acceptodds
Under review as a conference paper at ICLR 2027

NoRRA: Nonlocal Representation and Response Augmentation for Machine Learning Interatomic Models

Abstract

Machine-learning interatomic potentials (MLIPs) achieve efficient atomistic simulation by constructing each atomic representation from neighbors within a finite cutoff, but this locality can miss physical responses that depend on structure beyond the cutoff. Two complementary strategies address this limitation: nonlocal representation, which supplies local MLIPs with global structural context, and nonlocal energy response, which constructs an explicit nonlocal interaction energy from latent variables such as charges or dipoles. We introduce NoRRA (Nonlocal Representation and Response Augmentation), a plug-in framework that couples both pathways in a local MLIP. Reciprocal-space features inject global periodic context before a replaceable backbone, while a separate nonlocal-response readout predicts constrained latent charges and evaluates an explicit interaction energy through a learnable sum-of-Gaussians kernel. On the same Point Edge Transformer (PET) backbone, NoRRA adds only parameters and reduces SPICE relative-energy MAE from the baseline value of to  kJ/mol. This framework also transfers to MatterSim. Across molecular, ionic, interfacial, and electrochemical systems, NoRRA improves energy and force accuracy and downstream physical properties, including molten-salt thermal expansion, the liquid structure of NaCl, and the long-range force tails of ion pairs. Component ablations show that the two pathways are complementary, with the largest gains in ionic regimes where Coulomb interactions dominate.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning interatomic potentials, long-range interactions, nonlocal representation, nonlocal response, electrostatics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_pZWhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents NoRRA, a framework for modeling long-range interactions in machine-learned interatomic potentials. It uses species-weighted structure factors to augment the atom and pair features of a local model with information about the atomic arrangement in the full structure. The model also predicts latent atomic charges, constrained to sum to the specified total charge, and uses them to compute an additional interaction energy with a learnable sum-of-Gaussians kernel. The authors evaluate implementations based on PET and MatterSim architectures on various molecular and materials datasets. They assess energy and force errors and report additional tests of molten-salt structure and thermal expansion, together with the longitudinal–transverse optical phonon splitting in NaCl.

Critical strengths

The paper addresses the challenge of capturing long-range effects in local interatomic potential models and evaluates the proposed approach across several datasets and physical observables. The ablation study (Section 4.5) examines the contributions of the non-local input features and the interaction energy modeled by the sum-of-Gaussians kernel. Their combination gives the lowest reported energy and force errors on the three systems included in this comparison (Table 4).

Critical weaknesses

I think the manuscript contains some promising ideas, but they need to be developed further and tested more carefully. Despite the many calculations, it remains unclear what the method adds and whether the results support the claims. I believe substantial revision is needed before the paper is suitable for publication.

Detailed comments:

W1. I find it difficult to understand what the authors consider the main contribution. As the authors acknowledge, nonlocal features, interaction energies based on latent charges, and learnable sum-of-Gaussians kernels have already been explored in previous work, including https://doi.org/10.1063/1.5128375, https://proceedings.mlr.press/v202/kosmala23a.html, and https://journals.aps.org/prl/abstract/10.1103/ssp9-7s81. NoRRA passes nonlocal atom and pair embeddings to an existing backbone and predicts an additional interaction energy from latent atomic charges. It remains unclear what is new in the embeddings, their use by the backbone, or their combination with the interaction energy. The paper should explain whether its contribution is a new method, better results from combining existing methods, or a better understanding of when this combination is useful.

W2. The ablation in Section 4.5 is useful, but I find it difficult to relate its results to the rest of the paper. It shows that combining the two components improves energy and force errors on three systems. However, it does not show whether the same conclusion holds for the main benchmarks or the models used to calculate physical properties. The phonon calculation appears to use a separate PET model with the SOG interaction-energy component but without the nonlocal input features, so it does not test the proposed combination. The authors should explain which conclusions from the ablation apply to the other results and test the individual components under the conditions used for those calculations.

