{"id":"971b0b30-ef6a-48f2-81c8-161b37e4e5a0","arxiv_id":"2507.14302","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LES augments short-range MLIPs with long-range electrostatics learned from energies and forces alone, improving accuracy and enabling Born effective charge and dipole prediction.","lead":"This paper packages the Latent Ewald Summation (LES) method as a library that adds long-range electrostatics to machine learning interatomic potentials using only energy and force training data. The authors show accuracy gains across four MLIP architectures and train a scalable organic-molecule model with better bulk liquid predictions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BEC/dipole inference is a derivative of a learned charge map; without a uniqueness/identifiability check, the matches to DFT could reflect correlated but non-unique charges, so the central claim needs one targeted test.","rationale":"The reader's weakest_assumption identifies the same core issue: field-independence and lack of an identifiability proof for latent charges. My analysis agrees that the strongest claim—that LES reliably infers BECs/dipoles from E/F data—depends on a uniqueness assumption that the paper does not establish. The paper's own Discussion admits the field-dependence limitation, so the concern is explicit in the manuscript. However, I do not see an internal inconsistency that would justify rejection: the benchmarks are real, the code is available, and the small-system results are internally consistent. The right verdict is CONDITIONAL: the central claim should be accepted only with an identifiability or charge-label validation test, or with softened language that BECs are 'computed' rather than 'uniquely inferred.' I therefore agree with the reader and recommend keeping the CONDITIONAL verdict.","tokens_in":22041,"tokens_out":1621,"duration_ms":541186,"concrete_test":"Run a charge-label/identifiability test on the water or dipeptide dataset: (1) compute DFT reference atomic charges (e.g., Hirshfeld or DDEC) and BECs for a held-out set; (2) train 3-5 LES-augmented models with identical hyperparameters but different random seeds; (3) compare the spread of predicted BECs/dipoles across seeds against the RMSE to DFT. If seed-to-seed BEC variation is comparable to or larger than the error vs. DFT, the agreement is not unique. Also, if possible, repeat with the LES charge head replaced by a separately initialized head while keeping the host MLIP fixed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that LES infers electrostatics, polarization, and BECs from energy and force data alone rests on the assumption that the latent charges q_i are identifiable from E and F labels. In LES, q_i is a learned function of local invariant features (Sec. II), so the long-range energy is a sum of pair/decomposed Coulomb terms. Because a gauge-like transformation of charges that preserves the total electrostatic energy (and its gradients) leaves E and F unchanged, many charge maps can fit the same labels. The paper never tests this identifiability assumption; it only reports that predicted charges/BECs match DFT for selected models. The reader's weakest_assumption already flags this and the field-dependence issue is explicitly conceded in the Discussion, but the paper still claims reliable BEC and dipole prediction. Without an explicit uniqueness/robustness test or a comparison against charge labels, the claim that LES 'infers' the correct electrostatics is too strong. A concrete check would be to train on a dataset where reference charges are available, then compare the LES charges/BECs against reference, and also to test whether different random seeds or feature maps that give near-identical E/F RMSE produce significantly different BECs. If BECs vary widely across equally well-fitting models, the BEC agreement is not evidence for unique inference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents LES as a standalone PyTorch module that adds an explicit Ewald or direct-sum long-range electrostatic term to short-range MLIPs by learning 'latent charges' q_i^LES from per-atom local invariant features. The authors integrate LES with MACE, NequIP, CACE, and CHGNet and benchmark these on bulk water, polar dipeptides, and Au2/MgO(001), reporting reductions in energy and force RMSEs as well as Born effective charges that agree with DFT references. They then train MACELES-OFF on the SPICE dataset and report test errors comparable to MACE-OFF23, accurate dipole and BEC predictions, improved bulk-liquid densities and enthalpies of vaporization, an IR spectrum for water, and qualitatively reasonable biomolecular simulations. The central claim is that long-range electrostatics, polarization, and Born effective charges can be inferred from energy and force data without auxiliary charge or dipole labels.","tokens_in":22353,"tokens_out":11488,"duration_ms":120531,"significance":"If the central claim holds, this is a valuable contribution: it makes long-range electrostatics a drop-in addition to diverse MLIP architectures, removes the need for charge or dipole training labels, and