acceptodds
Under review as a conference paper at ICLR 2027

NEMORA: Neural Equivariant Multipole Operators for Long-Range Atomistic Learning

Abstract

Equivariant graph neural networks have emerged as foundational architectures for machine-learned interatomic potentials, approaching quantum-chemical accuracy at a fraction of the computational cost. These models describe local atomic environments accurately, but finite spatial cutoffs truncate long-range information flow, and stacking message-passing layers can lead to over-smoothing and over-squashing. Existing long-range extensions either prescribe a fixed analytical propagation kernel, restrict long-range communication to scalars or degree-preserving channels, or incur super-linear computational cost. Combining learnable long-range equivariant transport with multiscale many-body expressivity and efficient scaling remains a central challenge. We introduce Neural Equivariant Multipole Operators (NEMORA), a neural equivariant extension of the Fast Multipole Method (FMM) for learning long-range tensorial representations. NEMORA generalizes the FMM's analytical multipole expansion and translation operators to learned equivariant counterparts on an adaptive spatial hierarchy. Its operators couple angular degrees and form many-body interactions across length scales, retaining the FMM's hierarchical organization and analytical radial factors as physical inductive biases while learning data-dependent long-range couplings. NEMORA is linear in time and memory, allowing it to treat larger systems than other long-range methods, and it augments both symmetry-constrained and unconstrained short-range backbones. On non-local benchmarks, it reduces energy and force errors relative to the short-range backbones by up to three orders of magnitude, and it is competitive with or better than existing long-range extensions in accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
equivariant graph neural networks, geometric deep learning, machine-learned interatomic potentials, long-range interactions, Fast Multipole Method, multiscale representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_25hJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed NEMORA, using learnable FMM to achieve equicvariant Long-Range Atomistic Learning. NEMORA allow different angular mixing during long-range propogating, and using density trick for multi-body interaction. Experiment results indicate that it increased the performance.

Critical strengths
  1. Provides extensive experiments that demonstrate the effectiveness of NEMORA across multiple benchmarks and different backbones.
  2. has low computational complexity and can scale to large scale systems.
Critical weaknesses
  1. FMM is already an effective method for long-range interactions. This paper lacks the abaltion with Fixed FMM (or, in other word, fixed NEMORA). Also, FMMCLASSIC outperforms the full model on some benchmarks, which makes this comparison important.
  2. Lack the accuracy performance evaluation on large scale system of NEMORA.
Questions for authors
  1. The authors provide many experiments and show good results. However, these experiments do not sufficiently demonstrate that the complex learnable mechanisms in NEMORA are necessary. Could the authors provide additional ablation experiments to clarify this (W1)?
  2. From my perspective, NEMORA shares some core ideas with existing hierarchical message passing methods. Could the authors compare NEMORA with these methods?
  3. How does the full NEMORA perform on large scale systems in terms of accuracy (W2)?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 25hJ

We thank the reviewer for three clear questions at the core of the method. New results: Appendix J. We uploaded the revised manuscript and, as supplementary material, a comparison with the submission. We also edited the text for clarity and brevity; the method and theory are unchanged, and all changes are marked in the comparison.

R1 – Ablation with a fixed FMM (W1, Q1). Yes: where the long-range physics is more than a sum of Coulomb terms, the learnable FMM is far more accurate than a fixed FMM trained on the same data. We ran the ablation you ask for, with every operator fixed to its classical value and the same backbone, data and schedule (Section 5.2, Appendix J.1).

Force RMSE (meV/Å), three seeds each:

Modelcumulene IDSODOODcarbon chainNaCl
MACE86.0102.171.915.043.9
MACE–LES (Coulomb)84.8102.272.420.27.2
Fixed FMM, scalar sources86.4102.372.419.87.3
Fixed FMM, quadrupolar sources86.5102.472.16.17.2
Learnable FMM (NEMORA)28.558.464.12.11.9
  • The fixed FMM matches the Coulomb baseline within 2%.
  • The learnable FMM is 3 to 10 times more accurate on in-distribution cumulene, the carbon chain and NaCl, and 43% and 11% ahead on unseen chain lengths.
  • On cumulene the energy depends on the relative twist of the end groups through the conjugated chain. No Coulomb kernel carries that at any multipole order.
  • FMMCLASSIC (fixed radial functions, learned sources and readout) sits between the two: it helps where electrostatics dominate (Au₂/MgO) and costs accuracy elsewhere (NaCl 2.78 against 1.89). Section 5 now states which member suits which physics.

