{"id":"0930e5cc-028d-417a-a070-c1438a266d58","arxiv_id":"2610.23643","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DPA4C-LR combines a strictly local equivariant backbone with analytically derived long-range electrostatics and Born effective charges, scaling to 5.2 million atoms on four GPUs while accurately inferring electrical response from energy and force training.","lead":"This paper presents a method to integrate long-range electrostatics and electrical response into strictly local machine learning interatomic potentials while preserving spatial decomposability for distributed computation. It enables million-atom simulations under electric fields, bringing atomistic modeling closer to device-relevant scales.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's weakest assumption—that locally predicted latent charges are insufficient for nonlocal charge redistribution—is precisely stated and acknowledged by the authors. I agree this is the principal limitation, but it does not undermine the central claim for the systems tested (insulators and dielectrics). The paper provides adequate evidence (BEC RMSE, IR spectra, P–E loops) that the method works where its assumptions hold. The technical derivations appear correct, and the code/data release supports reproducibility. Therefore, the reader's verdict of ACCEPT remains appropriate; no adjustment is warranted.","tokens_in":21832,"tokens_out":4045,"duration_ms":34712,"concrete_test":"Reproduce the DPA4C-LR scaling benchmark on 8 or 16 GPUs for the same water system (1.96 million atoms) to verify that the parallel efficiency does not drop by more than 10 percentage points relative to the 4-GPU measurement (43%), thereby confirming that the PME communication overhead does not degrade scalability at larger GPU counts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper presents a technically sound and clearly argued method for incorporating long-range electrostatics and electrical response into strictly local equivariant MLIPs while preserving spatial decomposability. The central innovation—analytical differentiation to decouple the global electrostatic solve from local backpropagation—is well-motivated and validated. The benchmarks for bulk water, PMN, LNOC, and Si/diamond are thorough, and the limitations (nonlocal charge redistribution, medium-range coupling, reliance on a fitted ε∞) are explicitly acknowledged. No internal inconsistency or unsupported leap undermines the central claim. The framework’s accuracy relative to DFPT for water and PMN is convincing, the IR spectra provide independent validation, and the scalability results, though limited to four GPUs, are consistent with the backbone’s known behavior. The LNOC conductivity is a qualitative demonstration, but the paper does not overclaim its certainty. Overall, the argument holds up under scrutiny.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper presents a machine learning interatomic potential framework (DPA4C-LR) that augments a strictly local equivariant backbone (DPA4C) with long-range electrostatics and Born effective charge (BEC) inference while preserving spatial decomposability for distributed computation. The key innovation is analytical differentiation of the long-range energy, so that forces, stresses, and BECs decompose over spatial partitions like the short-range part. The method is validated on bulk water (BEC RMSE 0.05 e, IR spectra), relaxor ferroelectric Pb(Mg1/3Nb2/3)O3 (P–E hysteresis), amorphous solid electrolyte Li0.9NbO0.9Cl4.1 (ionic conductivity 6.8±0.7 mS/cm), and Si/diamond interfacial thermal transport (YB-corrected electrostatics). Scalability up to 5.2 million atoms on four GPUs with 43% parallel efficiency is demonstrated.","tokens_in":22133,"tokens_out":2984,"duration_ms":29215,"significance":"If the claims hold, this work removes a key obstacle to using equivariant MLIPs for device-scale simulations: it shows that long-range electrostatic interactions and electrical response can be added without sacrificing the domain decomposition that makes short-range models scalable. The explicit BEC expression (Eq. 10) is cell-size independent and distributable, which avoids finite-size artifacts of prior latent Ewald approaches. The applications (relaxor hysteresis, ionic conductivity, interfacial thermal transport) are at scales relevant to device design and are enabled by the method. The code and data are publicly available. The paper is a strong candidate for publication in a high-impact