{"id":"524aed61-de6e-4cf1-a744-8c148f2bc53d","arxiv_id":"2608.03287","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An equivariant graph neural network predicts energies, forces, Born effective charges, atom-resolved charges and magnetic moments, and simulates field-driven polaron, phonon, and ionic dynamics.","lead":"This paper introduces EFR-GNN, a machine-learning model that simulates how electric fields move atoms and localized charges in materials over long times. It is tested on three cases: hole-polaron hopping in MgO, terahertz-driven phonon rotation in GaAs, and silver-ion drift in superionic silver iodide.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Field-driven dynamics rely on unvalidated linear-response coupling (PES + BEC·E); at applied fields up to 0.05 V/Å this may break down, so central claim lacks support.","rationale":"The reader's weakest_assumption correctly identifies the linear-response approximation as the central vulnerability. The paper's central claim is that EFR-GNN provides a unified framework for field-driven dynamics, but the field-coupling mechanism is strictly first-order (field-free PES + BEC·E). No finite-field DFT benchmark is provided to show that this approximation holds at the applied fields (0.01–0.05 V/Å). In MgO, the field-induced energy difference across a hop is comparable to the hopping barrier, a regime where nonlinear electronic effects are plausible. The authors' own Discussion acknowledges that field-induced charge redistribution is omitted, which directly undercuts the 'atom-resolved tracking' aspect under strong fields. The q_eff fit is a post-hoc rationalization that cannot serve as validation because it absorbs errors in the force model. The consistency between zero-field and drift mobility is encouraging but not a direct test of the linear-response assumption. The most decisive remedy is a direct finite-field DFT comparison, which would settle whether the method's central claim holds at the demonstrated field strengths. The GaAs thermostat issue is a real but separate concern about dephasing artifacts; it is not the primary load-bearing assumption because even if the thermostat were removed, the linear-response coupling might still be inadequate. Thus the reader's CONDITIONAL verdict is appropriate, and the proposed test would either confirm the approximation or reveal a fundamental limitation.","tokens_in":14405,"tokens_out":4939,"duration_ms":57591,"concrete_test":"Perform finite-field DFT (e.g., VASP with an applied sawtooth electric field) on ~50–100 configurations sampled from the field-driven MLMD trajectories for each system and field strength (MgO at ±0.01, ±0.03, ±0.05 V/Å; GaAs at 0.01 V/Å; AgI at ±0.03 V/Å). Compare the DFT energies and forces against EFR-GNN predictions (field-free PES gradient + predicted BEC·E). If the RMSE in forces exceeds the field-free force MAE (e.g., 13.7 meV/Å in MgO) or systematic energy deviations appear at high fields, the linear-response assumption is violated. As a secondary check, rerun the GaAs dephasing simulations in NVE to verify the thermostat is not dominating the damping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"EFR-GNN's field-driven forces are computed by adding a predicted Born-effective-charge term to a field-free potential-energy surface (Discussion: 'EFR-GNN propagates atomic motion on a field-free ground-state potential-energy surface, with field-induced forces introduced through configuration-dependent BECs'). This is a first-order, linear-response approximation. The manuscript gives no direct validation against finite-field DFT: all training data are field-free, and the only field-dependent tests are internal consistency (zero-field mobility vs. drift mobility) and a post-hoc fit of q_eff to the observed forward/backward hopping ratios. The q_eff fit cannot validate the force model because it absorbs any error in the BEC-field coupling or the PES. At the largest MgO field (0.05 V/Å), the electrostatic energy across a nearest-neighbor hop is q_eff E d ≈ 0.5 e × 0.05 V/Å × 2.95 Å ≈ 74 meV, comparable to the 85 meV barrier. In this regime higher-order polarizability or field-induced charge redistribution could reshape the barrier beyond the linear term. The paper explicitly disclaims such effects: 'field-induced charge redistribution or carrier relocalization at fixed geometry' are not captured. Therefore the central claim—supporting accurate long-time field-driven molecular dynamics with atom-resolved electronic tracking—is not fully substantiated for the field strengths used.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces EFR-GNN, a modular equivariant graph neural network with three parameter-separated branches: a field-free potential-energy surface (PES), atom-resolved Born effective charge tensors (BECs), and Bader charges/local magnetic moments. Field-induced forces are added as configuration-dependent BEC·E terms on top of the field-free PES. The method is demonstrated on three systems: hole-doped MgO (field-driven small-polaron hopping), GaAs (THz resonant excitation and helicity control of the Γ-TO phonon), and α-AgI (superionic Ag+ transport and field-driven drift). The paper reports low energy/force/BEC/charge MAEs, reproduces activation energies and temperature trends, and explicitly states the linear-response and non-propagated-electronic-descriptor limitations of the framework.","tokens_in":14742,"tokens_out":5303,"duration_ms":66926,"significance":"If the central claims hold, EFR-GNN is a valuable contribution to machine-learned molecular dynamics, combining