R2 – Accuracy on large systems (W2, Q3). The accuracy holds. Models trained on periodic cells of at most 1,024 atoms are tested on unseen cells of up to 32,768 atoms, 32 times larger, and the error stays flat. The submission showed this to 4,096 atoms (Table 3 there); the revision extends it to 32,768 (Table 2, Appendix J.3).

Periodic ionic melt, force RMSE (meV/Å); cells above 1,024 atoms unseen:

Model2561,0242,0484,09616,38432,768
MACE168.6265.2329.8363.5489.8480.9
MACE–LES9.809.889.959.959.94out of memory
Fixed FMM, scalar sources9.649.699.789.799.799.83
Fixed FMM, dipolar sources8.418.468.538.548.548.59
  • From 256 to 32,768 atoms the fixed FMM moves from 8.41 to 8.59; MACE degrades from 169 to 481 and MACE–LES runs out of memory.
  • The long-range energy of this melt is the Coulomb sum itself, so fixed operators are the right model for it. The gain from learned operators is where the physics is not Coulombic (R1).
  • These cells are up to 88 times larger than the largest benchmark of Table 1 (370 atoms). The water model runs stably at 6,144 atoms.
  • On the 370-atom nanotube the learnable FMM lowers the force MAE of the local model by 29%.

R3 – Comparison with hierarchical message passing (Q2). NEMORA shares the hierarchy and differs in what the hierarchy carries. Hierarchical message passing (EGHN, MuSE) summarizes a cluster by a scalar-weighted average of its atoms, which discards how they are arranged about the centre. Multipole pooling keeps that arrangement, direction included, and the far field depends on it.

  • Example: two water molecules 10 Å apart, one rotated in place. The interaction energy swings from −6.7 to +8.6 meV. Scalar pooling sees a neutral cluster and returns zero at every orientation. Multipole pooling to ℓ ≤ 2 follows the swing to 0.8%.
  • Matched baseline (new, Appendix J.2): same backbone, tree, cluster pairs, schedule and seeds, with mean pooling and generic learned equivariant messages in place of the multipole transport.

Force RMSE (meV/Å), mean ± standard error of three seeds:

Modelcarbon chainNaClcumulene IDSODOOD
MACE–RANGE12.42.9395.6113.094.9
Generic hierarchy, same tree2.67 ± 0.382.93 ± 0.1534 ± 576 ± 4170 ± 20
NEMORA, same recipe2.07 ± 0.051.89 ± 0.0537 ± 669 ± 1964 ± 4
  • The hierarchy carries much of the gain, as you anticipated. Multipole transport adds 35% on NaCl and 22% on the carbon chain.
  • On cumulene the two are level in distribution; NEMORA is ahead on unseen lengths with a two to five times smaller spread.
  • MuSE (code not public; published numbers, gain over each method's own backbone, Appendix J.6): NEMORA's force gain is 26% against 8–9% on both DNA systems of MD22 and 29.4% against 22.5% on the nanotube; MuSE's is larger on the two smallest molecules.
  • NEMORA also counts every atom pair exactly once on a tree that deepens with system size, and reduces to the classical FMM when its operators take their classical values.

Thanks to the authors for the detailed response. My major concerns have been addressed, and I have raised my score accordingly.

One suggestion is to provide clearer guidance in the main text on how to choose between different NEMORA variants for different application scenarios. This would help users to better use it for their specific scenarios.

AuthorsOpenReview ↗

Thank you, and guidance on choosing NEMORA variants

Dear Reviewer 25hJ, thank you for your detailed review, your follow-up, and for raising your score. Your questions on the fixed-FMM ablation, accuracy on large systems, and comparison with hierarchical message passing led to new experiments that strengthened the paper. We agree with your suggestion: in the final version we will add explicit guidance in the main text on which NEMORA variant to use: the fixed FMM where the long-range energy is essentially Coulombic (e.g., ionic melts, water), and the learnable operators where it is not (e.g., cumulene, carbon chains). We would be glad to clarify anything further before the discussion closes on Nov 18. We greatly appreciate your support.

