Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Learning Non-Local Molecular Interactions via Equivariant Local Representations and Charge Equilibration

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read This paper claims that inserting a charge-equilibration layer into equivariant graph neural network potentials lets effectively local models capture long-range electrostatics and charge transfer, and demonstrates this on benchmarks where…

desk verdict CELLI is a useful, well-engineered Qeq layer for equivariant GNNs with strong benchmark gains, but the paper over-claims attribution to the Qeq solve without an ablation that isolates it. read the letter →

arxiv 2501.19179 v2 pith:GEC54J3D submitted 2025-01-31 physics.chem-ph cs.LGphysics.comp-ph

classification physics.chem-phcs.LGphysics.comp-ph
keywords chargeequilibrationlong-rangeinteractionsequivariantgraphneuralnetworksmachinelearningpotentialselectrostaticstransfermoleculardynamics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes CELLI, a model-agnostic neural-network layer that attaches to equivariant graph neural network potentials and gives them a non-local charge equilibration channel. The layer predicts environment-dependent atomic partial charges by minimizing a Qeq-style energy, computes a Gaussian-charge Coulomb energy from those charges, and feeds the charge environment back into the network's latent features. On benchmark systems designed to be unsolvable by strictly local methods, CELLI-enhanced Allegro and MACE reduce energy and force errors relative to their baselines, in some cases by large margins, while adding only modest computational cost. If correct, this offers a simple route for local machine-learning potentials to handle charge transfer, long-range electrostatics, and charge-state dependence, broadening their use to systems where those effects dominate.

What carries the argument

The central object is the charge equilibration layer, which takes scalar latent edge features, predicts per-edge contributions to electronegativity and hardness, assembles atomic electronegativities $\chi_i$ and hardnesses $J_i$, defines Gaussian charge radii from covalent radii, and solves for the partial charges $Q_i$ that minimize $U_{\mathrm{Qeq}} = U_{\mathrm{Coul}} + \sum_i (\chi_i Q_i + \frac{J_{ii}}{2} Q_i^2)$ subject to charge conservation. The resulting charges enter a learned Coulomb energy via a Gaussian-density interaction potential and are embedded into the latent feature update, so the layer carries global information while remaining a drop-in block for strictly local equivariant models.

What would settle it

Run CELLI on a system whose long-range electrostatics are dominated by anisotropic polarization or higher multipole moments, such as a water cluster in a strong non-uniform external field, and compare energy and force errors to a strictly local baseline; if the errors do not drop even when the model's charges fit reference values well, the claim that the scalar-charge Qeq channel captures the relevant long-range physics is falsified.

Watch

Extended reading notes

Core claim

CELLI generalizes the classical charge equilibration method to modern equivariant graph neural networks by splitting the total energy into a Coulombic part and a learned correction. The Coulombic part comes from partial charges obtained by globally minimizing a quadratic charge energy subject to charge conservation, with electronegativities, chemical hardnesses, and Gaussian charge radii learned from local edge features and species. These equilibrated charges are then embedded into the scalar latent features that the next layer consumes, so the short-range learned correction depends on the non-local charge environment. On carbon chains, silver clusters, sodium chloride clusters, and gold dimers on MgO(001) surfaces, CELLI-enhanced Allegro and MACE outperform their local baselines and often match or beat previous long-range-capable machine-learning potentials.

Load-bearing premise

The paper's non-local channel is a single scalar partial charge per atom, and the central assumption is that these scalar charges, together with the Gaussian-charge Coulomb energy, are expressive enough to represent the long-range physics that matters in the target systems.

Editorial extensions

If this is right

  • Strictly local equivariant models that incorporate CELLI can model long-range charge transfer and electrostatic interactions on the tested benchmarks, reaching errors several orders of magnitude below the local baseline in some cases.
  • Because CELLI adds only a marginal runtime overhead, it can make a small strictly local model competitive with much larger message-passing models on chemically diverse datasets such as OE62.
  • Integrating CELLI into message-passing models such as MACE improves accuracy on charge-transfer benchmarks, in some cases reducing errors by nearly a factor of ten.
  • The explicit charge predictions give CELLI-enhanced models interpretability and the ability to distinguish charge states, as demonstrated on charged silver and sodium chloride clusters.
  • CELLI-enhanced models produce stable 1 ns molecular dynamics simulations on SPICE subsets, suggesting the added non-local channel does not introduce simulation artifacts.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the non-local channel is a single scalar charge per atom, the method likely cannot represent anisotropic polarization or dipolar response; a tensor-valued charge or dipole layer would be a natural extension.
  • The explicit, differentiable charges could enable simulations in external electric fields and prediction of infrared spectra, applications the paper lists as future work but does not demonstrate.
  • The weak correlation between charge error and energy/force error in the MACE comparison suggests the performance gain may come substantially from the charge-conditioned latent update rather than from the physical accuracy of the charges; an ablation that removes the charge labels would isolate the mechanism.
  • CELLI could be combined with a dispersion correction to cover long-range van der Waals interactions, which are not part of the current energy model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces CELLI, a charge equilibration layer for equivariant graph neural network potentials. CELLI predicts environment-dependent electronegativities and hardnesses from latent edge features, solves a learned Qeq linear system to obtain partial charges, computes a Coulomb energy from those charges, and feeds the charges back into the latent feature updates of the GNN. The authors integrate CELLI into Allegro and MACE and evaluate it on four long-range benchmark systems from Ko et al. (carbon chains, silver clusters, NaCl clusters, gold dimers on MgO), on the OE62 organic-molecule dataset, and on SPICE-based molecular dynamics simulations. They report large accuracy improvements over strictly local baselines, favorable comparisons with 4G-HDNNP, LRSR, CACE-LR, and SpookyNet, modest computational overhead, and stable MD trajectories.