electric-field response with atom-resolved electronic-structure descriptors in one framework. The three applications span distinct physical regimes (localized carrier hopping, coherent phonon control, ionic transport) and yield physically plausible results, including the directional hopping asymmetry in MgO, helicity-controlled coherent phonon rotation in GaAs, and the experimental activation energy trend in α-AgI. The paper is also commendable for its explicit limitations paragraph, which distinguishes the linear-response force model from full field-dependent electronic structure. The main risk is that the field-driven dynamics are not directly validated against finite-field first-principles references, so the quantitative claims rest on internal consistency and post-hoc fits rather than on an independent test.","major_comments":[{"comment":"The field-driven force model is a linear-response approximation: F_field = Z*·E added to a field-free PES, and all training data are field-free. No direct validation against finite-field DFT is provided. At the largest MgO field (0.05 V/Å), q_eff E d ≈ 74 meV versus an 85 meV barrier, so nonlinear effects are plausible in this regime. Please add a benchmark against finite-field DFT (e.g., field-induced forces on representative configurations or a finite-field barrier) or explicitly restrict the claims to a field range where linear response is justified. As written, the central claim of accurate long-time field-driven dynamics is not fully substantiated.","section":"Discussion; Methods; Fig. 3(a)"},{"comment":"The effective driving charge q_eff is obtained by fitting the model's forward/backward hopping ratio to the very simulated ratios it is then used to explain. This is a compact parametrization, not an independent validation of the BEC-force coupling. Please state explicitly that q_eff is a fitted descriptor (not a predicted charge), and, if possible, compute the field-induced hopping bias directly from the BEC branch without free parameters, or compare q_eff with a value inferred solely from the BEC tensors. The current text implies more explanatory power than the fitting procedure supports.","section":"Fig. 3(d) and nearest-neighbor model"},{"comment":"The coherent-phonon dephasing time is extracted from the post-pulse decay in NVT simulations with a Nosé–Hoover-chain thermostat. Thermostatting can artificially damp coherent oscillations, potentially biasing the reported ~2.0 ps dephasing time. A control in NVE (or with the thermostat applied only before the pulse) is needed to confirm that the decay is intrinsic. Without this, the quantitative agreement with the experimental 2.1 ps may be fortuitous.","section":"GaAs, Fig. 4(e,f)"}],"minor_comments":[{"comment":"The sentence beginning 'For each primitive-cell Ga–As basis pair l, u_s,l,α denotes…' is missing a verb and is hard to parse; please rephrase.","section":"Fig. 4(d) text"},{"comment":"The pulse carrier frequency f=7.9 THz is chosen to match the model's own finite-temperature Γ-TO frequency (Fig. 4c). This is reasonable given the good agreement with experiment, but the text should state clearly that the resonance is by construction, not a prediction. The off-resonant control (3.95 THz and 15.8 THz) is a good check and should be highlighted.","section":"GaAs THz pulse setup"},{"comment":"Some reference DOIs appear to be placeholders (e.g., Refs. 11, 12, 19, 35–37). Please verify all DOIs before publication.","section":"References"},{"comment":"'Data available from the corresponding authors upon reasonable request' is weak for a methods paper. Consider depositing the trained models and key datasets in a public repository to strengthen reproducibility.","section":"Data availability"},{"comment":"Minor typo: 'sufficiently' should be 'sufficiently'.","section":"α-AgI subsection"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising framework and three diverse applications, but the lack of finite-field DFT validation and the circular q_eff fit are load-bearing for the central claims. If the authors can add a concrete finite-field test (even on a single system) and clarify the thermostat effect on dephasing, the paper would be suitable for publication. I would also encourage code/data release, as the manuscript currently offers only 'available upon request,' which may limit impact. The references are mostly appropriate, but some DOIs look like placeholders."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2608.03287. The genuinely new thing here is architectural: three parameter-separated equivariant branches in one GNN that output field-free energy/forces, Born effective charges, and DFT-supervised Bader charges/magnetic moments, with field forces added as BEC·E. That combination isn't in the cited literature, and the paper uses it well on three distinct systems. Polaron hopping in MgO, THz-driven coherent phonons in GaAs, and Ag+ transport in α-AgI all behave sensibly; the AgI mobility overshoot against experiment is admitted, and the phonon dephasing at 300 K (≈2.0 ps) matches the measured ≈2.1 ps. The modularity is a real contribution: you can activate only the branches you need.