Reviewer_yyPEOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NEMORA extends local interatomic potentials to include information from distant atoms. It groups atoms at several spatial scales and learns how these groups exchange information, following the structure of the fast multipole method. This information is used to modify the local model’s atomic features and energy prediction. The paper also evaluates a variant that predicts atomic charges and, for water, dipoles from a local MACE model and calculates their electrostatic interactions using the analytical fast multipole method. The authors report energy and force errors on molecular and ionic systems, examine water simulations, and measure how runtime and memory use increase with system size.

Critical strengths

I identified the following strengths:

S1. The combination of equivariant features and an FMM-based hierarchy is a promising approach to nonlocal atomistic modeling. The results show meaningful improvements over local models on several datasets. The held-out cumulene lengths in Table 1 provide a useful test beyond fitting angular profiles, although MACE–MFN remains more accurate.

S2. The comparisons across four baseline models in Table 2 provide evidence for the usefulness of directional nonlocal information. The component ablations in Tables 9–10 are also useful, including the reported cases where simpler variants achieve lower errors.

S3. The authors check how accurately the analytical FMM variant calculates energies and forces for fixed charges. They also acknowledge that the learned operators do not retain the convergence guarantees of classical FMM.

Critical weaknesses

I think the manuscript contains some interesting ideas, but I find it difficult to understand what we learn from the presented results. Several parts of the method and the interpretation of the results need to be explained and tested more carefully. Detailed comments are provided below.

W1. The evidence for the impact of directional nonlocal information appears to be more convincing than the evidence for the particular FMM-based approach. As the authors acknowledge, learned FMM operators and multiscale equivariant models have already been explored, including https://arxiv.org/abs/2509.20591 and https://arxiv.org/abs/2605.25710. The proposed combination is interesting, but the paper should establish more clearly what it adds beyond these approaches.

W2. Several results are obtained without the proposed learned operators. In particular, the periodic-ion and water experiments use a more conventional approach. A local MACE model predicts atomic charges and, for water, dipoles, whose electrostatic interactions are then calculated using analytical FMM operations. This version removes the proposed learned operations and several other components, so its results do not establish their benefit. Furthermore, the largest reported improvement comes from the synthetic ionic benchmark, whose reference data include long-range Coulomb interactions. Adding these interactions substantially improves on the local model, but LES achieves similarly low errors. The authors should distinguish the benefit of including long-range electrostatics from the benefit of their proposed method.

W3. The cumulene results need a more careful interpretation. Reproducing the dependence on the relative orientation of the terminal groups and transferring the odd-even dependence to unseen chain lengths are different tasks. The former can be represented using relatively simple global equivariant features and tensor products. The latter additionally requires information about chain length and how it affects the coupling of the terminal groups. The paper does not explain how NEMORA captures this dependence or whether its hierarchy is necessary.

W4. It is unclear whether the learned model produces smooth energies and forces when the hierarchy changes. Section 4.2 uses discrete decisions to assign atoms to groups, split these groups, and distinguish near and far interactions. Differentiating the model for a fixed hierarchy does not show that energies and forces remain continuous when these decisions change. The authors should consider testing this directly. The water simulations do not address this point because the authors appear to use analytical FMM operations rather than the proposed learned operations, as discussed in W2.

W5. The runtime comparison does not establish an advantage over efficient particle-mesh methods. Section 5.6 uses direct Ewald implementations of LES and LOREM. However, the LOREM paper already reports PME timings in Appendix E of https://arxiv.org/html/2507.19382v2. An SPME implementation for an interatomic model is also available with https://doi.org/10.1021/acs.jctc.5c01828; see also https://github.com/nec-research/ictp. Finally, the difference between O(N)\mathcal{O}(N) and O(Nlog⁡N)\mathcal{O}(N\log N) does not determine which implementation is faster at practical system sizes.