Significance. If the reported gains are causally due to the non-local charge equilibration, CELLI is a useful architectural contribution: it shows a model-agnostic way to add long-range electrostatics and charge transfer to modern equivariant GNNs while retaining end-to-end training, interpretable charges, and compatibility with local parallelization. The paper's strengths include parameter-matched local baselines, a wide range of benchmark systems including periodic and charged cases, explicit charge outputs that permit physical inspection, and a stability test in MD. The authors also honestly acknowledge that their charge errors are often larger than SpookyNet's even when their energy errors are lower, and that increasing cutoff changes some rankings. However, the absence of an ablation isolating the Qeq solve from the added charge-embedding MLPs, together with the lack of error bars and the depth/cutoff confounds in some comparisons, means the central claim that the non-local Qeq mechanism drives the improvements is not yet fully established.

major comments (4)
  1. [Results, 'Long-range interactions for message-passing models'; Methods, 'Hyperparameters'] The paper lacks an ablation that isolates the Qeq solve from the added charge-embedding machinery. The baseline Allegro and MACE models are obtained by replacing CELLI with an additional interaction layer, but this removes not only the equilibration but also the charge-dependent feature update yij = MLP_Q(Qi, Qj, ci, cj) and the MLP_x update. To attribute the benchmark gains to non-local equilibration, the authors should compare against a variant that keeps the same charge-embedding MLPs and the Coulomb term but replaces the equilibrated charges with charges predicted directly from local features (or otherwise bypasses the Qeq linear solve). Without this ablation, the 'overcoming inherent locality' claim is not yet supported, since the gains could come from the extra local charge-embedding capacity.
  2. [Results, 'Long-range interactions for message-passing models'; Table 2] The paper's own observation that charge errors do not consistently correlate with energy or force errors undermines the proposed mechanistic role of the Qeq charges. For carbon chains, CELLI(6) has a much larger charge RMSE than SpookyNet (1.273 vs 0.117 me) yet a lower energy RMSE (0.128 vs 0.364 meV/atom); for NaCl, the better-charge CELLI(6) (13.91 me) has a slightly worse energy RMSE than CELLI(2) (15.52 me, 0.097 vs 0.104 meV/atom). The authors should quantify, per system or per sample, whether charge accuracy is predictive of energy/force accuracy, and should state whether the reported gains remain if the charge-dependent features are fed with un-equilibrated local charges. This is load-bearing for the central claim that the non-local charge equilibration, rather than the Coulomb term with imperfect charges or the additional local embeddings, is the active ingredient.
  3. [Tables 1, 2, and 3] All benchmark results are reported as single point estimates without error bars or standard deviations over training seeds. Differences such as 0.599 vs 0.609 meV/atom in Table 1 and 55.3 vs 55.1 meV in Table 3 are too small to interpret without uncertainty estimates. In addition, Table 3 reports only energy MAE/RMSE and parameter counts, not force errors, so the claim of significant accuracy gains on OE62 is incomplete. The authors should report means and standard deviations over at least three independent training runs and include force RMSE for the OE62 comparison.
  4. [Results, 'Long-range interactions for message-passing models'; Table 2] The MACE comparisons confound model depth with the effect of CELLI. CELLI(6) has six message-passing layers, while baseline MACE has two, and the authors themselves note that the deeper models may achieve low errors because of their depth rather than CELLI. At matched depth, CELLI(2) improves over baseline MACE(2) on NaCl and AuMgO but performs worse on carbon chains (energy RMSE 0.398 vs 0.335 meV/atom). The statement that integrating CELLI can overcome the inherent locality of MACE on the benchmark systems is therefore too strong; the authors should report matched-depth comparisons for all systems and qualify the per-system claims.
minor comments (6)
  1. [Table 1 caption] The word 'Table' is misspelled as 'T able' in the caption.
  2. [Results, 'Benchmark systems with strictly local models'] The sentence 'The results confirm CELLI’s ability to model can effectively capture long-range charge transfer and electrostatics' is grammatically garbled and should be rewritten.
  3. [Discussion] The phrase 'In the feature, we plan' should be 'In the future, we plan'.
  4. [Abstract and Results] The abstract's claim of 'state-of-the-art results for strictly local models' should be qualified, since Table 3 includes non-local baselines and Table 1 shows that CACE-LR achieves a lower AuMgO error when its larger cutoff is used.
  5. [Figure 3 caption] The caption should clarify whether (a) shows runtime per structure for a single sample and (b) shows the optimal runtime over the listed batch sizes, and should state whether the timings include the Qeq solve and the SPME calculation.
  6. [Code Availability] The adapted models and training scripts are not publicly released; for a methods paper presenting a new architecture, releasing a documented implementation of CELLI (or at least the model definition) would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CELLI's Qeq charges are trained against separate energy, force, and charge targets, and the benchmark gains are empirical rather than built into the loss.