W3. The SPICE results are also difficult to interpret because the training settings differ. NoRRA uses an energy-to-force loss weight of 0.25:1, while the PET and PET-LES baseline models use 0.01 or 0.05. Giving more weight to energies could itself reduce the energy error. The authors should explain why these weights differ and provide comparisons with the same training settings. I could not find the corresponding SPICE force errors. The reported force-sign accuracies do not show whether the lower energy errors come at the expense of force accuracy.

W4. I would also like to see how the PET baseline compares with the models reported in https://arxiv.org/abs/2601.16195v3. The published numbers use different test sets and error measures, so they cannot be compared directly. Evaluating these PET models on the same test set would help establish whether the reported gains depend on the particular baseline used here. Differences in model size, training time, and computational cost should be taken into account too.

W5. I do not find the argument in lines 57–65 convincing. Writing the total energy as a sum of atomic contributions does not prevent correct long-range behavior when those contributions depend on nonlocal features. LODE and related approaches explicitly connect such features to electrostatic asymptotics (https://doi.org/10.1039/d0sc04934d). The authors should explain which limitation of these approaches NoRRA resolves, especially since its learned SOG kernel does not itself guarantee the Coulomb asymptote.

W6. The authors use a fixed reciprocal-space cutoff, but it is unclear whether the sums are sufficiently converged for different simulation cells. A cutoff that works for one cell may not be sufficient for a larger cell. Related concerns about the dependence on summation parameters and transfer between systems are discussed in https://openreview.net/pdf?id=ssfrYXdQ4x. The authors should show how the predicted energies, forces, and stresses depend on the cutoff and clarify whether it controls numerical accuracy or is part of the learned model. They should also test whether repeating the same periodic structure in a larger cell preserves the energy per atom, forces, and stress.

W7. The SPICE models also use different charge constraints. Table 6 specifies charges that sum to the total system charge for NoRRA, while the LES charges are constrained to sum to zero. The authors should explain this difference and its effect on the comparison. PET-QEq uses a hard 12 Å interaction cutoff and is fine-tuned for 30 epochs from an epoch-150 checkpoint, whereas the other SPICE models are trained for 207 epochs (Appendices C.4 and D.1). These differences should be justified. The QEq results also need to be interpreted with this cutoff in mind, since they do not establish how NoRRA compares with QEq models that include interactions beyond 12 Å.

W8. Several experiments do not clearly support the authors’ explanations. The Au/MgO cutoff test uses an adaptive neighborhood targeting eight neighbors, so increasing the cutoff does not necessarily increase the structural information available to the model. The ion-pair test checks only force signs. The authors could also compare force magnitudes and their dependence on distance, which does not require isolated-ion reference energies. Finally, the OMat results use different datasets and training stages of 1, 6, and 300 epochs. It is therefore unclear whether the larger improvements on ionic systems come from their stronger long-range interactions or from differences in the data and training.

W9. The presentation of the MatterSim results is misleading. The main table includes only cases where both energy and force errors improve, while Appendix F reports worse energy errors elsewhere and different training schedules for Au/MgO. It is also unclear whether MatterSim uses both NoRRA components. These results do not establish that the reported benefits hold across backbones.

W10. The paper should state how forces and stresses are calculated for each experiment. The RDF simulations use nonconservative forces (Appendix E.3), and Appendix B.5 allows forces calculated with the predicted charges held fixed. Since these charges depend on atomic positions, this dependence must be accounted for when differentiating the energy. The authors should also check whether the simulation results change with the way they compute forces.

W11. Appendix B.5 describes direct summation, PME, and NUFFT as alternative ways of calculating the interaction energy. The authors should show whether they give the same energies and forces for the same structures. Any resulting differences, and which implementation is used for each experiment, should be explained.

W12. The supplement contains saved results and plotting scripts, but no NoRRA implementation, trained models, or complete settings to reproduce the experiments.