journal.","major_comments":[{"comment":"The high-frequency permittivity ε∞ for water (3.07) and for PMN (√ε∞ = 3.80) is obtained by a least-squares fit of predicted charges against DFPT BECs (Fig. 2d). This introduces a parameter fitted to the same kind of reference data that is later used for BEC validation. While the independent IR spectra (Fig. 2e) provide corroboration, the BEC RMSE values (0.05 e for water, 0.22 e off-diagonal for PMN) are not fully independent. The authors should clarify whether the reported RMSE is computed after the ε∞ fit, and discuss the sensitivity of the BEC predictions to the fitted ε∞. For the LNOC application, no DFPT reference is available; is ε∞ estimated from literature, a different fit, or treated as a free parameter? The text (line 156) says '√ε∞ = 2.49 (Fig. S2)', but this figure is not included in the manuscript I reviewed. Please provide the determination procedure for ε∞ in LNOC.","section":"§III.A and §III.B, Eq. (9)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The strong scaling efficiency from 1 to 4 GPUs is 43% (speedup 1.73× on 4 GPUs). This is modest and the paper does not discuss how performance would degrade on larger GPU counts (e.g., 64 or 256 GPUs) where communication overhead becomes dominant. The weak scaling data are only shown up to 4 GPUs. For the claim 'scalable to multi-million-atom systems across GPUs', the reader needs a clearer discussion of the scaling bottlenecks (global PME FFT, ghost-zone communication) and how the architecture would fare on the 1000+ GPU setups that short-range DPA4C achieves (Ref. 24). The authors should either provide scaling data for more GPUs or explicitly qualify the scalability statement to the 2–4 GPU regime.","section":"§III.A, Fig. 2f–h and §IV","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The LNOC training and validation use MACE-MH-1 labels rather than DFT. The authors cite Wang et al. to argue that MACE-MH-1 dynamics reproduce experimental local structure, but the force RMSE (79 meV/Å) is evaluated against MACE-MH-1, not DFT. The paper should explicitly state that all LNOC results are contingent on the accuracy of the foundation potential, and the conductivity prediction (6.8±0.7 mS/cm) is a zero-shot prediction of the combined DPA4C-LR + MACE-MH-1 pipeline, not a DFT-validated result. This is not a fatal flaw—the method itself is validated on DFT for water and PMN—but the text at lines 157–159 currently states the LNOC model was 'trained on the energies and forces of the remaining frames, without supervision of charges or BECs', which may give the impression of a DFT-level validation. A clear caveat should be added.","section":"§III.C, §IV","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The decomposition of Z* into a direct charge contribution and a response contribution is clear. However, the derivation that the global charge-projection term cancels from the pairwise response expression (Eq. 11) would benefit from a short justification in the main text or a cross-reference to the SI section where the full derivation is given. Currently the text simply states the result.","section":"§II.C, Eq. (10)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The validation of the Yeh–Berkowitz correction shows that the uncorrected 3D Ewald energy deviates by 0.88 eV at the production vacuum gap. It would be helpful to state in the caption what the total long-range energy of the 32 nm cell is, so the reader can judge the fractional error. Also, the text says 'the maximum absolute deviation from the 2D reference' but the inset shows values up to ~10^-9 eV; please confirm the y-axis label.","section":"§III.D, Fig. 5d","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison with LES (Eq. 15) and the cell-size independence of the real-space expression (Eq. 10) is compelling. However, the figure caption and the text should specify that the BEC shown is for Nb in a 15-atom Pb3(Mg1Nb2)O9 cell. Is this the only element tested? The claim that Eq. (10) gives a cell-size-independent BEC is central; a sentence stating that the same behavior holds for all atom types in the test would strengthen the point.","section":"§IV, Fig. 6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation for the high-frequency permittivity changes between √ε∞ (Eq. 9, line 156) and ε∞ (line 82 and Fig. 2d caption). This is not inconsistent but could be confusing. Please unify the notation in the text, using ε∞ consistently and stating explicitly when the square root is used.