full rationale

The paper does not define its target quantity into its inputs. The CELLI layer computes partial charges by solving the Qeq linear system (Eq. 7) from learned electronegativities and hardnesses, then adds the Coulomb energy of Eq. (5) and a separately learned correction Delta U. All parameters are trained with the multi-target loss of Eq. (9) that includes reference energies, forces, and Hirshfeld charges as independent labels; nothing in this construction forces the Qeq charges to reproduce the energy, so the reported error reductions are empirical outcomes rather than tautologies. The closest potential concern is attribution: the paper claims the improvements stem from the Qeq non-local channel, and its own Table 2 discussion notes that charge errors do not consistently correlate with energy or force errors, which weakens the mechanistic interpretation. However, a weak attribution argument is not circular reasoning. The paper also cites its own software and related prior work (e.g., chemtrain, chemtrain-deploy), but these citations are not load-bearing for the central claim: they support implementation details and scalability statements, and the benchmarks against Allegro, MACE, SpookyNet, DimeNet++, and other published methods provide independent external evidence. No equation in the paper reduces a prediction to a fitted input by construction, and no uniqueness claim or ansatz is imported solely from the authors' prior work. Overall, the derivation chain is self-contained in the relevant sense, and the central claim, whether or not fully established, is not circular.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

CELLI introduces no new physical entities; its parameters are learned network weights and per-species scales. The main assumptions are inherited from charge equilibration: the quadratic charge-energy expansion, Gaussian charge densities with covalent radii, and the choice of DFT plus atomic charge partitioning as ground truth. One assumption specific to this paper is that scalar charges are a sufficient non-local message, which is the weakest point. A separate concern is the attribution of gains to CELLI versus extra model capacity, since some baseline comparisons replace CELLI with an additional interaction layer or use different cutoffs.

free parameters (5)
  • learnable electronegativity scale f = learned during training
    Multiply factor on summed per-edge electronegativity contributions in Section 'Environment Embedding'; a learned scalar parameter that controls charge response.
  • species hardness offset ~J_i^Z = learned per species
    Species-dependent hardness contribution combined via softplus with environment hardness; affects the Qeq linear system Eq. (7).
  • radius scaling factor ~s_i = learned per species
    Scales covalent radii gamma_i used in Gaussian charge widths in Eq. (5); the paper notes appropriate radii are crucial for optimization convergence.
  • loss weights gamma_U, gamma_F, gamma_Q = problem-specific, hand-set
    Balances energy, force, and charge targets in Eq. (9); chosen per dataset and not reported for all runs.
  • model hyperparameters (number of interaction layers, hidden sizes, cutoffs) = chosen similar to Ko et al. or Kosmala et al.; see Supplementary Table 2
    Architecture choices affect comparisons; some baseline Allegro variants include an additional interaction layer, confounding attribution of gains to CELLI.
assumptions (5)
  • domain assumption The charge-core interaction energy can be approximated by a second-order expansion in charges: U_Qeq = U_Coul + sum_i (chi_i Q_i + J_ii/2 Q_i^2) (Eq. 6).
    Standard Qeq approximation inherited from Rappe and Goddard; the paper assumes this form is sufficiently expressive for the systems studied.
  • domain assumption Electrostatic energy between partial charges is modeled as Gaussian charge densities with widths determined by covalent radii, Eq. (5).
    The erf-screened Coulomb form with radii gamma_i based on Pyykko covalent radii; enters the Qeq matrix (Eq. 7) and thus the charges.
  • domain assumption Reference charges (Hirshfeld for benchmarks, MBIS for SPICE) are valid training targets for the charge prediction head.
    The loss Eq. (9) includes charge MSE; if these atomic charge partitioning schemes are ambiguous, the model's charge branch is trained on convention-dependent labels.
  • domain assumption DFT references (PBE for benchmarks and OE62, omegaB97M-D3BJ for SPICE) are accurate ground truths for the problems tested.
    All RMSE values are measured against these DFT labels; no uncertainty in the reference data is propagated.
  • ad hoc to paper The non-local information required for long-range accuracy can be conveyed entirely through scalar per-atom charges Qi and the charge-dependent edge update.
    CELLI's only global channel is the Qeq charge solve; the latent update uses (Qi, Qj, ci, cj). If long-range effects require dipoles or higher multipoles, this representation is insufficient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Non-Local Molecular Interactions via Equivariant Local Representations and Charge Equilibration." pith.science (2026). https://pith.science/paper/GEC54J3D