provides access to electrical response properties from standard E/F datasets. The manuscript is concrete and unusually reproducible: the LES library, integration patches, training scripts, and trained MACELES-OFF model are all released, and the benchmarks include independent DFT comparisons of BECs and dipoles, bulk-liquid thermodynamics, and long-timescale dynamics. The consistency of the accuracy gains across four architectures and the successful scale-up to the chemistry-diverse SPICE set are genuine strengths. The main caveats are that the 'inference' claim is not yet supported by an identifiability analysis of the latent charges, and the MACELES-OFF comparison is not an apples-to-apples short-range ablation. These issues are addressable but need to be fixed before the strong wording of the abstract is fully justified.","major_comments":[{"comment":"The central claim that LES 'infers' atomic charges, BECs, and dipoles from energy and force data requires an identifiability analysis that is not provided. Since q_i^LES is a learned function of local invariant features and the training labels are only total energies and forces, many different charge maps can leave E and F unchanged on the training distribution; Eqs. (4)-(5) define the reported dipole and BEC predictions directly as functions of this non-unique map. The external parity checks in Figs. 2b, 3b, and 6b-c show that a chosen model matches DFT BECs, but they do not establish that the E/F data uniquely determine those charges. Please add a concrete robustness test: train several LES models with different initializations or feature maps on the same dataset and show that near-identical E/F RMSEs imply similar charge/BEC predictions, or compare the learned latent charges against reference partial charges on a held-out set. Without such a test, 'LES infers ... BECs just by learning from energy and force training data' is too strong.","section":"Section II, Eqs. (4)-(5), and Section V"},{"comment":"The claim that MACELES-OFF is more accurate than 'its short-range counterpart' trained on the same dataset is not supported by a controlled comparison. MACELES-OFF uses r=4.5 Å with k=192, ℓ=1, float32, whereas MACE-OFF23(S) uses k=96, ℓ=0, float64 and MACE-OFF23(M) uses r=5.0 Å, k=128, ℓ=1. The reported improvements in Table III (e.g., PubChem force RMSE 35.26 vs 61.83 meV/Å) conflate the LES addition with changes in channel width, rotation order, cutoff, and numerical precision. Please train a short-range MACE baseline with the same hyperparameters and data split, or otherwise ablate the LES contribution. The same issue affects the 1FSV comparison in Fig. 10, which uses MACE-OFF24(M) trained on SPICE 2 rather than SPICE 1. Until such a baseline is provided, the abstract's superiority claim is not established.","section":"Section IV, Table II, and Fig. 5"},{"comment":"The statements that LES infers 'polarization' and enables simulations under applied electric fields are stronger than the implemented model, because the latent charges are fixed functions of local features and do not respond to an external or internal electric field. The Discussion concedes this ('the issue of the field-dependence of atomic charges, which LES currently does not handle'), but the Introduction and Abstract should not imply field-dependent polarization without qualification. Please either soften these claims or add a benchmark that quantifies the error incurred by the fixed-charge approximation in a system where field-induced charge redistribution is expected, such as an ion or a polarizable molecule in an external field.","section":"Section II and Section V"}],"minor_comments":[{"comment":"Energy RMSE values for the water benchmark are said to be uniformly low but are omitted from the figure; please report them in the text or in a table for reproducibility.","section":"Fig. 2c"},{"comment":"MACELES-OFF is not uniformly better than MACE-OFF23(L); for example, the QMugs energy RMSE is 0.94 vs 0.58 meV/atom. The text should mention this exception when describing the two models as comparable.","section":"Table III"},{"comment":"The RMSD discussion appears self-contradictory: a lower RMSD (4.43 Å) is described as drift/over-compaction while a higher RMSD (5.14 Å) is described as close to AMBER. Please clarify what is being measured and why the lower value is interpreted as a failure.","section":"Section IVc, Fig. 10"},{"comment":"The Ewald expression assumes charge neutrality or a compensating background for periodic systems, but the manuscript does not state whether the learned latent charges are constrained to sum to zero; please clarify this implementation detail.","section":"Section II, Eqs. (1) and (3)"},{"comment":"BEC prediction for periodic systems requires the high-frequency permittivity ε∞ as an external input; the text should state prominently in the Abstract or Results that this parameter is obtained from experiment or DFPT and is not inferred from energy/force labels alone.","section":"Eq. (5)"},{"comment":"The LES defaults σ=1 Å and dl=2 Å are used for all benchmarks, but no sensitivity analysis is provided for these Ewald parameters; a