W6. It is unclear whether the reported timings apply to the models used in the accuracy tests and water simulations. Figure 4 measures untrained models on argon configurations. These measurements show their computational cost, but do not connect that cost to the reported prediction errors. Appendix E also states that the implementation used for its reported accuracy calculates a smoothing correction over all atom pairs, so these tests do not establish linear scaling. The authors should report runtime and memory for the models they used in the main experiments, including all parts of the energy and force calculation.

W7. The water results do not show whether the improvement comes from FMM or from adding atomic dipoles. LES uses only charges, and the FMM model using only charges gives similar errors. The periodic FMM models also constrain the total charge to zero, while LES does not. The authors should consider comparing the methods using the same atomic charges or dipoles and the same charge constraints.

W8. The claims which involve over-smoothing and over-squashing are not supported by the experiments. A local model may fail simply because the relevant atoms lie outside its receptive field. Showing that a nonlocal model improves the errors does not establish that over-smoothing or over-squashing makes the baseline fail.

W9. The manuscript does not provide enough detail to reproduce the method and experiments. For example, some operations listed in Table 26 are not fully explained, and no code was provided for review.

Questions for authors

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

6 comments
AuthorsOpenReview ↗

Response to Reviewer yyPE Part 1

We thank the reviewer for a detailed and constructive report. The new results are in Appendix J of the revision. Part 2 covers R5–R9. We uploaded the revised manuscript and, as supplementary material, a comparison with the submission. We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

R1 – Contribution beyond learned FMM and multiscale models (W1). NEMORA rests on one observation: the FMM's own translations are Clebsch–Gordan tensor products weighted by a radial function, the form of an equivariant message. Learning the radial functions and couplings therefore keeps equivariance, conservative forces and the classical FMM as a limit.

  • Fognini et al. replace each translation by an unconstrained MLP on a fixed 2D grid, where nothing rotates and the output is not differentiated. That does not give invariant energies or conservative forces for atoms.
  • MuSE pools atoms with scalar distance weights at three fixed scales. It has no multipole expansion, and its range ends at 11.25 Å.
  • New matched test (Appendix J.2): a generic equivariant hierarchy with the same backbone, tree, cluster pairs, schedule and seeds, differing only in the transport.

Force RMSE (meV/Å), three seeds:

Modelcarbon chainNaClcumulene IDSODOOD
MACE15.043.986.0102.171.9
Fixed FMM19.87.386.4102.372.4
Generic hierarchy, same tree2.672.9334.375.869.6
NEMORA, same recipe2.071.8937.269.263.8

Multipole transport lowers the error by 35% on NaCl and 22% on the carbon chain. On cumulene the two are level in distribution and NEMORA is ahead on unseen lengths, with a two to five times smaller spread across seeds.

R2 – Results without the learned operators (W2). We agree, and the revision separates the two benefits. Long-range electrostatics is what the fixed FMM and LES both show, and they match. The benefit of learned operators is shown by training both on the same data (Appendix J.1).

Force RMSE (meV/Å), three seeds:

Modelcumulene IDcarbon chainNaCl
MACE86.015.043.9
MACE–LES84.820.27.2
Fixed FMM, scalar sources86.419.87.3
Fixed FMM, quadrupolar sources86.56.17.2
Learned NEMORA28.52.11.9
  • Learned NEMORA is 3 to 10 times more accurate than both, and 43% and 11% ahead on unseen cumulene lengths.
  • Quadrupolar sources help only on the carbon chain, so the gap is the learned transport, not the multipole order.
  • MD22, same comparison (one seed, force MAE): the learned model is 40% ahead of the fixed FMM on the nanotube (12.44 against 20.62 meV/Å) and 9% on stachyose, and comparable on the smaller molecules.
  • The periodic melt and water results use the fixed FMM, because their long-range energy is essentially the Coulomb sum. Every result now names its member of the family, and the three orders of magnitude on the melt are attributed to electrostatics, as you say. For learned NEMORA the largest gains over MACE are 25× on NaCl and 21× on Au₂/MgO.

R3 – Interpretation of cumulene (W3). Your distinction is what the data show, and the hierarchy carries the second task. The revised Section 5.1 separates the two.