Questions for authors

My questions follow directly from the weaknesses presented above and would be answered by addressing them.

Reviewer_e3vGOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper describes NoRRA, a method augmenting a local MLIP with two nonlocal pathways. The first is a pathway that computes species-weighted structure factors over a finite set of reciprocal modes, filters them with a Ewald-like kernel, and fuses the results into the initial node and edge embeddings of a replaceable backbone. The second is a pathway that routes a slice of backbone output through a charge predictor with constraints, then evaluates an explicit reciprocal-space interaction energy with a learnable sum-of-Gaussians kernel.

Critical strengths

The representation-vs-response taxonomy is a clean way to organize prior work, and coupling both in one plug-in is a natural, previously untested design. The breadth of evaluation is unusual: molecules, inorganic crystals, oxide catalysts, and finite-temperature observables rather than only supervised MAE. Inference overhead is reported honestly and scales favorably with system size.

Critical weaknesses

The SPICE comparison is not matched. Sec 4 states that for each comparison "the only architectural difference is whether NoRRA is enabled". However, Appendix C.1/D.1 states that PET-NoRRA trains different λE:λF=0.25:1\lambda_E:\lambda_F = 0.25:1 while the PET baseline and PET-LES rows keep λE=0.01\lambda_E=0.01 or 0.05, and the reported metric is an energy error. The PET-QEq row uses yet another protocol (30-epoch fine-tune from epoch 150 at 0.25:1). 2. There is no comparison with existing nonlocal methods on the systems where it matters. Au/MgO and Na8/9CI8+Na_8/_9CI_8^{+} come from the 4G-HDNNP benchmark. Water/Pt, Deep ⋅\cdot water1900 and Electrolyte-B come from SOG-Net. LES-BEC water from LES. Published numbers exist for all of them. A local baseline is known to fail on these systems, so beating it is expected. The open question is whether NoRRA beats 4G-HDNNP, SOG-Net, LES, or the Kim et al. (2025) wrapper, and the paper never asks it outside SPICE (where the comparison is confounded). 3. There are no seeds/multiple trials repeated evidence. It seems like all numbers reported are a single run with best-validation checkpoint selection.

Questions for authors

See above weaknesses.

Reviewer_XtyLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose Nonlocal Representation and Response Augmentation (NoRRA), which adds long-range physics to an interatomic potential in two ways: Fourier features of the whole cell as input, and an explicit energy built from predicted charges and a learnable Fourier-space kernel. The signal-processing side feels familiar, since it consists of non-uniform FFTs and a convolution kernel.

Critical strengths

The idea of the paper is sensible and the paper is very readable especially compared to the other papers in my stack.

Critical weaknesses

1- The main issue with this article is that NoRRA was trained with a different loss from the models it is compared against, and the difference favours the metric being reported.

Specifically, these models are trained with a weighted sum of two errors, λELE+λFLF\lambda_E \mathcal{L}_E + \lambda_F \mathcal{L}_F, one on energies and one on forces (Eq. 28). Appendices C.1, C.2 and D.1 state:

  • PET-NoRRA: energy:force weight 0.25:1.
  • PET baseline and PET-LES: they “keep their prescribed energy weights (0.01 or 0.05)”. The two LES rows in Table 1 are labelled 0.01:1 and 0.05:1.

This matters, because NoRRA puts 5 to 25 times more weight on energy. The SPICE metric in the abstract, 19.798 → 14.297 kJ/mol, is an energy error: relative energies between conformers of the same molecule.

As with the energy–force trade-off in the first paper, the weight moves a model along a trade-off curve: more weight on energy usually means better energies and somewhat worse forces. Part of NoRRA’s energy gain may therefore come from the loss weight rather than the architecture, and the paper gives no way to tell how much.

The wording obscures this. Section 4 says the baseline and NoRRA share “the same data split, preprocessing, backbone capacity, training budget, and checkpoint-selection rule; the only architectural difference is whether NoRRA is enabled”. That’s literally true, but the loss isn’t architecture, and it’s missing from every list of “matched” settings (Table 6, App. C.2).