short sensitivity test would strengthen the claim that the reported gains are not artifacts of these choices.","section":"Section VI, Methods"}],"recommendation":"major_revision","confidential_remarks":"This is largely an application and software-demonstration paper building on the authors' earlier LES papers [25-27]. The editor should ensure that the novelty relative to Ref. [26] is made explicit; the main new contributions here are the universal library integration and the SPICE-scale MACELES-OFF model. Given that framing, the decisive review question is whether the benchmarks are controlled enough to support the advertised claims. The two uncontrolled comparisons (identifiability of latent charges and the lack of a same-hyperparameter short-range baseline) are fixable and should be required before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nIf you work on MLIPs, this is the paper that turns Latent Ewald Summation from a method into something you can actually use. The standalone library and patches for MACE, NequIP, CACE, and CHGNet are real, the benchmarks across all four show consistent force/energy improvements, and the BEC predictions on water and dipeptides line up with DFT. The code is public. That alone makes it worth a look.\n\nWhat's genuinely new: the integration layer, the systematic multi-architecture benchmark, and the MACELES-OFF model trained on SPICE with the liquid and protein simulations. The liquid density and heat-of-vaporization numbers are noticeably better than MACE-OFF23(M), and the IR spectrum from predicted BECs is a nice demonstration.\n\nNow the soft spots, in proportion.\n\nThe headline MACELES-OFF comparison is not a controlled ablation. They compare against published MACE-OFF23(S/M/L), which differ in cutoff, channels, and irrep order. They say errors are halved versus 23(S), but 23(S) has k=96 and l=0, so part of the gain could be from the bigger network. They show MACELES-OFF is about as fast as 23(S), which is fine, but the accuracy claim is confounded. A same-hyperparameter short-range baseline would settle it. This is not fatal—the small-system benchmarks have proper paired baselines—but the abstract's 'more accurate than its short-range counterpart' oversells a comparison that isn't apples-to-apples.\n\nThe BEC/dipole inference is the deeper issue. Since BECs are derivatives of the latent charges, and the charges are only constrained by E and F, there is no proof they are identifiable. Different charge maps can give the same energy and forces, and the paper doesn't test seed-to-seed variation or compare against reference charges on a dataset that has them. However, the external DFT comparison for BECs and dipoles is independent evidence that the learned charges are physically meaningful on these test sets—so the practical claim holds, but 'infers correct electrostatics' should be softened until identifiability is checked. The Discussion does concede field-dependence of charges, which is honest.\n\nMinor: single-run RMSEs with no error bars, and the CHGNet baseline appears to be used as a pretrained model rather than retrained from scratch, which muddies that comparison a bit.\n\nWho is this for? Anyone building or using short-range MLIPs who needs electrostatics without changing their training data. It's a strong engineering contribution, and the public code means the claims are checkable. A serious referee should engage and should ask for the same-hyperparameter ablation and an identifiability test before the strongest claims are accepted.","headline":"Useful, reproducible long-range electrostatics augmentation library with real benchmarks; the large-scale comparison is confounded and the BEC inference claim needs an identifiability test.","tokens_in":22853,"tokens_out":2363,"would_cite":true,"duration_ms":23044,"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":"This paper claims that a standalone Latent Ewald Summation (LES) module can graft long-range electrostatics onto any short-range machine learning interatomic potential, inferring atomic charges, dipoles, and Born effective charges from…","keywords":["machine learning interatomic potentials","long-range electrostatics","Ewald summation","latent charges","Born effective charges","foundation models","molecular dynamics","SPICE dataset"],"falsifier":"Train an LES-augmented model on a dataset that includes configurations under applied electric fields or strongly polarizable ions where reference charges change with field, and compare predicted BECs or long-range energies to DFT; if the fixed local-feature charges cannot reproduce the field-induced changes, the central claim about general electrostatic inference fails. A second check is whether different random initializations of the same LES model yield different latent charges but nearly identical energies, which would reveal that energy and force labels do not uniquely determine the charges on which the BEC predictions depend.","tokens_in":21873,"feed_emoji":"⚛️","tokens_out":5781,"duration_ms":59908,"temperature":0.7,"pith_summary":"Most machine learning interatomic potentials cut off interactions at