function","Born effective charges","electrified interfaces","OH adsorption on Cu(100)","constant-potential molecular dynamics","implicit solvation"],"falsifier":"Run explicit constant-charge density-functional theory (not relying on the quadratic expansion) for OH on Cu(100) at several charges spanning the claimed transition, for example $q=0$, $+0.02$, $+0.05$, and $+0.1\\,e$ per Cu surface atom, and compare the relaxed adsorption sites and relative free energies with the RAZOR-MLIP predictions at the same $q$. If the explicit calculations do not show the bridge-to-hollow switch, or place it at a different charge, the second-order response assumption is falsified.","tokens_in":9788,"feed_emoji":"⚡","tokens_out":11187,"duration_ms":58882,"temperature":0.7,"pith_summary":"The paper introduces RAZOR, a response-augmented machine-learning scheme that makes fast interatomic-potential models aware of the excess charge carried by an electrified electrode. Instead of training on many charged configurations, it learns the work function, the first-order change in energy when a bias charge $q$ is added, and uses Born effective charges, the cross-derivatives of forces with respect to $q$, to stabilize that learning. With a constant electronic capacitance for the quadratic term, the resulting surrogate is correct to second order in $q$ at a modest increase in training cost. Applied to OH adsorbed on Cu(100) in implicit water, RAZOR-based molecular dynamics finds that the preferred adsorption site switches from bridge at negative charge to hollow at positive charge, which the paper connects to the experimentally observed pH-dependence of the adsorption site at fixed electrode potential.","feed_headline":"Machine learning explains copper's pH-dependent OH adsorption site","feed_subtitle":"RAZOR learns the work function as charge response, reproducing the observed bridge-to-hollow OH switch.","key_machinery":"The load-bearing object is the second-order Taylor expansion of the interface energy in excess bias charge $q$, combined with an equivalence between $q$-response and uniaxial electric-field response. For a planar metal electrode the relevant response is the scalar work function $\\phi_0^\\alpha=\\partial E^\\alpha/\\partial q|_{q=0}$, and its atomic derivatives are Born effective charges $Z^*_i$, which enter force labels as $\\partial F_i/\\partial q$ and stabilize learning of $\\phi_0^\\alpha$ exactly as force labels stabilize ordinary machine-learned potentials. The Helmholtz relation $P_z=\\epsilon_0 A\\,\\Delta\\phi_0$ lets the work function be treated as an extensive sum of atomic contributions, and the electronic capacitance $C_{\\mathrm{el},0}$ supplies the second-order term; because it is taken to depend only on composition, not on configuration, it drops out of fixed-composition force rankings and does not need to be learned per atom. Thermodynamic integration over $\\langle\\phi\\rangle_q$ and a Legendre transform then convert constant-charge simulations into constant-potential free energies.","core_discovery":"On the paper's own terms, the central claim is that the energetics of a charged metal/electrolyte interface can be captured by learning response coefficients at the charge-free state rather than by sampling many biased states. RAZOR writes the energy as $E^\\alpha(q)=E^\\alpha_0+\\phi^\\alpha_0\\,q+\\tfrac12\\,C^{-1}_{\\mathrm{el},0}\\,q^2$, where $\\phi^\\alpha_0$ is the work function (interfacial potential drop) and $C_{\\mathrm{el},0}$ the electronic capacitance; it then machine-learns $\\phi^\\alpha_0$ and the atomic Born effective charges $Z^*_i=-\\partial\\phi^\\alpha_0/\\partial r_i=\\partial F_i/\\partial q$ from local descriptors, while treating $C_{\\mathrm{el},0}$ as a composition-dependent constant. In the showcase system, 0.5 monolayer OH on Cu(100), RAZOR-MLIP molecular dynamics at 300 K shows a sharp bridge-to-hollow site switch within a narrow charge window near $q\\approx0.02$--$0.05\\,e$ per Cu surface atom, accompanied by a $\\sim1.5$ V work-function change; after Legendre transformation to constant potential and reference-electrode alignment, the predicted pH-dependence of the preferred site matches the two experimental studies cited by the paper.","pith_inferences":["Editorial: the most transferable output may be the local work-function decomposition itself; if it proves accurate on other surfaces, it could give a direct mapping from MLIP configurations to local electrochemical driving forces, a use the paper only names as future work.","Editorial: the configuration-independent treatment of $C_{\\mathrm{el},0}$ will likely be the first assumption challenged at low coverage or with mobile adsorbed water, where the double-layer response depends on local structure; making the capacitance a learned descriptor-dependent quantity is a natural stress test.","Editorial: the same response-learning logic could extend to semiconductor or two-dimensional-material electrodes, where Born effective charges remain well defined but the capacitance physics