@misc{pith2026250119179,
  author       = {Pith},
  title        = {Pith review of: Learning Non-Local Molecular Interactions via Equivariant Local Representations and Charge Equilibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GEC54J3D}},
  note         = {Machine review of arXiv:2501.19179}
}
read the original abstract

Graph Neural Network (GNN) potentials relying on chemical locality offer near-quantum mechanical accuracy at significantly reduced computational costs. Message-passing GNNs model interactions beyond their immediate neighborhood by propagating local information between neighboring particles while remaining effectively local. However, locality precludes modeling long-range effects critical to many real-world systems, such as charge transfer, electrostatic interactions, and dispersion effects. In this work, we propose the Charge Equilibration Layer for Long-range Interactions (CELLI) to address the challenge of efficiently modeling non-local interactions. This novel architecture generalizes the classical charge equilibration (Qeq) method to a model-agnostic building block for modern equivariant GNN potentials. Therefore, CELLI extends the capability of GNNs to model long-range interactions while providing high interpretability through explicitly modeled charges. On benchmark systems, CELLI achieves state-of-the-art results for strictly local models. CELLI generalizes to diverse datasets and large structures while providing high computational efficiency and robust predictions.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations

    physics.comp-ph 2025-06 conditional novelty 6.0 of 10

    A model-agnostic JAX-to-LAMMPS framework runs machine learning potentials in million-atom multi-GPU molecular dynamics with near-ideal strong and weak scaling.

Reference graph

Works this paper leans on

60 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    The Journal of Physical Chemistry A 127(11), 2417–2431 (2023) https://doi.org/10.1021/acs.jpca.2c06778

    Anstine, D.M., Isayev, O.: Machine learning interatomic potentials and long- range physics. The Journal of Physical Chemistry A 127(11), 2417–2431 (2023) https://doi.org/10.1021/acs.jpca.2c06778

  2. [2]

    https: //doi.org/10.48550/arXiv.2412.08541

    Frank, J.T., Chmiela, S., M¨ uller, K.-R., Unke, O.T.: Euclidean Fast Attention: Machine Learning Global Atomic Representations at Linear Cost (2024). https: //doi.org/10.48550/arXiv.2412.08541

  3. [3]

    Nature Communications 12(1), 398 (2021) https://doi.org/ 10.1038/s41467-020-20427-2

    Ko, T.W., Finkler, J.A., Goedecker, S., Behler, J.: A fourth-generation high- dimensional neural network potential with accurate electrostatics including non- local charge transfer. Nature Communications 12(1), 398 (2021) https://doi.org/ 10.1038/s41467-020-20427-2

  4. [4]

    The Journal of Chemical Physics 161(23) (2024) https://doi.org/10.1063/5.0235189

    R¨ ocken, S., Burnet, A.F., Zavadlav, J.: Predicting solvation free energies with an implicit solvent machine learning potential. The Journal of Chemical Physics 161(23) (2024) https://doi.org/10.1063/5.0235189

  5. [5]

    npj Computational Materials 10(1) (2024) https: //doi.org/10.1038/s41524-024-01251-4

    R¨ ocken, S., Zavadlav, J.: Accurate machine learning force fields via experimental and simulation data fusion. npj Computational Materials 10(1) (2024) https: //doi.org/10.1038/s41524-024-01251-4

  6. [6]

    Nature Communications 14(1), 579 (2023) https://doi.org/10.1038/ s41467-023-36329-y

    Musaelian, A., Batzner, S., Johansson, A., Sun, L., Owen, C.J., Kornbluth, M., Kozinsky, B.: Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications 14(1), 579 (2023) https://doi.org/10.1038/ s41467-023-36329-y