a few angstroms and miss the long-range electrostatics that dominate systems like water, charged peptides, and doped oxide surfaces. This paper argues that those electrostatics can be restored by a drop-in module, Latent Ewald Summation (LES), that learns a latent charge for every atom from local invariant features and adds the resulting Ewald or direct Coulomb energy to the host model's short-range energy. Because the charges come from energy and force labels already present in standard datasets, LES needs no extra supervision from atomic charges, Wannier centers, or dipoles. The paper shows improved accuracy across MACE, NequIP, CACE, and CHGNet on water, dipeptides, and gold-on-MgO, and builds a SPICE-trained organic model (MACELES-OFF) that beats its short-range baseline, predicts dipoles and Born effective charges, and describes bulk water and molecular liquids better. A sympathetic reader would care because this could turn existing short-range datasets and models into electrostatic-aware potentials without new data collection.","feed_headline":"LES adds long-range electrostatics to any MLIP, no charge labels","feed_subtitle":"It infers atomic charges from energy and force data alone, improving accuracy and Born effective charges.","key_machinery":"The load-bearing object is the latent charge $q_i^{\\mathrm{les}}$, predicted from the host MLIP's local invariant atomic features $B_i$ via a neural network. These charges enter the total energy as a long-range correction: $E_{\\mathrm{lr}} = \\frac{1}{2\\varepsilon_0 V}\\sum_{0<k<k_c} \\frac{e^{-\\sigma^2 k^2/2}}{k^2}|S(k)|^2$ for periodic systems, with $S(k)=\\sum_i q_i e^{i k\\cdot r_i}$, and the corresponding screened direct sum for finite systems; total energy is $E_{\\mathrm{sr}} + E_{\\mathrm{lr}}$ and forces and stresses come from automatic differentiation. The same $q_i$ charges, via the polarization $P=\\sum_i q_i r_i$ (or its reciprocal-space analogue), define the Born effective charge tensors in Eqs. (4)-(5), so electrical response properties are derivatives of an energy that was fit only to energies and forces.","core_discovery":"The central claim is that long-range electrostatics, polarization, and Born effective charges (BECs) can be inferred from total energy and force training data alone, and that this inference works as a universal augmentation for any short-range MLIP that produces atomic features. LES assigns each atom a latent charge $q_i^{\\mathrm{les}}$ through a neural network on the host model's invariant features; the long-range energy is computed by Ewald summation in periodic cells or direct Coulomb sums with an erfc screening factor in finite systems, and added to the short-range energy before forces and stresses are obtained by automatic differentiation. BECs follow from differentiating the predicted polarization with respect to atomic positions, using the high-frequency permittivity as an extra input. The paper reports that LES-augmented models consistently reduce energy and force errors across architectures and systems, predict BECs and dipoles in good agreement with reference DFT, and that a large model trained on SPICE (MACELES-OFF) is more accurate than MACE-OFF on the same data, with better liquid densities and heats of vaporization and a qualitatively correct free-energy surface for alanine dipeptide.","pith_inferences":["Beyond the paper: the non-uniqueness of latent charges is not resolved, so several charge assignments could give nearly identical energies and forces while differing per-atom, which matters for BEC predictions on out-of-distribution configurations.","Beyond the paper: adding dipole labels, which the paper notes are often available, is a natural next step that could pin down charges and improve BEC transferability without requiring periodic-system charge labels.","Beyond the paper: extending LES to field-dependent charges, flagged by the authors as ongoing work, would open the door to simulations under applied electric fields and strongly polarizable interfaces, exactly the regimes the current local-feature assumption may bias.","Beyond the paper: the gold-on-MgO results suggest LES could replace explicit charge training in older-generation neural network potentials, potentially improving systems where charge equilibration was previously needed."],"forward_implications":["Any existing energy-plus-force dataset can be upgraded to include long-range electrostatics without collecting charge, dipole, or Wannier-center labels.","LES-augmented models should improve accuracy most where the baseline perceptive field is short, for example low-body-order or small-cutoff models, as seen in the gold-on-MgO benchmarks.","Electrical response properties such as dipoles, Born effective charges, IR spectra, and dielectric response become available from models trained only on energies and forces.","A large chemically diverse LES model (MACELES-OFF) can beat its short-range counterpart on organic molecules, and improves bulk liquid density and enthalpy of vaporization predictions.","Because LES is a standalone PyTorch module, it can be patched into new and existing MLIP architectures