{"id":"d63160a0-280f-40da-a28a-f219c55ab197","arxiv_id":"2502.02413","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-valued machine-learned dipole model with oxidation-number corrections enables electric-field-driven molecular dynamics for liquids and solids, demonstrated on water and LiNbO3.","lead":"Researchers built a machine-learning model that tracks electric dipoles in periodic materials even when atoms move across cell boundaries, by adding an oxidation-number term to a MACE neural network. They used it to run nanosecond simulations of water and LiNbO3 under electric fields, showing effects like field-enhanced ordering and transient polarization changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear-coupling Hamiltonian (Eq. 6) with zero-field dipoles is the load-bearing approximation for the showcased strong-field results; at 0.15–0.30 V/Å, neglected electronic polarizability is a ~10% effect and is named by the authors as a likely cause of the missing ~1000 cm^-1 water band.","rationale":"The reader's weakest_assumption identifies the same point I would stress-test: the linear-coupling field Hamiltonian. I agree that this is the most load-bearing approximation because every high-field quantitative result passes through it, while the multi-valued dipole architecture (Eq. 4) is independently supported by the SI failure of the single-valued model and by the branch-matched dipole predictions. The paper's own caveat about the missing ~1000 cm^-1 band is the clearest internal signal that the approximation is not fully converged at the applied intensities. I do not see an internal inconsistency in Eq. 4; the oxidation-number term is a standard branch decomposition, and the training curves in Fig. S1 support it. The appropriate remedy is additional validation of field-dependent dipoles and BECs at selected snapshots, not rejection of the method. Thus the CONDITIONAL verdict stands.","tokens_in":23278,"tokens_out":15679,"duration_ms":163144,"concrete_test":"Select 50–100 snapshots from the 0.15 V/Å water trajectory and the 0.30 V/Å LiNbO3 pulse trajectory. For each snapshot, compute the dipole and Born effective charges with a finite-field DFT method (e.g., the constant-field approaches in Refs. [7–9] of the manuscript) and compare them with the zero-field-trained ML dipole/BEC predictions. If the mean field-induced dipole change exceeds the reported dipole RMSE (3.12 mD/atom for water, 11.38 mD/atom for LiNbO3) or the BEC change exceeds about 0.1 e, then the linear-coupling Hamiltonian is not quantitatively reliable at the advertised field strengths, and the missing ~1000 cm^-1 band and the LiNbO3 transient polarization should be revisited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The multi-valued dipole construction (Eq. 4) is physically motivated, and the SI's demonstration that a single-valued model cannot train on far-from-equilibrium data (Fig. S1) is credible evidence in its favor. The load-bearing step for the quantitative showcase, however, is the Hamiltonian used in all field-driven simulations: H = T + V(R) - E·mu(R) (Eq. 6), with mu and its derivative (the BEC, Eq. 8) taken from a model trained on zero-field DFT data. This is the electric-dipole approximation in the linear-response regime. The paper applies it at 0.15 V/Å for water and up to 0.30 V/Å for the LiNbO3 pulses. At these intensities the field is not negligible relative to intramolecular fields: a water molecule's polarizability volume (≈1.44 Å^3) yields an induced dipole of order 0.14 D at 0.3 V/Å, roughly 8% of the permanent dipole, and in LiNbO3 the electronic dielectric response contributes to the polarization at a level that is not negligible compared with the transient changes reported in Fig. 3c. The authors themselves note in Section II B that omission of second-order electric-field coupling may explain the missing ~1000 cm^-1 band in their high-field water IR spectra. Thus if nonlinear electronic polarization or field-dependent BECs are non-negligible, the dielectric function, electrofreezing rates, and LiNbO3 switching dynamics would shift. This would not invalidate Eq. 4 as a way to represent the topological branch structure, but it would undercut the 'quantum-mechanical accuracy' claimed at the showcased field strengths and would require adding polarizability terms or training dipoles under finite fields.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a machine-learning framework for electric-field-driven nuclear dynamics in liquids and solids. The key methodological contribution is a multi-valued dipole model: for periodic systems, the dipole is written as mu_tilde_MV(R) = mu_tilde(A) + e sum_I N_I R_I (Eq. 4), where the residual term mu_tilde depends on the local atomic environment and the oxidation-number term accounts for the branch structure of the polarization. The model is autodifferentiated to obtain Born effective charges. The authors showcase the approach on liquid water, computing the GHz-THz dielectric function, field-dependent IR spectra, electrofreezing, and nuclear quantum effects, and on LiNbO3, computing the ferroelectric-paraelectric transition and THz-driven phonon dynamics. They report that the method enables