Cumulene force RMSE (meV/Å):

ModelIDSOD (unseen lengths)OOD (unseen lengths)
MACE86.0102.171.9
LOREM (global features)29.4166.7133.1
EFA (global features)30.3158.293.9
Fixed FMM86.4102.372.4
NEMORA without hierarchy77.2151.2121.5
NEMORA28.558.464.1
  • Global tensorial features solve the orientation task and fail on unseen lengths. So does NEMORA without its hierarchy.
  • Mechanism: the end groups interact through far-field transfers whose separation and number of tree levels grow with the chain. Length enters through the radial factor and its learned modulation; the Clebsch–Gordan paths carry the orientations.
  • The operators must be learned: the fixed FMM on the same tree stays at the MACE error, because the coupling is not electrostatic.

R4 – Smoothness when the hierarchy changes (W4). Energies and forces are smooth. Membership, splitting and the near/far criterion are C² partitions of unity, as in regularized FMM (Shamshirgar et al., 2019), and forces are the exact gradient including the tree geometry (Section 4.5). We ran your direct test on the learned model with the tree rebuilt every step (Appendix J.4).

NVE energy drift (meV/atom/ps), 0.25 fs step:

SystemNEMORAMACE
Cumulene C₉, 300 K, 10 ps≤ 0.006≤ 0.006
Water, 384 atoms, 2 ps≤ 0.00080.0011
  • On cumulene the partition changes about 700 times in 10 ps with no jump in energy.
  • On DHA, moving one atom out of the support of a cluster, the force matches the finite difference of the energy to a relative 5 × 10⁻⁹, with second-order convergence.
AuthorsOpenReview ↗

Response to Reviewer yyPE Part 2

We continue the response below.

R5 – Runtime against particle-mesh methods (W5). We agree: at practical sizes the implementation decides the speed, not O(N) against O(N log N). Section 5.6 is corrected and no longer rests on direct Ewald. Timed against PME on the same GPU, NEMORA costs the same or less (Appendix J.7).

Energy and forces (ms), one B200, periodic argon:

Model1,0244,09616,38465,536 atoms
NEMORA, device time9.718.961.6287
MACE–LES with PME, device time13.431.6105409
NEMORA, complete step15.668.6173620
MACE–LES with PME, complete step17.3128205724
LOREM with PME (their Appendix E, H100, NaCl)–12.141.5 (17,576)101 (32,768)
  • PME is torch-pme, tuned to a relative force error of 10⁻³. The complete step adds the neighbour list, shared by both models, and the hierarchy, which is built on the GPU and takes 0.5–4% of the step. One run; the neighbour list dominates the spread.
  • LOREM is on other hardware and another system, so we read it per atom (about 3 μs against about 4 μs for NEMORA), not as a ranking.
  • The same code runs non-periodic systems, where particle-mesh methods do not apply.
  • What costs more is the equivariant content: 32 channels of degrees 0 to 2 where a mesh carries one scalar. This is a first implementation, and future work will accelerate it.

R6 – Timings for the models of the main experiments (W6). The cost of the architecture does not depend on the weights. To make the link direct we now time the trained models, and we restricted the smoothing correction to near-field pairs, which makes the fixed FMM linear end to end (Appendix J.7).

Trained MD22 models (42 and 60 atoms), one B200:

QuantityNEMORAMACE
Energy and forces (ms)9.6–10.72.3–2.4
Peak device memory (MB)261260

Fixed FMM, trained models, force RMSE (meV/Å):

Systemcorrection over all pairsnear field only
Water (768 atoms)11.00011.001
Ionic melt (4,096 atoms)8.5398.539
Ionic cluster (1,024 atoms)16.59716.597
  • On small molecules the added 7–8 ms is a fixed cost of the hierarchy, not a cost per atom. It is not yet optimized.
  • The correction is an erfc term below 10⁻¹¹ beyond 7 Å, so restricting it changes nothing.

R7 – Water: FMM or dipoles (W7). You are right: the gain on water comes from the dipoles. That is the point of the method. The FMM carries dipoles and higher multipoles through the same operators at linear cost; LES sums scalar charges.