  7. [7]

    https://arxiv.org/abs/2206.07697

    Batatia, I., Kov´ acs, D.P., Simm, G.N.C., Ortner, C., Cs´ anyi, G.: MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (2023). https://arxiv.org/abs/2206.07697

  8. [8]

    Nature Communications 13(1), 2453 (2022) https://doi.org/10.1038/s41467-022-29939-5

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt, T.E., Kozinsky, B.: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications 13(1), 2453 (2022) https://doi.org/10.1038/s41467-022-29939-5

Show all 60 references
  1. [9]

    Journal of the American Chemical Society 147(21), 17598–17611 (2025) https://doi.org/ 10.1021/jacs.4c07099

    Kov´ acs, D.P., Moore, J.H., Browning, N.J., Batatia, I., Horton, J.T., Pu, Y., Kapil, V., Witt, W.C., Magd˘ au, I.-B., Cole, D.J., Cs´ anyi, G.: MACE-OFF: Short- Range Transferable Machine Learning Force Fields for Organic Molecules. Journal of the American Chemical Society 1...

  2. [10]

    npj Computational Materials 11(1), 49 (2025) https://doi.org/10.1038/s41524-025-01535-3 19

    Yang, Z., Wang, X., Li, Y., Lv, Q., Chen, C.Y.-C., Shen, L.: Efficient equivariant model for machine learning interatomic potentials. npj Computational Materials 11(1), 49 (2025) https://doi.org/10.1038/s41524-025-01535-3 19

  3. [11]

    In: Precup, D., Teh, Y.W

    Gilmer, J., Schoenholz, S.S., Riley, P.F., Vinyals, O., Dahl, G.E.: Neural message passing for quantum chemistry. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 1263–...

  4. [12]

    In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J

    Kosmala, A., Gasteiger, J., Gao, N., G¨ unnemann, S.: Ewald-based long-range message passing for molecular graphs. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th Inter- national Conference on Machine Learning. Pr...

  5. [13]

    npj Computational Materials 10(1), 47 (2024) https://doi.org/10.1038/s41524-024-01225-6

    Shaidu, Y., Pellegrini, F., K¨ u¸ c¨ ukbenli, E., Lot, R., Gironcoli, S.: Incorporating long-range electrostatics in neural network potentials via variational charge equi- libration from shortsighted ingredients. npj Computational Materials 10(1), 47 (2024) https://doi.org/10....

  6. [14]

    Chemi- cal Reviews 121(16), 10142–10186 (2021) https://doi.org/10.1021/acs.chemrev

    Unke, O.T., Chmiela, S., Sauceda, H.E., Gastegger, M., Poltavsky, I., Sch¨ utt, K.T., Tkatchenko, A., M¨ uller, K.-R.: Machine Learning Force Fields. Chemi- cal Reviews 121(16), 10142–10186 (2021) https://doi.org/10.1021/acs.chemrev. 0c01111

  7. [15]

    Chemical Reviews 118(4), 1691–1741 (2018) https: //doi.org/10.1021/acs.chemrev.7b00305

    Zhou, H.-X., Pang, X.: Electrostatic Interactions in Protein Structure, Folding, Binding, and Condensation. Chemical Reviews 118(4), 1691–1741 (2018) https: //doi.org/10.1021/acs.chemrev.7b00305

  8. [16]

    Journal of Molecular Liquids 351, 118616 (2022) https://doi.org/10.1016/j.molliq.2022.118616

    Baskin, I., Epshtein, A., Ein-Eli, Y.: Benchmarking machine learning methods for modeling physical properties of ionic liquids. Journal of Molecular Liquids 351, 118616 (2022) https://doi.org/10.1016/j.molliq.2022.118616

  9. [17]

    Biophysical Chemistry 77(1), 49–68 (1999) https://doi.org/10.1016/ S0301-4622(99)00010-1

    Gromiha, M.M., Selvaraj, S.: Importance of long-range interactions in protein folding. Biophysical Chemistry 77(1), 49–68 (1999) https://doi.org/10.1016/ S0301-4622(99)00010-1

  10. [18]

    Galante, M., Tkatchenko, A.: Anisotropic van der waals dispersion forces in poly- mers: Structural symmetry breaking leads to enhanced conformational search. Phys. Rev. Res. 5, 012028 (2023) https://doi.org/10.1103/PhysRevResearch.5. L012028

  11. [19]

    https://arxiv.org/ abs/2205.14276

    Frank, J.T., Unke, O.T., M¨ uller, K.-R.: So3krates: Equivariant attention for inter- actions on arbitrary length-scales in molecular systems (2023). https://arxiv.org/ abs/2205.14276