nanosecond-scale simulations and that a full polarization switch in LiNbO3 is not achieved.","tokens_in":23613,"tokens_out":5958,"duration_ms":62344,"significance":"If the quantitative claims hold, the multi-valued dipole construction is a genuine and useful contribution: it addresses a known failure mode of single-valued ML dipole models in periodic systems with ionic transport or large displacements, and the learning-curve evidence in Supplementary Section S1 is compelling. The autodifferentiation of the model preserves the acoustic sum rule and avoids separate BEC fitting. The applications to water and LiNbO3 are ambitious and demonstrate a promising simulation capability. However, several load-bearing aspects of the validation and of the physical approximations need to be strengthened before the quantitative conclusions can be accepted.","major_comments":[{"comment":"The field-driven simulations use the linear-coupling Hamiltonian H = T + V - E·mu with field-independent Born effective charges obtained from a model trained on zero-field DFT data. This approximation is applied at fields up to 0.15 V/angstrom for water and up to 0.30 V/angstrom for LiNbO3. At these intensities the electronic polarizability is not negligible: an induced dipole of roughly 0.14 D at 0.3 V/angstrom is about 8% of the permanent dipole of a water molecule, and in LiNbO3 the electronic dielectric response is comparable to the transient polarization changes shown in Fig. 3c. The authors themselves note in Section II.B that omission of second-order electric-field coupling may explain the missing ~1000 cm^-1 band in the high-field water IR spectra. This concern directly affects the quantitative claims for the Stark shifts, electrofreezing rates, and phonon-driving dynamics. Please provide a quantitative estimate of the nonlinear field response, for example by computing dipoles and Born effective charges with finite-field DFT at the field strengths used, or clearly restrict the conclusions to the linear-response regime.","section":"Section II.B and Eq. (6)"},{"comment":"Table III reports water energy and force RMSE values marked with an asterisk, and the caption states that these are RMSE on the training dataset rather than on a held-out test set. The abstract and introduction claim 'quantum-mechanical accuracy', but for water only the dipole error is a test-set error. Training-set RMSE does not provide a valid estimate of generalization error. Please provide test-set errors for the water energy and forces, and state how the test set was constructed and whether any structures generated under applied electric fields were included in the test set.","section":"Table III and Section IV.C"},{"comment":"The water dipole model used for the main-text dielectric function and IR spectra was trained on dipoles of aperiodic water droplets computed with the revPBE functional. Supplementary Section S3 shows that a dipole model trained on periodic bulk-water dipoles yields a static dielectric constant closer to experiment and better reproduces previous ab initio results. The main-text quantitative analysis of the dielectric function and electrofreezing therefore uses the less accurate droplet-trained model. The authors should present the main-text results obtained with the periodic-data model, or quantify the differences and justify the choice of the droplet-trained model for the central claims.","section":"Section IV.C and Supplementary Section S3"},{"comment":"The water dipole model is trained on dipoles computed with the revPBE functional on droplet structures, whereas the MLIP interatomic potential is trained on energies and forces from revPBE0-D3 using the data of Ref. [75]. Since the field-dependent forces in Eq. (7) combine V and mu from different electronic-structure descriptions, the consistency of this mixed-functional approach is not established. A validation on bulk water configurations with dipoles computed at the revPBE0-D3 level would help rule out systematic biases in the Born effective charges and hence in the field-induced forces.","section":"Section IV.C (water computational details)"}],"minor_comments":[{"comment":"The right-hand side of Eq. (1) is typeset as 'el'; it should read 'e l', where e is the elementary charge and l is a lattice vector.","section":"Eq. (1)"},{"comment":"The text after Eq. (3) contains a duplicated word: 'the more general case case where Nat atoms...'.","section":"Eq. (3)"},{"comment":"The main text states that PIMD simulations use 8 ring-polymer replicas, while Supplementary Section S4 refers to 'the mean over the 32 beads' for the quantum IR spectra. Please clarify the number of beads used for each set of simulations and explain any difference.","section":"Section IV.C and Supplementary Section S4"},{"comment":"The experimental data in Fig. 2a are at 20°C while the simulations are at 300 K; this temperature difference should be stated explicitly in the caption or text.","section":"Fig. 2a"},{"comment":"Reference [81] is given as a GitHub release name without a version or DOI; please provide a citable version or a repository URL with a stable identifier.