with minimal changes."],"supporting_citations":[{"why":"Introduces the Latent Ewald Summation method that defines latent charges and the Ewald long-range energy.","marker":"[25]"},{"why":"Shows that charges and long-range interactions can be learned from energies and forces, providing the basis for the benchmark claims.","marker":"[26]"},{"why":"Derives Born effective charge prediction from LES polarization, supporting the electrical-response claims.","marker":"[27]"},{"why":"Provides the Au2-MgO(001) dataset and the fourth-generation HDNNP comparison used in the benchmarks.","marker":"[9]"},{"why":"Supplies the SPICE dataset used to train MACELES-OFF.","marker":"[42]"},{"why":"Defines the MACE-OFF baseline models whose test errors and liquid properties are compared against MACELES-OFF.","marker":"[43]"},{"why":"Supplies the RPBE-D3 bulk water dataset and the reference Born effective charges used for water benchmarks.","marker":"[51]"},{"why":"Provides the MACE architecture into which LES is patched.","marker":"[33]"},{"why":"Provides the NequIP architecture into which LES is patched.","marker":"[32]"},{"why":"Provides the CHGNet architecture used as a host baseline.","marker":"[48]"}],"fun_headline_variants":["LES infers electrostatics for any MLIP, no charge labels","Universal electrostatics: LES learns from energy and forces","Latent Ewald Summation: Electrostatics for all MLIPs","LES adds long-range electrostatics to any MLIP without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Each atom's latent charge is assumed to depend only on local invariant features, with no response to the environment's electric field; the paper explicitly says LES does not handle field-dependent atomic charges.","fun_headline_variants_meta":{"raw":{"variants":["LES infers electrostatics for any MLIP, no charge labels","Universal electrostatics: LES learns from energy and forces","Latent Ewald Summation: Electrostatics for all MLIPs","LES adds long-range electrostatics to any MLIP without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1767,"prompt_tokens":1017,"completion_tokens":750,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":674}},"tokens_in":633,"tokens_out":750,"duration_ms":7849,"temperature":1.0,"reasoning_tokens":674,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:00:02.020787+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an LES-augmented model on a dataset that includes configurations under applied electric fields or strongly polarizable ions where reference charges change with field, and compare predicted BECs or long-range energies to DFT; if the fixed local-feature charges cannot reproduce the field-induced changes, the central claim about general electrostatic inference fails. A second check is whether different random initializations of the same LES model yield different latent charges but nearly identical energies, which would reveal that energy and force labels do not uniquely determine the charges on which the BEC predictions depend.","supporting_citations":[{"cited_title":"Latent ewald summation for machine learning of long-range interactions,","cited_arxiv_id":null,"evidence_quote":"Introduces the Latent Ewald Summation method that defines latent charges and the Ewald long-range energy."},{"cited_title":"Learning charges and long-range interactions from energies and forces","cited_arxiv_id":"2412.15455","evidence_quote":"Shows that charges and long-range interactions can be learned from energies and forces, providing the basis for the benchmark claims."},{"cited_title":"A fourth-generation high- dimensional neural network potential with accurate electrostatics including non-local charge transfer,","cited_arxiv_id":null,"evidence_quote":"Provides the Au2-MgO(001) dataset and the fourth-generation HDNNP comparison used in the benchmarks."},{"cited_title":"Equiformer: Equivariant graph attention transformer for 3d atomistic graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the SPICE dataset used to train MACELES-OFF."},{"cited_title":"Spice, a dataset of drug-like molecules and peptides for training machine learning potentials,","cited_arxiv_id":null,"evidence_quote":"Defines the MACE-OFF baseline models whose test errors and liquid properties are compared against MACELES-OFF."},{"cited_title":"Short solvent model for ion correlations and hydrophobic association,","cited_arxiv_id":null,"evidence_quote":"Supplies the RPBE-D3 bulk water dataset and the reference Born effective charges used for water benchmarks."},{"cited_title":"E (3)-equivariant graph neural networks for data- efficient and accurate interatomic potentials,","cited_arxiv_id":null,"evidence_quote":"Provides the MACE architecture into which LES is patched."},{"cited_title":"Atomic cluster expansion for accurate and transferable interatomic potentials,","cited_arxiv_id":null,"evidence_quote":"Provides the NequIP architecture into which LES is patched."},{"cited_title":"A universal graph deep learning interatomic potential for the periodic table,","cited_arxiv_id":null,"evidence_quote":"Provides the CHGNet architecture used as a host baseline."}],"review_version":1}