\n\nThe soft spots are real but not disqualifying. The field enters as a first-order BEC·E coupling on a field-free PES, trained entirely on field-free DFT. There is no direct check against finite-field DFT. At 0.05 V/Å in MgO the electrostatic energy across a hop is ~74 meV, comparable to the 85 meV barrier, so higher-order polarization could matter. The paper is candid about this in the Discussion, so it's a stated limitation rather than a hidden one, but it means the central claim of accurate field-driven dynamics is only validated internally (zero-field vs drift mobility consistency). Second, the q_eff = 0.49 e fitted to the forward/backward hopping ratios and then used to 'rationalize' those ratios is more of a compact fit than an explanation; the authors do note it absorbs lattice relaxation and recrossing, so fine if framed as interpolation. Third, the GaAs dephasing times come from NVT runs; the thermostat can contribute to the decay. An NVE control would tighten this. Finally, no code or data is shipped, only 'available on reasonable request,' so none of this is independently reproducible as-is.\n\nNet: this is a competent, honest integration of existing capabilities with a useful modular design. It deserves peer review. I'd want to see finite-field validation, an NVE dephasing control, and ideally a public release of code/data, but those are revision requests, not reasons to reject. If I were working on MLMD for polarons or field-driven ionic transport I'd cite it despite those caveats.","headline":"A solid modular MLMD paper that does what it says within linear response, but the field-coupling is only validated internally and no code/data are shipped.","tokens_in":15240,"tokens_out":2858,"would_cite":true,"duration_ms":29879,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EFR-GNN extends machine-learning molecular dynamics to electric fields by learning Born effective charge tensors and atom-resolved charges, and it demonstrates field-driven polaron drift, THz phonon control, and superionic transport in long","keywords":["electric-field-driven molecular dynamics","graph neural network interatomic potential","Born effective charges","equivariant graph neural networks","hole polaron transport","terahertz coherent phonons","superionic conductors"],"falsifier":"Run direct finite-field first-principles molecular dynamics on hole-doped MgO at E = 0.05 V/Å (using a sawtooth field or Berry-phase polarization) and compare per-atom forces with EFR-GNN's prediction, field-free force plus $Z_i^*\\mathbf{E}$. If the BECs change with field strength or the polaron redistributes charge at fixed geometry, the forces will deviate beyond the model's force MAE, and the predicted drift mobility would be wrong.","tokens_in":14246,"feed_emoji":"⚡","tokens_out":7547,"duration_ms":85531,"temperature":0.7,"pith_summary":"The paper is trying to establish that a single equivariant graph neural network can serve as a machine-learning force field for finite-temperature dynamics under static or time-dependent electric fields, while also resolving where electronic charge and spin are localized. The central move is to keep the potential-energy surface field-free and introduce the field purely through learned Born effective charge tensors, so the field-induced force on each atom is a predicted tensor contracted with the applied field. On top of that, an atom-resolved branch predicts Bader charges and local magnetic moments, which lets the simulation track a localized carrier as it hops. The paper claims this combination reproduces known physics in three materials: field-rectified hole-polaron drift in MgO, resonant terahertz excitation and helicity-controlled rotation of the Γ-point transverse-optical phonon in GaAs, and temperature-dependent plus field-driven Ag+ transport in α-AgI. The Discussion explicitly limits the framework to a configuration-dependent linear-response regime, stating that the field does not reshape the ground-state potential-energy surface and that charges and moments are structure-dependent descriptors rather than propagated electronic degrees of freedom.","feed_headline":"One graph network unifies field-driven forces, charges, and spin","feed_subtitle":"Tracks polarons, phonons, and ions in MgO, GaAs, and AgI across nanosecond-scale simulations.","key_machinery":"The carrying mechanism is the modular, parameter-separated equivariant graph neural network: three branches share a graph representation but have separate weights, so the field-free PES, the atom-resolved electronic-state descriptors, and the response tensors can be trained independently and activated as needed. The load-bearing identity is the force decomposition $F_i = -\\partial U/\\partial \\mathbf{r}_i + Z_i^*\\,\\mathbf{E}$, where $U$ is a learned field-free potential and $Z_i^*$ is the predicted $3\\times 3$ Born effective charge tensor, a quantity that measures how strongly an atom's force responds to a uniform electric field. This identity converts an arbitrary external field into per-ato","core_discovery":"EFR-GNN is a graph neural network with three parameter-separated equivariant branches processing the same atomic graph. Branch 1 predicts the field-free energy and forces; Branch 2 predicts atom-resolved Bader charges and local magnetic moments; Branch 3 predicts the Born effective charge tensor of each atom. The field-induced force is $F_i^\\text{field}=Z_i^*\\,\\mathbf{E}$, and the total force is the sum of the field-free and field-induced contributions, with the predicted tensors acoustic-sum-rule corrected so a uniform field produces no net translation. The paper's claim is that this modular design is enough to describe non-equilibrium field-driven dynamics that previously required either f","pith_inferences":["Editorial inference: the same three-branch design should transfer to other polar or ionic materials whose field response is dominated by the linear BEC coupling, including ferroelectric switching and electrochemical interfaces, provided DFT labels