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The methodological core is interesting and likely publishable after revision. The main risk is the strong-field linear-coupling approximation, and the authors' own caveat about the missing water band supports that concern. I would also encourage the editor to require test-set errors for the water MLIP and a discussion of the mixed revPBE/revPBE0-D3 training, since these directly affect the 'quantum-mechanical accuracy' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. First, the core method idea is sound: adding the oxidation-number term e sum N_I R_I to a MACE dipole model correctly captures the branch structure of the polarization, and the SI learning curves make a convincing case that a single-valued model cannot train on far-from-equilibrium structures. That is a genuine advance, and it should transfer to any equivariant architecture. Second, the showcase applications are good but not as clean as the abstract suggests; the water results in particular rely on a linear dipole-field coupling at fields where electronic polarizability is roughly a 10% effect, and the authors are honest that this likely explains the missing ~1000 cm^-1 band.\n\nWhat the paper does well: the theory is careful, the oxidation numbers are computed from DFT rather than fitted to the target property, the LiNbO3 phonon-driving results are physically plausible and reproduce the experimental finding of no full polarization switch, and the code situation is decent (i-PI extension released, MACE fork available). The phase-transition temperature lands close to experiment, though the authors themselves flag the partial luck in that.\n\nSoft spots, in proportion. Table III reports energy and force RMSE on the training set for water; that should be test-set or at least clearly labeled. The dielectric function has no error bars, and the main-text water results use a droplet-trained dipole model even though the SI shows a periodic-trained model reproduces the experimental dielectric constant more closely; using the better model would strengthen the central figure. Eight-bead PIMD is light for room-temperature water, and with no convergence check the NQE enhancement claim is suggestive, not proven. The load-bearing approximation is the linear-coupling Hamiltonian of Eq. 6; at 0.15-0.30 V/Å the neglected electronic polarizability shifts the quantitative field-driven results, which is exactly what the missing ~1000 cm^-1 band hints at. None of this invalidates Eq. 4 as a way to represent the topological branch structure; it just means the quantitative field-driven numbers may move when second-order terms are added.\n\nWho this is for: anyone working on ML dielectric response, field-driven molecular dynamics, or polar materials. It deserves a serious referee and should be publishable after a revision that addresses the water-model choice, the error bars, and the PIMD convergence check. I would engage with it.","headline":"The multi-valued dipole construction is the real advance here; the water and LiNbO3 applications are compelling but lean on a linear-coupling approximation that deserves scrutiny.","tokens_in":24195,"tokens_out":1702,"would_cite":true,"duration_ms":20730,"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":"By treating the dipole as multi-valued and encoding atomic oxidation numbers into the model, this paper makes machine-learned electric-field-driven molecular dynamics accurate for periodic systems far from equilibrium, demonstrated on…","keywords":["multi-valued dipole","electric-field-driven molecular dynamics","Born effective charges","dielectric response of water","electrofreezing","nuclear quantum effects","lithium niobate ferroelectric transition","THz phonon driving"],"falsifier":"Take the water or LiNbO$_3$ configurations from a field-driven trajectory and recompute the dipole with first-principles finite-field or Berry-phase methods at applied fields 0.05–0.30 V/Å; if $\\boldsymbol{\\mu}(\\mathbf{E})$ deviates from $\\boldsymbol{\\mu}(0)+\\mathbf{Z}^*\\cdot\\mathbf{E}$ beyond the model's training error, or if a second-order coupling term is needed to recover the experimentally reported ~1000 cm$^{-1}$ water band, the quantitative predictions of the method would need revision in that regime.","tokens_in":23018,"feed_emoji":"⚡","tokens_out":8547,"duration_ms":80838,"temperature":0.7,"pith_summary":"Machine-learned models of the electric dipole normally fail in periodic systems as soon as atoms move far from their equilibrium sites, because the dipole is a multi-valued function: the same atomic environment can sit on different branches differing by a polarization quantum. This paper establishes a way to learn that multi-valued function by adding an explicit oxidation-number term, $e\\sum_I N_I \\mathbf{R}_I$, to the neural-network dipole. With this modification, automatic differentiation produces Born effective charges that stay correct across large displacements, so the model can drive nanosecond-scale molecular dynamics under static and time-dependent electric fields at first-principles quality. The authors demonstrate the framework on liquid water, where they compute dielectric response from GHz to THz and an electrofreezing transition enhanced by nuclear quantum effects, and on