  12. [20]

    https://doi.org/10.48550/ arXiv.2211.16684 20

    Yu, H., Hong, L., Chen, S., Gong, X., Xiang, H.: Capturing Long-Range Inter- action with Reciprocal Space Neural Network (2022). https://doi.org/10.48550/ arXiv.2211.16684 20

  13. [21]

    https://doi.org/10.48550/arXiv.2409.17622

    Wang, Y., Cheng, C., Li, S., Ren, Y., Shao, B., Liu, G., Heng, P.-A., Zheng, N.: Neural P3M: A Long-Range Interaction Modeling Enhancer for Geometric Gnns (2024). https://doi.org/10.48550/arXiv.2409.17622

  14. [22]

    https://doi.org/10.48550/arXiv.2304.13542

    Li, Y., Wang, Y., Huang, L., Yang, H., Wei, X., Zhang, J., Wang, T., Wang, Z., Shao, B., Liu, T.-Y.: Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation (2024). https://doi.org/10.48550/arXi...

  15. [23]

    https://arxiv.org/abs/2502.13797

    Caruso, A., Venturin, J., Giambagli, L., Rolando, E., No´ e, F., Clementi, C.: Extending the RANGE of Graph Neural Networks: Relaying Attention Nodes for Global Encoding (2025). https://arxiv.org/abs/2502.13797

  16. [24]

    https://arxiv.org/abs/2502

    Zhdanov, M., Welling, M., Meent, J.-W.: Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems (2025). https://arxiv.org/abs/2502. 17019

  17. [25]

    The Journal of Chemical Physics 153(3), 034702 (2020) https://doi.org/10.1063/5.0005084 https://pubs.aip.org/aip/jcp/article- pdf/doi/10.1063/5.0005084/13464163/034702 1 online.pdf

    Rowe, P., Deringer, V.L., Gasparotto, P., Cs´ anyi, G., Michaelides, A.: An accurate and transferable machine learning potential for carbon. The Journal of Chemical Physics 153(3), 034702 (2020) https://doi.org/10.1063/5.0005084 https://pubs.aip.org/aip/jcp/article- pdf/doi/10...

  18. [26]

    Artrith, N., Morawietz, T., Behler, J.: High-dimensional neural-network poten- tials for multicomponent systems: Applications to zinc oxide. Phys. Rev. B 83, 153101 (2011) https://doi.org/10.1103/PhysRevB.83.153101

  19. [27]

    Coste, A., Slejko, E., Zavadlav, J., Praprotnik, M.: Developing an implicit solva- tion machine learning model for molecular simulations of ionic media. J. Chem. Theory Comput. 20(1), 411–420 (2023) https://doi.org/10.1021/acs.jctc.3c00984

  20. [28]

    Journal of Chemical Theory and Computation 20(5), 2088–2097 (2024) https://doi.org/10.1021/acs.jctc.3c01254

    Song, Z., Han, J., Henkelman, G., Li, L.: Charge-optimized electrostatic interac- tion atom-centered neural network algorithm. Journal of Chemical Theory and Computation 20(5), 2088–2097 (2024) https://doi.org/10.1021/acs.jctc.3c01254

  21. [29]

    Pl´ e, T., Lagard` ere, L., Piquemal, J.-P.: Force-field-enhanced neural network inter- actions: from local equivariant embedding to atom-in-molecule properties and long-range effects. Chem. Sci. 14, 12554–12569 (2023) https://doi.org/10.1039/ D3SC02581K

  22. [30]

    https://arxiv.org/abs/2412.15455

    Kim, D., King, D.S., Zhong, P., Cheng, B.: Learning charges and long-range interactions from energies and forces (2024). https://arxiv.org/abs/2412.15455

  23. [31]

    Nature Communications 12(1), 7273 (2021) https://doi.org/10

    Unke, O.T., Chmiela, S., Gastegger, M., Sch¨ utt, K.T., Sauceda, H.E., M¨ uller, K.-R.: Spookynet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nature Communications 12(1), 7273 (2021) https://doi.org/10. 1038/s41467-021-27504-0 21

  24. [32]

    Ghasemi, S.A., Hofstetter, A., Saha, S., Goedecker, S.: Interatomic potentials for ionic systems with density functional accuracy based on charge densities obtained by a neural network. Phys. Rev. B 92, 045131 (2015) https://doi.org/10.1103/ PhysRevB.92.045131

  25. [33]

    Nature Communications 13(1), 1572 (2022) https: //doi.org/10.1038/s41467-022-29243-2

    Gao, A., Remsing, R.C.: Self-consistent determination of long-range electrostatics in neural network potentials. Nature Communications 13(1), 1572 (2022) https: //doi.org/10.1038/s41467-022-29243-2

  26. [34]