  • Same scalar charges, water with 384 atoms: LES 15.75, fixed FMM 15.88 meV/Å.
  • FMM against Ewald at identical fixed charges and dipoles (water, 4,095 atoms): force difference 4.07, 0.19, 0.03 meV/Å at p = 3, 6, 8.
  • Neutrality: LES omits the k = 0 term, which is a neutralizing background, and periodic NEMORA removes the net monopole in the same way. Switching the projection off changes the melt force error by less than 10⁻³ meV/Å.

Section 5.5 now attributes the gain to the dipoles.

R8 – Over-smoothing and over-squashing (W8). Agreed; the claim is narrowed. Both terms are removed from the abstract and remain in the introduction only as cited background that we do not test. The paper now claims what the experiments show: a finite receptive field limits local models, and NEMORA removes that limit.

R9 – Reproducibility (W9). Agreed that some operations of Table 26 were named without being defined. Appendix L.5 now gives one equation per setting: the gated refinement step, cumulant modulation, density-mean aggregation, the self-Kelvin scalar and the channel gates. The "Neumann iteration" was a single step (two on NaCl) and is renamed. The configuration of every benchmark is in the appendix. The code and the periodic water dataset will be released with the published paper.

AuthorsOpenReview ↗

Follow-up before the discussion deadline

Dear Reviewer yyPE, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns, in particular the fixed- versus learned-operator comparison, the matched hierarchical baseline, the smoothness tests with the tree rebuilt every step, the PME timings, and the attribution of the water gains to dipoles (Appendices J.1, J.2, J.4, J.7). We would be glad to clarify anything further before then. Thank you again for your detailed review.

Reviewer_KujqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a neural version of the fast multipole method. More precisely, it parameterizes the tensor charge in the Laplace expansion as equivariant operators and computes them under top-down adaptive subdivision.

Critical strengths
  1. The idea of neural FMM is clearly novel and the writing is clean, as all components are complete like Laplace expansion and computing clusters (or subdivisions)
  2. Results on MD22 show the effectiveness of this method. And Figure 4 justifies the O(N)O(N) time complexity claim of this method.
Critical weaknesses

I do not see major weaknesses in this paper. There are some subtleties I would like to mention

  1. Line 207: I don't see it as necessary to put the Ewald sum related works here
  2. Section 4.2: How does the method handle the near-degeneracy of PCA?
  3. I would like to see the subdivision method over the Laplace expansion (eq. 3) without neural parameterization. It would be clearer for me to see if such a subdivision method is effective
  4. For comparison over crystalline materials, I recommend that the authors compare with neural parameterizations with Ewald summation, e.g., Ewald-MP [1]
  5. Cite and discuss [2] [3] [4]

[1] Kosmala, Arthur, et al. "Ewald-based long-range message passing for molecular graphs." International Conference on Machine Learning. PMLR, 2023.

[2] Taniai, Tatsunori, et al. "Crystalformer: Infinitely connected attention for periodic structure encoding." International Conference on Learning Representations. Vol. 2024. 2024.

[3] Lin, Yuchao, et al. "Efficient approximations of complete interatomic potentials for crystal property prediction." International conference on machine learning. PMLR, 2023.

[4] Nie, Jianan, et al. "ReciNet: Reciprocal space-aware long-range modeling for crystalline property prediction." arXiv preprint arXiv:2502.02748 (2025).

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Kujq

We thank the reviewer for the careful review and for recognizing the novelty of the neural equivariant FMM. The new results are in Appendix J of the revision. We uploaded the revised manuscript and, as supplementary material, a comparison with the submission. We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

R1 – Ewald related work at line 207. Agreed. We removed the passage from Section 4.1; Ewald-based models are discussed in Related Work.

R2 – Near-degeneracy of PCA. Near-degenerate axes do not change the predictions on any benchmark structure. The operators are exactly equivariant for any assignment of atoms to clusters, so degeneracy could act only through the assignment, and a rotation-invariant rule fixes it (Section 4.5, Appendix J.5).