LiNbO$_3$, where they reproduce the ferroelectric-to-paraelectric transition and show that ultrafast THz phonon driving reaches the paraelectric state only transiently without completing a polarization switch.","feed_headline":"Oxidation-number term unlocks field-driven MD for polar materials","feed_subtitle":"Water dielectric spectra and LiNbO3 phase transitions simulated with quantum accuracy for nanoseconds.","key_machinery":"The load-bearing object is a multi-valued dipole model: a dipole function $\\tilde{\\mu}_{\\mathrm{MV}}(\\mathbf{R})$ that is allowed to take different values at periodically equivalent atomic positions, with the branch structure encoded through atomic oxidation numbers $N_I$ via $\\tilde{\\mu}_{\\mathrm{MV}}=\\tilde{\\mu}(\\mathbf{A})+e\\sum_I N_I\\mathbf{R}_I$. The first term is learned by an equivariant message-passing neural network; the second term supplies the non-periodic part whose gradient gives the $N_I\\mathbf{1}_{3\\times3}$ contribution to the Born effective charges. This construction makes the dipole continuous along paths where atoms cross cell boundaries, lets training include strongly out-of-equilibrium structures, and guarantees charge conservation and translation invariance.","core_discovery":"The paper's central claim is that incorporating the multi-valued nature of the dipole makes machine-learned electric-field-driven molecular dynamics generally applicable to molecules, liquids, solids, and disordered systems, in and out of equilibrium. Standard equivariant models treat the dipole as a single-valued function of the atomic environment, which forces a discontinuous branch switch across periodic boundaries and causes autodifferentiated Born effective charges to deviate once displacements are large. The paper instead writes the modeled dipole as $\\tilde{\\mu}_{\\mathrm{MV}}(\\mathbf{R})=\\tilde{\\mu}(\\mathbf{A})+e\\sum_I N_I \\mathbf{R}_I$, with the oxidation numbers $N_I$ fixed by a line integral of the dipole around a closed path, so equivalent atomic positions can carry different dipole values on the same smooth branch. Automatic differentiation then gives $\\mathbf{Z}^*_I = \\frac{1}{e}\\frac{\\partial \\tilde{\\mu}}{\\partial \\mathbf{R}_I} + N_I \\mathbf{1}_{3\\times3}$, which preserves the acoustic sum rule and translation invariance. The demonstration is that this model, coupled to the linear-coupling Hamiltonian $H=T+V-\\mathbf{E}\\cdot\\boldsymbol{\\mu}$, reproduces water's dielectric dispersion and electrofreezing and LiNbO$_3$'s phase transition and non-equilibrium phonon dynamics over nanoseconds.","pith_inferences":["A reader can push further: the missing ~1000 cm$^{-1}$ water libration band at high fields, which the paper attributes to neglected second-order field coupling, is a concrete spectral test of the linear-coupling approximation; a finite-field DFT calculation of the dipole along the same trajectories would reveal whether $\\mathbf{Z}^*$ changes appreciably at these intensities.","The same oxidation-number construction should transfer to other polar and superionic systems, such as nanoconfined water or superionic ice, where electric-field tuning of proton transport is of interest; the paper's own discussion points in this direction.","The topological character of the line integral behind $N_I$ connects this ML architecture to the modern theory of polarization and quantized charge transport, so the recipe may also be used to enforce correct branch behavior in ML models of Wannier centers, dielectric tensors, and higher-order multipoles.","A direct test of the method's generality would be to train the multi-valued dipole on a known ionic conductor, run field-driven dynamics, and compare predicted ionic conductivity with experiment; success would confirm that the framework is not limited to water-like or ferroelectric polarizations."],"forward_implications":["For liquid water, the approach yields a dielectric function spanning 0.001–135 THz with a simulated onset of Debye relaxation that tracks experiment, plus IR spectra whose field-induced shifts (libration blue-shift, OH red-shift) reflect a Stark-like response.","Nuclear quantum effects, included via path-integral dynamics, make water easier to polarize under an applied field and strengthen the electrofreezing signature at larger intensities, reversing the zero-field ordering of diffusion coefficients.","For LiNbO$_3$, the model reproduces a second-order ferroelectric-to-paraelectric transition with a Curie temperature close to the experimental 1413 K, and ab initio-quality phonon bands over the full Brillouin zone.","Under ultrafast 18 THz pulses, the simulations show non-linear coupling of the driven QIR mode to the 7.4 THz polarization-reversal mode and only three other $A_1$ modes; the polarization is transiently driven to zero at 0.15–0.20 V/Å and relaxes in about 300 fs, while at 0.25–0.30 V/Å the system heats above $T_C$ and no full coherent switch occurs.","Because the model learns only the dipole and autodifferentiates, the acoustic sum rule and translational invariance of the Born effective charges are satisfied by construction, extending previous dipole and Born-charge ML models to cases where atoms diffuse or move far