    Scientific Data 7(1), 58 (2020) https://doi.org/10

    Stuke, A., Kunkel, C., Golze, D., Todorovi´ c, M., Margraf, J.T., Reuter, K., Rinke, P., Oberhofer, H.: Atomic structures and orbital energies of 61,489 crystal- forming organic molecules. Scientific Data 7(1), 58 (2020) https://doi.org/10. 1038/s41597-020-0385-y

  27. [35]

    Scientific Data 10(1), 11 (2023) https://doi.org/10.1038/ s41597-022-01882-6

    Eastman, P., Behara, P.K., Dotson, D.L., Galvelis, R., Herr, J.E., Horton, J.T., Mao, Y., Chodera, J.D., Pritchard, B.P., Wang, Y., De Fabritiis, G., Markland, T.E.: SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials. Scientific Data ...

  28. [36]

    Pyykk¨ o, P., Atsumi, M.: Molecular Single-Bond Covalent Radii for Elements 1–

  29. [37]

    The Journal of Chemical Physics 103(19), 8577–8593 (1995) https://doi.org/10.1063/1.470117

    Essmann, U., Perera, L., Berkowitz, M.L., Darden, T., Lee, H., Pedersen, L.G.: A smooth particle mesh Ewald method. The Journal of Chemical Physics 103(19), 8577–8593 (1995) https://doi.org/10.1063/1.470117

  30. [38]

    https://arxiv.org/abs/2506.04055

    Fuchs, P., Chen, W., Thaler, S., Zavadlav, J.: chemtrain-deploy: A parallel and scalable framework for machine learning potentials in million-atom MD simulations (2025). https://arxiv.org/abs/2506.04055

  31. [39]

    Computer Physics Communications 271, 108171 (2022) https: //doi.org/10.1016/j.cpc.2021.108171

    Thompson, A.P., Aktulga, H.M., Berger, R., Bolintineanu, D.S., Brown, W.M., Crozier, P.S., in ’t Veld, P.J., Kohlmeyer, A., Moore, S.G., Nguyen, T.D., Shan, R., Stevens, M.J., Tranchida, J., Trott, C., Plimpton, S.J.: LAMMPS - a flexible simulation tool for particle-based mate...

  32. [40]

    Journal of Chemical Theory and Computation 19(2), 562–579 (2023) https://doi.org/10.1021/acs.jctc.2c00661

    Th¨ urlemann, M., B¨ oselt, L., Riniker, S.: Regularized by Physics: Graph Neural Network Parametrized Potentials for the Description of Intermolecular Inter- actions. Journal of Chemical Theory and Computation 19(2), 562–579 (2023) https://doi.org/10.1021/acs.jctc.2c00661

  33. [41]

    22 https://doi.org/10.26434/chemrxiv-2024-bdfr0-v3

    Kabylda, A., Frank, J.T., Dou, S.S., Khabibrakhmanov, A., Sandonas, L.M., Unke, O.T., Chmiela, S., M¨ uller, K.-R., Tkatchenko, A.: Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields (2025). 22 https://doi.org/10.26434/chemrxiv-2024-bdfr0-v3

  34. [42]

    Chemical Science 8(10), 6924–6935 (2017) https://doi.org/10.1039/C7SC02267K

    Gastegger, M., Behler, J., Marquetand, P.: Machine learning molecular dynamics for the simulation of infrared spectra. Chemical Science 8(10), 6924–6935 (2017) https://doi.org/10.1039/C7SC02267K

  35. [43]

    Chemical Science 12(34), 11473–11483 (2021) https://doi.org/10.1039/D1SC02742E

    Gastegger, M., Sch¨ utt, K.T., M¨ uller, K.-R.: Machine learning of solvent effects on molecular spectra and reactions. Chemical Science 12(34), 11473–11483 (2021) https://doi.org/10.1039/D1SC02742E

  36. [44]

    https://doi.org/ 10.48550/arXiv.2102.03150

    Sch¨ utt, K.T., Unke, O.T., Gastegger, M.: Equivariant Message Passing for the Prediction of Tensorial Properties and Molecular Spectra (2021). https://doi.org/ 10.48550/arXiv.2102.03150

  37. [45]

    Jour- nal of Computational Physics 73(2), 325–348 (1987) https://doi.org/10.1016/ 0021-9991(87)90140-9

    Greengard, L., Rokhlin, V.: A fast algorithm for particle simulations. Jour- nal of Computational Physics 73(2), 325–348 (1987) https://doi.org/10.1016/ 0021-9991(87)90140-9

  38. [46]

    The Journal of Chemical Physics 110(6), 2810–2821 (1999) https://doi.org/10

    Martyna, G.J., Tuckerman, M.E.: A reciprocal space based method for treating long range interactions in ab initio and force-field-based calculations in clusters. The Journal of Chemical Physics 110(6), 2810–2821 (1999) https://doi.org/10. 1063/1.477923