  • Rule: split axes are chosen by the invariant ratios λ_k/λ_max, and each axis is oriented by the sign of the third moment of the cluster (fifth as fallback). Periodic systems are split in the lattice frame.
  • Rotation audit: 51,440 rotated copies of cumulene, carbon chain, NaCl, Au₂/MgO, AT-AT and the buckyball catcher. The partition is unchanged in all of them. Energies and forces change by less than 10⁻⁵ of the test error in double precision and at most 2% in single precision.
  • Crossings: torsion scans of eight cumulene chains through the crossing of the two transverse eigenvalues at 90°, resolved to 10⁻⁴ degrees. No energy step above 10⁻⁶ meV; forces match finite differences to 0.002 meV/Å.
  • Dynamics: in NVE runs of Ac-Ala3-NHMe and the double-walled nanotube, where 84% of test frames have two eigenvalues within 1%, the total energy stays within 2.3 meV.

Only at an exactly degenerate covariance, a set of measure zero, are the principal axes not unique, as for any construction based on principal axes. A displacement of 10⁻³ Å removes it, and the appendix states this.

R3 – Subdivision without neural parameterization. Yes, the subdivision is effective on its own. With analytical operators it converges to the Ewald and direct sums, and trained with learned charges it matches the Ewald-based baseline. This variant produced the periodic results of the submission; the revision presents it in Section 4.1 and adds the tests below (Appendix J.1).

Force error against Ewald or direct sum (meV/Å), fixed sources, p = 2 / 4 / 6:

Systemadaptive treeuniform octree
Water (4,095 atoms)12.5 / 1.14 / 0.1412.5 / 0.96 / 0.08
Ionic melt (16,384 atoms)14.4 / 1.07 / 0.1010.0 / 0.69 / 0.05
Ionic cluster (4,096 atoms)8.2 / 0.63 / 0.068.6 / 0.58 / 0.05
  • The error falls about tenfold for every two orders in p, from 1,024 to 65,536 atoms.
  • At fixed sources a uniform octree is as accurate, as it should be for an exact expansion. The adaptive tree is faster: 178 against 231 ms on 16,383 water atoms at p = 4.
  • With learned operators the adaptive tree also matters for accuracy. Force RMSE, adaptive against octree: carbon chain 2.07 against 3.21, NaCl 1.89 against 3.56, cumulene 57.2 against 88.4 meV/Å.
  • Trained with learned charges, the fixed FMM is within 2% of MACE–LES on the periodic melt and keeps its error from 256 to 4,096 atoms.

R4 – Ewald-based neural parameterizations. We added Ewald-MP on the same MACE backbone as NEMORA, and NEMORA is more accurate on every benchmark tested (Appendix J.8). The submission already compares with two Ewald-based models, MACE–LES and LOREM, including on periodic Au₂/MgO.

Force RMSE (meV/Å):

ModelNaClcarbon chaincumulene IDSODOOD
MACE43.915.086.0102.171.9
Ewald-MP (MACE backbone)8.02.885.8105.578.2
NEMORA1.892.0728.558.464.1
  • Ewald-MP uses the filter design of Kosmala et al. (2023) and the schedule of the MACE models; one seed on NaCl and the carbon chain (validation frames), three seeds on cumulene (standard deviation 0.6–2.7). On cumulene it uses the settings of the authors' molecular configuration.
  • Ewald-MP helps where the long-range physics is electrostatic, and NEMORA is still more accurate there.
  • On cumulene, whose signal is the relative orientation of the end groups, Ewald-MP is level with MACE and does not transfer to longer chains. NEMORA improves on every split.
  • NEMORA also applies unchanged to non-periodic systems, where reciprocal-space methods do not.

R5 – Additional references. Thank you; all three are now cited and discussed in Related Work. Crystalformer evaluates attention over the infinite lattice, Lin et al. approximate complete lattice sums of interatomic potentials, and ReciNet adds reciprocal-space features to local message passing. They target crystal property prediction. NEMORA learns energies and forces for periodic and non-periodic systems with a real-space hierarchy at linear cost, and is complementary. Crystal property prediction with NEMORA is a natural next step.

Reviewer_KujqOpenReview ↗

Response to Authors

I believe my questions are well addressed. I think the score of this paper would be between 3 and 4, but there is no such option for me to do so. I will keep my positive score.