from equilibrium."],"supporting_citations":[{"why":"Rigorous definition of oxidation states in solids; supplies the relation between closed-path dipole changes and the integer multiples $N$ used in Eq. (4).","marker":"[35]"},{"why":"Supplies the equivariant message-passing architecture that the paper modifies to output the multi-valued dipole.","marker":"[39, 40]"},{"why":"Earlier machine-learned electric-field response model for condensed phases; provides the linear-coupling ansatz and a direct comparison for the water IR spectra.","marker":"[20]"},{"why":"Adapts branch-matching and fluctuation-dissipation procedures for dielectric response of BaTiO$_3$ that the current implementation builds on.","marker":"[13]"},{"why":"Simulation driver that implements the equations of motion for nuclei coupled to static and time-dependent electric fields.","marker":"[67]"},{"why":"Provides the revPBE0 water dataset used to train the interatomic potential for the water simulations.","marker":"[75]"},{"why":"Proposed the ultrafast THz excitation of the QIR mode in LiNbO$_3$ as a route to polarization switching; motivates the driven-phonon simulations.","marker":"[58]"},{"why":"Experiment reporting only transient and partial ultrafast polarization reversal in LiNbO$_3$; the simulations match this observation.","marker":"[32]"},{"why":"Berry-phase implementation used to compute dipole labels for LiNbO$_3$ training data.","marker":"[74]"},{"why":"Experimental dielectric permittivity of water used as the reference for the GHz–THz comparison.","marker":"[24]"}],"fun_headline_variants":["Multi-valued dipole model enables field-driven MD for polar systems","Smooth dipole branch unlocks nanosecond field-driven MD","Machine-learned dipole with oxidation numbers drives polar dynamics","Field-driven MD for liquids and solids via multi-valued dipole","Autodiff dipole model simulates field-driven phase transitions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the linear electric-dipole coupling in $H=T+V-\\mathbf{E}\\cdot\\boldsymbol{\\mu}$, with field-independent Born effective charges, remains a faithful description of the nuclear dynamics up to field strengths of 0.30 V/Å and frequencies of 18 THz; if nonlinear polarizability or field-dependent charge responses become important at these intensities, the quantitative field-driven results for water and LiNbO$_3$ would shift, even though the multi-valued dipole architecture itself would keep working.","fun_headline_variants_meta":{"raw":{"variants":["Multi-valued dipole model enables field-driven MD for polar systems","Smooth dipole branch unlocks nanosecond field-driven MD","Machine-learned dipole with oxidation numbers drives polar dynamics","Field-driven MD for liquids and solids via multi-valued dipole","Autodiff dipole model simulates field-driven phase transitions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1634,"prompt_tokens":1029,"completion_tokens":605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":526}},"tokens_in":645,"tokens_out":605,"duration_ms":6436,"temperature":1.0,"reasoning_tokens":526,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T12:13:33.061392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the water or LiNbO$_3$ configurations from a field-driven trajectory and recompute the dipole with first-principles finite-field or Berry-phase methods at applied fields 0.05–0.30 V/Å; if $\\boldsymbol{\\mu}(\\mathbf{E})$ deviates from $\\boldsymbol{\\mu}(0)+\\mathbf{Z}^*\\cdot\\mathbf{E}$ beyond the model's training error, or if a second-order coupling term is needed to recover the experimentally reported ~1000 cm$^{-1}$ water band, the quantitative predictions of the method would need revision in that regime.","supporting_citations":[{"cited_title":"Jiang, S","cited_arxiv_id":null,"evidence_quote":"Rigorous definition of oxidation states in solids; supplies the relation between closed-path dipole changes and the integer multiples $N$ used in Eq. (4)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier machine-learned electric-field response model for condensed phases; provides the linear-coupling ansatz and a direct comparison for the water IR spectra."},{"cited_title":"Litman, V","cited_arxiv_id":null,"evidence_quote":"Simulation driver that implements the equations of motion for nuclei coupled to static and time-dependent electric fields."},{"cited_title":"Cheng, E","cited_arxiv_id":null,"evidence_quote":"Provides the revPBE0 water dataset used to train the interatomic potential for the water simulations."},{"cited_title":"Subedi, Proposal for ultrafast switching of ferro- electrics using midinfrared pulses, Phys","cited_arxiv_id":null,"evidence_quote":"Proposed the ultrafast THz excitation of the QIR mode in LiNbO$_3$ as a route to polarization switching; motivates the driven-phonon simulations."},{"cited_title":"Carbogno, N","cited_arxiv_id":null,"evidence_quote":"Berry-phase implementation used to compute dipole labels for LiNbO$_3$ training data."},{"cited_title":"Artemov, The electrodynamics of water and ice , Vol","cited_arxiv_id":null,"evidence_quote":"Experimental dielectric permittivity of water used as the reference for the GHz–THz comparison."}],"review_version":1}