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption for Fig. 2 says 'DPA4C-LR for bulk water. (a)–(c) Test-set RMSEs ... (d) Diagonal BECs ... (e) Simulated IR spectra ...' but panels (f)–(h) are not described. Please add a description of the scaling panels, including the number of GPUs and the system sizes shown.","section":"§III.A, Fig. 2 caption","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Dear colleague,\n\nHere's my read on arXiv:2610.23643. The core idea—decoupling the global Ewald solve from local automatic differentiation so that long-range electrostatics and Born effective charges stay spatially decomposable—is a genuine technical advance. It solves a real bottleneck for scaling MLIPs with field response. The implementation on DPA4C plus PME shows strong-scaling efficiency around 40% on four GPUs, reaching 5.2M atoms. The water benchmark is the cleanest: trained only on energies and forces, it reproduces BECs with RMSE 0.05 e and field-dependent IR spectra. That part is solid.\n\nThe three applications demonstrate reach. The PMN relaxor shows a finite-size effect that disappears at 29k atoms, exactly the sort of problem this method is built for. The LNOC conductivity estimate (6.8 ± 0.7 mS/cm) is plausible but rests on MACE-MH-1 labels, not DFT; the authors note the comparison is qualitative, which is fair. The Si/diamond interface validates the Yeh–Berkowitz correction under 2D PBCs and yields consistent thermal conductances.\n\nNow soft spots. The epsilon_infinity fitting uses a least-squares fit against DFPT BECs, which introduces a mild circularity—the same BECs used for scaling are also used to assess accuracy. This is minor and acknowledged in practice. More substantively, the LNOC claim of a seven-fold narrower confidence interval from field-driven MD is real, but the underlying dynamics depend on the quality of the MACE-MH-1 training data; any systematic error there propagates. The scalability results are limited to four GPUs; beyond that, communication overhead may matter, though the method is architecturally ready.\n\nThe authors state two honest limitations: the single-cutoff receptive field and inability to handle nonlocal charge redistribution. These bound the method's applicability to systems without metal–electrolyte interfaces or strong medium-range correlations.\n\nWho should read this? Anyone building long-range MLIPs for device-relevant simulations, especially if they care about field coupling. The paper is technically sound and deserves a serious referee. The code and data are public, which strengthens the case.\n\nMy read: accept with moderate confidence.","headline":"Analytic factorization of global electrostatics and local backpropagation enables distributed long-range MLIPs with BEC inference; well-validated, with honest limitations.","tokens_in":22650,"tokens_out":2103,"duration_ms":22995},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Long-range electrostatics and field response are added to strictly local MLIPs while keeping spatial decomposability, scaling to 5.2M atoms and inferring Born charges from energy/force training alone.","keywords":["machine learning interatomic potentials","long-range electrostatics","Born effective charges","domain decomposition","equivariant neural networks","scalable molecular dynamics","field-driven simulations","relaxor ferroelectrics"],"falsifier":"A direct test would be to apply DPA4C-LR to a metal surface under bias and compare the field-induced charge rearrangement with DFT: if the predicted BECs and forces deviate systematically from DFT for the first few atomic layers, the local charge assumption is insufficient for that class of systems.","tokens_in":23341,"feed_emoji":"⚡","tokens_out":4029,"duration_ms":40183,"temperature":0.7,"pith_summary":"The paper develops a method to incorporate long-range electrostatics and electrical response into strictly local equivariant machine learning interatomic potentials (MLIPs) while preserving the domain decomposition that enables efficient parallel computation. By analytically separating the long-range forces and Born effective charges (BECs) into a global electrostatics calculation and strictly local derivatives of learned latent charges, the approach avoids the global computational graph that plagues prior methods. The resulting model, DPA4C-LR, reaches 5.2 million atoms on four GPUs with 43% parallel efficiency and infers BECs accurately from training on only energies and forces. The work demonstrates three large-scale applications—recovering