differs from that of a metallic double layer."],"forward_implications":["RAZOR can be added to any existing machine-learned interatomic potential, giving finite-bias energetics with only a modest premium in training data (roughly three charged-state DFT calculations per structure, reduced in practice by reusing the neutral-density SCF initialization).","Constant-charge molecular dynamics at applied bias becomes feasible for time scales of hundreds of picoseconds, long enough to observe adsorbate site interconversion that direct ab initio molecular dynamics cannot afford.","The learned decomposition of the work function into atomic contributions provides a route to local interfacial potential drops without relying on arbitrary electronic charge-density partitioning schemes.","For OH on Cu(100), the charge-induced site switch translates into a non-Nernstian pH-dependence of the adsorption site, matching the experimentally inferred preference for bridge sites in alkaline conditions and hollow sites in acidic conditions."],"supporting_citations":[{"why":"Supplies the local-environment neural-network architecture that RAZOR builds on for the base energy and force model.","marker":"[49]"},{"why":"Provides the prior machine-learning treatment of electric-field response through Born effective charges that RAZOR adapts to interface charge.","marker":"[31]"},{"why":"Supplies the implicit-solvent ab initio thermodynamics reference used for the charge-to-potential mapping and for validation.","marker":"[61]"},{"why":"Gives the series-capacitor decomposition of the electronic capacitance that justifies treating $C_{\\mathrm{el},0}$ as dominated by the double-layer term.","marker":"[44]"},{"why":"Reports the experimental pH-dependence of the OH adsorption site on Cu(100) that RAZOR rationalizes.","marker":"[63]"},{"why":"Second experimental study of OH adsorption on Cu(100) used to confirm the bridge-to-hollow site-switch interpretation.","marker":"[71]"}],"fun_headline_variants":["RAZOR ML decodes copper's pH-dependent OH site switch","AI predicts OH flips sites on copper as pH changes","Machine learning cracks copper's OH binding pH puzzle","Charged-up ML explains copper's OH site switch","RAZOR learns work function to unravel Cu's OH pH dependence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the energy of the charged interface is exactly quadratic in the added electrode charge $q$, with a fixed curvature, over the whole charge range studied; if higher-order terms or a curvature that depends on atomic arrangement matter, the predicted adsorption-site switch could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["RAZOR ML decodes copper's pH-dependent OH site switch","AI predicts OH flips sites on copper as pH changes","Machine learning cracks copper's OH binding pH puzzle","Charged-up ML explains copper's OH site switch","RAZOR learns work function to unravel Cu's OH pH dependence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001105,"raw_usage":{"total_tokens":4590,"prompt_tokens":911,"completion_tokens":3679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":3596}},"tokens_in":527,"tokens_out":3679,"duration_ms":21055,"temperature":1.0,"reasoning_tokens":3596,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:07:32.182898+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run explicit constant-charge density-functional theory (not relying on the quadratic expansion) for OH on Cu(100) at several charges spanning the claimed transition, for example $q=0$, $+0.02$, $+0.05$, and $+0.1\\,e$ per Cu surface atom, and compare the relaxed adsorption sites and relative free energies with the RAZOR-MLIP predictions at the same $q$. If the explicit calculations do not show the bridge-to-hollow switch, or place it at a different charge, the second-order response assumption is falsified.","supporting_citations":[{"cited_title":"Batzner, A","cited_arxiv_id":null,"evidence_quote":"Supplies the local-environment neural-network architecture that RAZOR builds on for the base energy and force model."},{"cited_title":"Falletta, A","cited_arxiv_id":null,"evidence_quote":"Provides the prior machine-learning treatment of electric-field response through Born effective charges that RAZOR adapts to interface charge."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the implicit-solvent ab initio thermodynamics reference used for the charge-to-potential mapping and for validation."},{"cited_title":"Binninger, Phys","cited_arxiv_id":null,"evidence_quote":"Gives the series-capacitor decomposition of the electronic capacitance that justifies treating $C_{\\mathrm{el},0}$ as dominated by the double-layer term."},{"cited_title":"Kunze, V","cited_arxiv_id":null,"evidence_quote":"Reports the experimental pH-dependence of the OH adsorption site on Cu(100) that RAZOR rationalizes."},{"cited_title":"Cruickshank, D","cited_arxiv_id":null,"evidence_quote":"Second experimental study of OH adsorption on Cu(100) used to confirm the bridge-to-hollow site-switch interpretation."}],"review_version":1}