AuthorsOpenReview ↗

Thank you for your review and support

Dear Reviewer Kujq, thank you for your careful review, for your follow-up, and for indicating that you see the paper between a 3 and a 4. Your questions on PCA near-degeneracy, the subdivision without neural parameterization, and the Ewald-MP comparison led to new experiments that strengthened the paper. We would be glad to clarify anything further before the discussion closes on Nov 18. We greatly appreciate your continued support.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Area Chair and All Reviewer

We thank the Area Chair and the three reviewers for their careful and constructive reports. We believe these reports and the subsequent experiments and edits we made have improved the paper. We uploaded the revised manuscript and, as supplementary material, a comparison with the submission. The new experiments the reviewers asked for are in Appendix J. We also edited the text throughout for clarity and brevity; the method and theory are unchanged, and all text changes are marked in the supplementary comparison.

The reviews converge on five questions. Short answers are below; numbers and tables are in the individual replies, with points numbered R1, R2, … for each reviewer.

1. Are the learned operators necessary, or does a fixed FMM suffice? (25hJ R1, yyPE R2, Kujq R3)

  • We trained NEMORA with every operator fixed to its analytical value on the same benchmarks as the learned model (Appendix J.1).
  • The fixed FMM matches MACE–LES, as an exact Coulomb evaluator should.
  • Learned NEMORA is 3 to 10 times more accurate on in-distribution cumulene, the carbon chain and NaCl, and 43% and 11% ahead on unseen cumulene lengths.
  • The analytical variant produced the periodic results of the submission; the revision names it throughout.

2. What does the FMM add beyond hierarchical message passing? (25hJ R3, yyPE R1)

  • Directional information. A generic hierarchy pools a cluster to a weighted average; a multipole expansion keeps how the atoms are arranged about the centre, and the FMM transports it.
  • Against a hierarchical message-passing baseline matched in backbone, tree, cluster pairs, schedule and seeds, multipole transport lowers the force error by 35% on NaCl and 22% on the carbon chain (Appendix J.2).

3. Are energies and forces smooth when the hierarchy changes? (yyPE R4, Kujq R2)

  • Cluster membership, splitting and the near/far criterion are smooth partitions of unity, and the forces are the exact gradient of the energy.
  • With the tree rebuilt at every step, the partition changes about 700 times in 10 ps on cumulene with no jump in energy (Appendix J.4).
  • The partition is unchanged in all 51,440 rotated copies of the benchmark structures (Appendix J.5).

4. How does the cost compare with particle-mesh methods? (yyPE R5, R6)

  • We agree that at practical sizes the implementation decides the speed. Section 5.6 is corrected.
  • Timed against MACE–LES with PME on the same GPU, a complete NEMORA step, including neighbor-list and hierarchy construction, takes 620 ms against 724 ms at 65,536 atoms (Appendix J.7).
  • Future work will accelerate the method further.

5. Does the accuracy hold on large systems? (25hJ R2)

  • Yes. On the periodic ionic melt, models trained on cells of at most 1,024 atoms keep a flat error on unseen cells of up to 32,768 atoms, 32 times larger, where the Ewald-based baseline runs out of memory (Section 5.5, Appendix J.3).

We also followed the reviewers where they asked for more precision:

  • The water improvement is attributed to the atomic dipoles, which the FMM carries at linear cost (yyPE R7).
  • The abstract no longer invokes over-smoothing or over-squashing (yyPE R8).
  • The cumulene discussion separates orientation dependence from length transfer (yyPE R3).
  • Ewald-MP is added as a baseline on the same backbone, and the references proposed by Reviewer Kujq are cited and discussed (Kujq R1, R4, R5).

We are glad to run further experiments or clarify any point during the discussion.

AuthorsOpenReview ↗

Final revision and thank you

We thank the Area Chair and all three reviewers for their careful reading and constructive discussion. We have uploaded a final revision addressing the remaining suggestions: guidance on choosing analytical or learnable operators (Section 4.3), discussion of computational cost across system sizes and the different MACE implementations used in the PME comparison (Section 5.6), and clarification of GPU hierarchy construction and complete-step timings (Appendix J.7). These final edits clarify the presentation without changing the method or experimental results. We appreciate the reviewers’ follow-up assessments and the time they devoted to helping us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.