the slim polarization hysteresis of relaxor PMN, predicting ionic conductivity in amorphous solid electrolytes, and modeling interfacial thermal transport—that were previously inaccessible to short-range MLIPs or limited by finite-size effects.","feed_headline":"Long-range MLIPs hit 5.2M atoms with field response","feed_subtitle":"Analytic charge separation gives parallel MLIPs that infer Born charges and scale to multi-million atoms.","key_machinery":"The central mechanism is the analytic separation of the long-range electrostatic contribution into a global classical electrostatics solve (using particle mesh Ewald or Yeh–Berkowitz correction) and strictly local derivatives of learned latent charges. The Born effective charge (BEC) is computed using a real-space expression (Eq. 10) that depends only on pairwise distance vectors and local charge gradients, making it cell-size-independent and distributable. This factorization, combined with the strictly local equivariant DPA4C backbone, allows the long-range augmentation to inherit the domain decomposition and communication pattern of the short-range model. The local latent charges are read from a descriptor within a single cutoff radius and projected to conserve total charge.","core_discovery":"The central claim is that long-range electrostatics and electrical response can be added to strictly local MLIPs without sacrificing the domain decomposition that makes them scalable, by analytically separating the long-range forces and Born effective charges (BECs) into a global electrostatics calculation and strictly local derivatives of learned latent charges. This formulation preserves spatial decomposability for energies, forces, stresses, and BECs, enabling distributed evaluation across multiple GPUs. The method outperforms prior long-range approaches: it avoids the finite-size effects of latent Ewald summation's BEC expression and removes the need for a whole-cell computational graph that hinders parallelization. Training on only energies and forces, the model infers BECs accurately and reproduces field-dependent infrared spectra. The authors demonstrate the framework on three systems—relaxor ferroelectric PMN, amorphous solid electrolyte LNOC, and Si/diamond interface—showing that multi-million-atom simulations with field coupling are now routine.","pith_inferences":["If the analytic separation is as general as claimed, it may allow other physically motivated long-range corrections (e.g., van der Waals, multipolar electrostatics) to be added to local MLIPs without breaking domain decomposition, opening the door to multi-physics models at scale.","The limitation to local charge readout suggests a natural next challenge: incorporating nonlocal charge equilibration (e.g., via a short-range neural network predictor of chemical hardness) while retaining decomposability through iterative local updates.","The demonstration on three distinct problems with different boundary conditions (3D periodic, slab, nonperiodic) suggests the method is robust, but the true test of transferability will be its performance on systems with strong charge flow, such as metal surfaces under bias, where the authors acknowledge the local charge approximation may fail."],"forward_implications":["Electrical-response inference from only energy and force labels becomes practical at device-relevant scales, enabling field-driven molecular dynamics for ferroelectric switching, ion transport, and dielectric response in billion-atom systems with further hardware scaling.","The real-space BEC expression eliminates finite-size errors present in latent Ewald summation, so BEC accuracy no longer depends on supercell size.","The factorization is general and can be applied to any strictly local backbone (e.g., Atomic Cluster Expansion), not just DPA4C.","The approach enables routine simulation of relaxor ferroelectrics, amorphous solid electrolytes, and interfacial thermal transport with ab initio accuracy at tens-of-nanometer scales.","By preserving spatial decomposability, the method naturally extends to distributed and GPU-accelerated environments, promising billion-atom simulations with long-range physics."],"supporting_citations":[{"why":"Provides the strictly local equivariant DPA4C backbone that this work extends.","marker":"[24]"},{"why":"Latent Ewald summation framework; the present approach improves upon it by avoiding the whole-cell computational graph.","marker":"[34]"},{"why":"Supplies the real-space BEC expression (Eq. 10) used in this work to