A few additional worries

a- Som fair runs do exist but are not shown. Appendix C.1 says NoRRA was also trained at 0.05:1 and 0.1:1, kept only as “archived weight-sweep diagnostics”. At 0.05:1 it would compare directly with PET-LES at 0.05:1 (16.960). Why are the authors not showing these results ?

b- One comparison is matched on weight. PET-QEq uses 0.25:1 too (App. D.1), and NoRRA still beats it, 14.297 vs 18.270. But QEq was trained differently (30 extra epochs continuing from epoch 150), so even that comparison is not clean.

c- The confound may be smaller than it looks. Going from 0.01 to 0.05 worsened PET-LES slightly (16.606 → 16.960), so a higher weight doesn’t automatically help. But nobody tried the baselines at 0.25. Also, no SPICE force errors are reported, so we cannot even check if NoRRA paid for its energies with worse forces.

Some other major problems

2- The MatterSim table is biased: by the authors’ own statement, it lists only systems where both energy and force improve. Electrolyte-B is reported on a subset, while on the full set its energy gets 9.2% worse.

3- There are no multiple runs: Tables 1–4 have no error bars, and the dedicated systems report validation error at the best validation checkpoint.

  • Every number in Tables 1–4 is a single value, with no ± and no mention of averaging over runs.
  • The checkpoint rules speak of “each run’s lowest validation-energy checkpoint” (App. C.4), in the singular.
  • The only replication is the RDF: four MD simulations from different starting frames (Fig. 5), all run with the same trained model. That measures simulation noise, not training noise.

It matters a lot because energy errors for example can sometimes vary 10-50% across runs.

Likely robust (large gains)Could be noise (small gains)
Au/MgO: energy −64%, force −81%Full OMat24: energy −2.1%, force −1.4% (validation)
Na$_{8/9}ClCl_8^+$: energy −49%, force −79%OC22 force: −5% (ID), −2% (OOD)
Water/Pt force: −9.6%
Electrolyte-B energy: −10%

4- The physics explanations conflict with their own data:

  • Melt volumes: they argue NoRRA restores cohesion, yet its melts are 7–11% less dense than the baseline.
  • LO–TO match: the near-perfect result is mostly built in. The experimental inputs alone give 38.1 THz² for LO²−TO², against 37.7 from their values.
  • Cutoff scan: PET’s adaptive cutoff means the scan doesn’t really enlarge the neighbourhood.

5- The Fourier features depend on how the cell is described: I checked that for perfect NaCl, all retained modes are zero from 216 atoms upwards, so the “global context” vanishes on large ideal cells.

Questions for authors

Fair comparison

  1. On SPICE, PET-NoRRA is trained with an energy:force weight of 0.25:1, while the PET baseline and PET-LES keep 0.01 or 0.05 (App. C.1, C.2, D.1), and the headline metric is an energy error. Could you report all arms at the same weight, in particular the archived NoRRA runs at 0.05:1 and 0.1:1, together with the SPICE force errors of every arm?
  2. Panel C of Table 3 lists only "the systems in which both energy and force improve" (App. F). Could you report every MatterSim system you trained, on the full data? On the full Electrolyte-B set, energy worsens by 9.2%, and LES-BEC water (+10.4% energy) is absent from the main text.
  3. Several comparisons differ in more than NoRRA: Au/MgO MatterSim-NoRRA uses a three-stage continuation while the baseline uses a single cosine run; the KF electrolyte compares a 1-block NoRRA with a 3-block baseline; best epochs differ between arms. Could you retrain these pairs under identical schedules and sizes?
  4. Long-range baselines (PET-LES, PET-QEq) appear only on SPICE. Could you compare with LES, SOG-Net, 4G-HDNNP or Ewald message passing on the periodic and dedicated systems, where your claims are strongest?
  5. OC22 results are reported at epoch 10. How do both arms compare at convergence, and against published OC22 models?