for the new chemistries are available.","Editorial inference: because Branch 2 descriptors are not conditioned on the applied field, the framework cannot describe field-induced charge redistribution or carrier relocalization at fixed geometry; under strong fields or near dielectric breakdown this is where the predictions would first fail.","Editorial inference: the fitted effective driving charge of 0.49 e depends on the Bader partitioning and on the nearest-neighbor hopping model, so it should not be interpreted as a directly measurable physical charge; other partitions or models would shift it.","Editorial inference: the helicity-controlled rotating phonon results suggest a simulation-side protocol for preparing chiral phonon states, and a direct comparison with polarization-resolved electro-optic experiments would be a natural next test."],"forward_implications":["Static-field molecular dynamics becomes feasible over hundreds of picoseconds for polaronic and ionic systems where DFT trajectories would be prohibitively expensive.","The MgO result gives a microscopic decomposition of field-driven polaron drift: the field biases forward relative to backward nearest-neighbor hops, and the bias can be quantified by a simple 12-site statistical model.","Resonant THz control of a specific phonon mode, including helicity-selected rotation direction, can be simulated at finite temperature and related to experimental dephasing times.","The same trained model can run either field-free or field-driven MD, and the charge/moment branch can be switched off when only forces are needed.","A unified framework of this kind opens the possibility of studying couplings between carrier localization and lattice response that require both electric-field and electronic-state resolution."],"supporting_citations":[{"why":"Provides a unified differentiable-learning formulation of electric response, the capability dimension EFR-GNN extends to atom-resolved charges.","marker":"[16]"},{"why":"Introduces a universal machine-learning framework for atomistic response to external fields, motivating the field-dependent force route used here.","marker":"[17]"},{"why":"Shows perturbed neural network potentials with explicitly predicted Born effective charges for condensed-phase field response, the direct antecedent of Branch 3.","marker":"[18]"},{"why":"Demonstrates neural-network Born effective charge prediction for ion mobility under electric fields, a core benchmark for the BEC-to-force route.","marker":"[20]"},{"why":"Shows equivariant graph convolutional networks can represent Born effective charge tensors, the architecture family used by Branch 3.","marker":"[21]"},{"why":"Establishes MLMD tracking of small-polaron hopping with atom-resolved charges and magnetic moments, which EFR-GNN makes field-driven.","marker":"[26]"},{"why":"Applies charge-aware MLMD to polaron transport in TiO2, supplying the transfer-tracking analysis used for the MgO polaron.","marker":"[27]"}],"fun_headline_variants":["EFR-GNN: one model for field, charge, and spin dynamics","Neural net tracks polarons, phonons, and ions under electric fields","Unified GNN predicts field responses in MgO, GaAs, and AgI","Field-driven dynamics unified in a single graph neural network","Polaron, phonon, and ion dynamics unified by EFR-GNN"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise, stated in the Discussion, is that the electric field acts only through configuration-dependent Born effective charge tensors on a field-free ground-state potential-energy surface, with atom-resolved charges and magnetic moments treated as structure-dependent descriptors that do not themselves respond to the field at fixed geometry.","fun_headline_variants_meta":{"raw":{"variants":["EFR-GNN: one model for field, charge, and spin dynamics","Neural net tracks polarons, phonons, and ions under electric fields","Unified GNN predicts field responses in MgO, GaAs, and AgI","Field-driven dynamics unified in a single graph neural network","Polaron, phonon, and ion dynamics unified by EFR-GNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000827,"raw_usage":{"total_tokens":3457,"prompt_tokens":755,"completion_tokens":2702,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2613}},"tokens_in":499,"tokens_out":2702,"duration_ms":20427,"temperature":1.0,"reasoning_tokens":2613,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:38:42.204377+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run direct finite-field first-principles molecular dynamics on hole-doped MgO at E = 0.05 V/Å (using a sawtooth field or Berry-phase polarization) and compare per-atom forces with EFR-GNN's prediction, field-free force plus $Z_i^*\\mathbf{E}$. If the BECs change with field strength or the polaron redistributes charge at fixed geometry, the forces will deviate beyond the model's force MAE, and the predicted drift mobility would be wrong.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a unified differentiable-learning formulation of electric response, the capability dimension EFR-GNN extends to atom-resolved charges."},{"cited_title":"& Jiang, B","cited_arxiv_id":null,"evidence_quote":"Introduces a universal machine-learning framework for atomistic response to external fields, motivating the field-dependent force route used here."},{"cited_title":"Polaron Transport in TiO$_{2}$ from Machine Learning Molecular Dynamics","cited_arxiv_id":"2606.01763","evidence_quote":"Applies charge-aware MLMD to polaron transport in TiO2, supplying the transfer-tracking analysis used for the MgO polaron."}],"review_version":1}