compute BECs without finite-size effects.","marker":"[29]"},{"why":"Previous work on inferring BECs from MLIPs; provides the benchmark protocol and the phase-corrected LES BEC expression that this work replaces.","marker":"[32]"},{"why":"Particle mesh Ewald method used for long-range electrostatics in periodic systems.","marker":"[37, 38]"},{"why":"Yeh–Berkowitz correction for slab geometry, enabling 2D periodic boundary condition treatment.","marker":"[42]"},{"why":"Provides the PMN dataset and FIRE-Swap method for generating B-site ordering in relaxor simulations.","marker":"[48]"},{"why":"MACE-MH-1 foundation potential used as source of training labels for the LNOC amorphous electrolyte model.","marker":"[54]"}],"fun_headline_variants":["Analytic charge separation yields parallel MLIP at 5.2M atoms","MLIPs infer Born charges from forces alone, scale to 5.2M atoms","Multi-million-atom electrostatic MLIPs with field response","Parallel long-range MLIP reaches 5.2M atoms, learns electrical response","5.2M-atom MLIP reproduces polarization and IR spectra under fields"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The model assumes that latent charges read from a strictly local descriptor (within a single cutoff radius) are sufficient to represent the electrostatic response, which may fail for systems with nonlocal charge redistribution such as metals or metal–electrolyte interfaces.","fun_headline_variants_meta":{"raw":{"variants":["Analytic charge separation yields parallel MLIP at 5.2M atoms","MLIPs infer Born charges from forces alone, scale to 5.2M atoms","Multi-million-atom electrostatic MLIPs with field response","Parallel long-range MLIP reaches 5.2M atoms, learns electrical response","5.2M-atom MLIP reproduces polarization and IR spectra under fields"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":7e-05,"raw_usage":{"cost":6.973344e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":6.973344e-05,"upstream_inference_prompt_cost":3.278016e-05,"upstream_inference_completions_cost":3.695328e-05},"total_tokens":1695,"prompt_tokens":1084,"completion_tokens":611,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":522}},"tokens_in":1084,"tokens_out":611,"cost_source":"openrouter_usage","duration_ms":6857,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":522,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-19T13:41:30.258065+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A direct test would be to apply DPA4C-LR to a metal surface under bias and compare the field-induced charge rearrangement with DFT: if the predicted BECs and forces deviate systematically from DFT for the first few atomic layers, the local charge assumption is insufficient for that class of systems.","supporting_citations":[{"cited_title":"Universal Machine-learning Molecular Dynamics at the Speed of Empirical Potentials","cited_arxiv_id":"2608.19041","evidence_quote":"Provides the strictly local equivariant DPA4C backbone that this work extends."},{"cited_title":"Latent Ewald summation for machine learning of long-range interactions, npj Comput. Mater.11, 80","cited_arxiv_id":null,"evidence_quote":"Latent Ewald summation framework; the present approach improves upon it by avoiding the whole-cell computational graph."},{"cited_title":"qNEP: A highly efficient neuroevolution potential with dynamic charges for large-scale atomistic simulations, J. Chem. Theory Comput.22, 4787","cited_arxiv_id":null,"evidence_quote":"Supplies the real-space BEC expression (Eq. 10) used in this work to compute BECs without finite-size effects."},{"cited_title":"Ma- chine learning interatomic potential can infer electrical response, npj Comput. Mater.11, 384","cited_arxiv_id":null,"evidence_quote":"Previous work on inferring BECs from MLIPs; provides the benchmark protocol and the phase-corrected LES BEC expression that this work replaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Yeh–Berkowitz correction for slab geometry, enabling 2D periodic boundary condition treatment."},{"cited_title":"Intrin- sic structure of relaxor ferroelectrics from first principles, Phys. Rev. B114, 024204","cited_arxiv_id":null,"evidence_quote":"Provides the PMN dataset and FIRE-Swap method for generating B-site ordering in relaxor simulations."},{"cited_title":"P.; Simm, G.; Ortner, C.; Cs´ anyi, G","cited_arxiv_id":null,"evidence_quote":"MACE-MH-1 foundation potential used as source of training labels for the LNOC amorphous electrolyte model."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}