  39. [47]

    The Journal of Physical Chemistry 95(8), 3358–3363 (1991) https: //doi.org/10.1021/j100161a070

    Rappe, A.K., Goddard III, W.A.: Charge equilibration for molecular dynamics simulations. The Journal of Physical Chemistry 95(8), 3358–3363 (1991) https: //doi.org/10.1021/j100161a070

  40. [48]

    Computer Physics Communications 104(1), 59–69 (1997) https://doi.org/10.1016/S0010-4655(97)00041-6

    Nakano, A.: Parallel multilevel preconditioned conjugate-gradient approach to variable-charge molecular dynamics. Computer Physics Communications 104(1), 59–69 (1997) https://doi.org/10.1016/S0010-4655(97)00041-6

  41. [49]

    Journal of Chemical Theory and Computation 20(16), 7264–7271 (2024) https://doi.org/10.1021/acs.jctc.4c00334

    Gubler, M., Finkler, J.A., Sch¨ afer, M.R., Behler, J., Goedecker, S.: Accelerating fourth-generation machine learning potentials using quasi-linear scaling particle mesh charge equilibration. Journal of Chemical Theory and Computation 20(16), 7264–7271 (2024) https://doi.org/...

  42. [50]

    https://doi.org/10.5281/zenodo.10975225

    Eastman, P., Behara, P.K., Dotson, D., Galvelis, R., Herr, J., Horton, J., Mao, Y., Chodera, J., Pritchard, B., Wang, Y., De Fabritiis, G., Markland, T.: SPICE 2.0.1. https://doi.org/10.5281/zenodo.10975225 . https://doi.org/10. 5281/zenodo.10975225

  43. [51]

    Computer Physics Communications 310, 109512 (2025) https://doi.org/10.1016/j.cpc.2025.109512

    Fuchs, P., Thaler, S., R¨ ocken, S., Zavadlav, J.: Chemtrain: Learning deep potential models via automatic differentiation and statistical physics. Computer Physics Communications 310, 109512 (2025) https://doi.org/10.1016/j.cpc.2025.109512

  44. [52]

    A framework for differentiable physics*

    Schoenholz, S.S., Cubuk, E.D.: JAX, M.D. A framework for differentiable physics*. Journal of Statistical Mechanics: Theory and Experiment 2021(12), 23 124016 (2021) https://doi.org/10.1088/1742-5468/ac3ae9

  45. [53]

    https://github.com/mariogeiger/ allegro-jax Accessed 2025-01-27

    Mario, G., Daigavane, A.: Allegro-jax. https://github.com/mariogeiger/ allegro-jax Accessed 2025-01-27

  46. [54]

    Nature Communications 12(1), 6884 (2021) https://doi.org/10.1038/s41467-021-27241-4

    Thaler, S., Zavadlav, J.: Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting. Nature Communications 12(1), 6884 (2021) https://doi.org/10.1038/s41467-021-27241-4

  47. [55]

    arXiv (2022)

    Gasteiger, J., Giri, S., Margraf, J.T., G¨ unnemann, S.: Fast and Uncertainty- Aware Directional Message Passing for Non-Equilibrium Molecules. arXiv (2022). https://doi.org/10.48550/arXiv.2011.14115

  48. [56]

    https://github.com/ACEsuit/mace-jax Accessed 2025-06-10

    Mario, G., Daigavane, A.: Mace-jax. https://github.com/ACEsuit/mace-jax Accessed 2025-06-10

  49. [57]

    Europhysics Letters (EPL) 26(8), 583–588 (1994) https://doi.org/10.1209/0295-5075/26/8/005

    Ercolessi, F., Adams, J.B.: Interatomic Potentials from First-Principles Calcula- tions: The Force-Matching Method. Europhysics Letters (EPL) 26(8), 583–588 (1994) https://doi.org/10.1209/0295-5075/26/8/005

  50. [58]

    arXiv (2017)

    Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. arXiv (2017). https://doi.org/10.48550/arXiv.1412.6980

  51. [59]

    Journal of Chemical Theory and Computation 8(10), 3637–3649 (2012) https://doi.org/10.1021/ct3000876 24

    Goga, N., Rzepiela, A.J., de Vries, A.H., Marrink, S.J., Berendsen, H.J.C.: Effi- cient Algorithms for Langevin and DPD Dynamics. Journal of Chemical Theory and Computation 8(10), 3637–3649 (2012) https://doi.org/10.1021/ct3000876 24

  52. [118]

    1002/chem.200800987

    Chemistry – A European Journal 15(1), 186–197 (2009) https://doi.org/10. 1002/chem.200800987

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.