Statistics and selection

  1. Tables 1–4 report single values without any measure of variability. How many training runs underlie each entry? Could you give mean ± standard deviation over at least three seeds, so that the smaller gains (OC22 forces, Water/Pt forces, full OMat24) can be judged against run-to-run variation?
  2. For the dedicated systems, errors are reported on the validation set at the best validation checkpoint. Could you report an independent test set, and define the "energy-calibrated" checkpoint rule? Why do the Water/Pt baseline errors differ so much between Table 3 (6.55 / 42.0) and Table 4 (2.04 / 84.7), and why does Table 7 not match Table 4, as its caption says it should?
  3. The ionicity trend (Table 2A) compares domains trained for 1, 6 and 300 epochs, on 769k, 238k and 1.4k structures. Could you repeat it at equal budgets, with a quantitative measure of ionicity? How does it square with Electrolyte-B having the smallest energy gain in Table 3?

Mechanism

  1. Could you plot the learned kernel Kψ(k)K_\psi(k) against 4π/k24\pi/k^2, and ϕψ(r)\phi_\psi(r) against 1/r1/r, for each dataset? A finite sum of Gaussians has no 1/r1/r tail, so over what range is the learned response actually long-range?
  2. Could you add a control that supplies global information without electrostatics (mean-pooled features or random Fourier features of matched size), to test whether the gains are electrostatic? Is there an ablation on OC22?
  3. PET's adaptive cutoff targets eight neighbours (App. D.4), so the cutoff scan of Fig. 2 does not enlarge the neighbourhood. Could you repeat it with a fixed, full neighbour ball?
  4. Fig. 6 shows no reference curve: what is the DFT equilibrium Au–O distance in each case? Please also check the dopant label.
  5. Are the SPICE-1.1 ion pairs of Table 1B part of the training data? Could you report force magnitudes against DFT, and the sign accuracy for same-sign pairs separately (about 0.69 if all groups weigh equally)?
  6. In App. G.1, ∣ELR∣/∣Etotal∣=0.37%|E_{LR}|/|E_{total}| = 0.37\% may still exceed the errors being corrected. Could you compare ELRE_{LR} with the interaction energy or the error instead? In Fig. 3, how strong is the charge correlation within each element?

Downstream physics

  1. Table 8 shows NoRRA melts 7–11% less dense than the baseline, while the text argues NoRRA restores Coulomb cohesion. What are the reference volumes, can you give αV\alpha_V with uncertainties, and how does αV\alpha_V compare at the reference density?
  2. Eq. 38 with experimental Z∗Z^* and ϵ∞\epsilon_\infty largely fixes ωLO2−ωTO2\omega_{LO}^2 - \omega_{TO}^2 (about 38 THz²). Could you report TO and LO separately, for both models, from the same checkpoint as the other results, ideally with model-predicted Born charges?
  3. The MD uses non-conservative forces. Do the downstream results hold with conservative forces, and is energy conserved in NVE? Is the RDF difference (0.041 ± 0.010 vs 0.034 ± 0.013) significant?

Invariance and cost

  1. Reciprocal modes are truncated on integer indices (∥n∥≤4\|n\| \le 4), so kmaxk_{max} shrinks with cell size and depends on the lattice basis. For perfect rock-salt NaCl, all retained modes vanish from 216 atoms upwards. Is E(supercell) = N × E(cell)? Is the model invariant to the choice of lattice basis, and does it transfer across cell sizes?
  2. How is the cell chosen for gas-phase SPICE molecules, and is rotation invariance preserved for the new reciprocal channels?
  3. Table 4B shows +71% latency at 64 atoms, a typical training size. Could you report training time, memory and latency at training sizes, against other long-range methods, to support the "lightweight" claim?
  4. The supplement regenerates tables from frozen numerical exports. Will you release training code, configurations and checkpoints so that the results can be reproduced, not only re-tabulated?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.