acceptodds
Under review as a conference paper at ICLR 2027

Distance-Dependent Angular Resolution for Equivariant Interatomic Potentials

Abstract

Equivariant machine-learned interatomic potentials (MLIPs) typically use the same angular resolution on every edge, regardless of interatomic distance. We introduce radial gates with learnable cutoff radii that adapt angular resolution to interatomic distance within each message-passing block. We further introduce a band mixture-of-experts (Band-MoE) that uses these learned radial cutoffs to route edges deterministically to parameter-sharing experts with different maximum retained degrees. Separately, we also show that a reflection-symmetric simplification of the edge-frame operator can lead to 25.9% parameter reduction. We reach batched inference speedup with 50% training memory reduction, while maintaining accuracy close to the noise thresholds for forces and energy at maximum angular degree of 4. We show that at higher angular degree (), our method outperforms the baseline model by with an inference speedup of . We evaluate the method on non-periodic molecules from OMol25 and periodic inorganic crystals from OMat24. The Band-MoE and the learnable radial cutoff methodologies are applicable to other MLIPs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational efficiency, equivariant neural networks, machine learning interatomic potentials

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ntwTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAL, which uses learnable radial gating to adaptively adjust the retained spherical harmonic orders based on interatomic distances. It further employs a parameter-sharing Band-MoE with different maximum angular orders to reduce edge-wise computation in equivariant machine learning interatomic potentials (MLIPs). The authors also introduce an independent reflection symmetry constraint to reduce the number of model parameters. The paper evaluates accuracy and efficiency on OMol25 and OMat24. On OMol25, under the Lmax⁡=4L_{\max}=4 direct force prediction setting, the combined approach achieves 2.10× higher batched inference throughput than the dense baseline while maintaining comparable accuracy.

Critical strengths
  1. Figure 1 is intuitive and clearly illustrates the proposed method, making it easy to understand.

  2. Adopting different angular degrees for different cutoff regions is a meaningful research.

Critical weaknesses
  1. I am not fully convinced by the experimental setup and results presented in the paper. Figure 2 uses pretrained models from different sources. I suggest that the authors provide details on the optimizers and training budgets used for each model, and clearly distinguish the contributions of training strategies (e.g., optimizers) from those of RAL itself. Furthermore, Table 1 shows that RAL leads to accuracy degradation in most configurations (I also have concerns about the Lmax⁡=8L_{\max}=8 results; see Question 3). In particular, the degradation in both energy and force accuracy is more pronounced under the important conservative setting. This further weakens the general applicability of the claim that RAL achieves acceleration with only a minor loss in accuracy.

  2. I remain concerned about the applicability of Band-MoE to molecular dynamics (MD) simulations. Specifically, suppose the messages from two experts are mixed as follows:

    m(r)=[1−α(r)]m1(r)+α(r)m2(r).\mathbf{m}(r) = [1-\alpha(r)]\mathbf{m}_1(r) + \alpha(r)\mathbf{m}_2(r).

Taking the derivative with respect to rr yields:

dmdr=(1−α)dm1dr+αdm2dr+α′(r)(m2−m1).\frac{d\mathbf{m}}{dr} = (1-\alpha)\frac{d\mathbf{m}_1}{dr} + \alpha\frac{d\mathbf{m}_2}{dr} + \alpha'(r)\bigl(\mathbf{m}_2-\mathbf{m}_1\bigr).

The stability of NVE MD simulations depends not only on the continuity of the potential energy surface but also on whether the magnitudes of its higher-order derivatives remain sufficiently controlled. Although Band-MoE employs a transition window and other mechanisms to ensure that the potential energy surface is continuously differentiable, these mechanisms do not provide an upper bound of its higher-order derivatives. In particular, the additional perturbation term α′(r)(m2−m1)\alpha'(r)\bigl(\mathbf{m}_2-\mathbf{m}_1\bigr) further raises my concerns. Unfortunately, the paper does not quantify this effect, and the existing MD experiments are conducted only under the NVT ensemble. Therefore, the impact of expert switching on NVE stability remains unverified.

Questions for authors
  1. In Figure 2, are the optimizers used for the proposed method consistent with those used for the other baselines?
  2. Why is the accuracy degradation more pronounced under the conservative setting?
  3. In Table 1, why does the baseline with Lmax⁡=8L_{\max}=8 (Line 319) exhibit worse accuracy than that with Lmax⁡=2L_{\max}=2 (Line 318)? Based on my experience, larger Lmax⁡L_{\max} values may require different optimal training hyperparameters than smaller ones. Moreover, models with more trainable parameters tend to be more prone to training instability on OMol25. Therefore, I suggest that the authors perform additional hyperparameter tuning to ensure a fair comparison.
  4. See Weakness 2. Can the authors provide guarantees regarding the smoothness of Band-MoE?
4 comments

We thank the Reviewer for the comments and great insights. We addressed all concerns and we believe that edits made the manuscript significantly better in both delivery and content than before.

I am not fully convinced by the experimental setup and results presented in the paper. Figure 2 uses pretrained models from different sources. I suggest that the authors provide details on the optimizers and training budgets used for each model, and clearly distinguish the contributions of training strategies (e.g., optimizers) from those of RAL itself. Furthermore, Table 1 shows that RAL leads to accuracy degradation in most configurations (I also have concerns about the results; see Question 3). In particular, the degradation in both energy and force accuracy is more pronounced under the important conservative setting. This further weakens the general applicability of the claim that RAL achieves acceleration with only a minor loss in accuracy.

For Figure 2 we didn’t re-train the literature models but we used official released checkpoints and evaluated them under the same settings (same hardware, same data). UMA, AllScAIP and eSCN are using AdamW with cosine scheduler, while MACE models use AMSGrad Adam, we’ve specified the details in the manuscript.

I remain concerned about the applicability of Band-MoE to molecular dynamics (MD) simulations. ... [reduced for to save characters] ... This further weakens the general applicability of the claim that RAL achieves acceleration with only a minor loss in accuracy. See Weakness 2. Can the authors provide guarantees regarding the smoothness of Band-MoE?

We thank the Reviewer for the insightful comment (similar to Reviewer CWky). We performed 78 new NVE MD (following the eSEN examples) simulations on the MD22 and the OMol25 datasets’ molecules and added the results to the manuscript. In short: the conservative RAL+BandMoE shows the same trajectories as the dense baseline models and none of the conservative trajectories failed under 100ps at 2000K temperature. We believe that these results reflect that our RAL approach does not introduce instabilities to the models.

We also thank the reviewer for raising a concern about the smoothness guarantees. The quintic transition is C^2, with first and second derivatives vanishing at the window boundaries. Assuming the experts and downstream operations are C^2, the energy of the conservative Band-MoE model is C^2, so forces and Hessian remain continuous across expert transitions. Third and higher derivatives need not be continuous at the window boundaries.

In Figure 2, are the optimizers used for the proposed method consistent with those used for the other baselines?

Yes, and we update the manuscript for better readability.

Why is the accuracy degradation more pronounced under the conservative setting?

We agree with the reviewer that the conservative results were lacking details, therefore, we performed more experiments on the conservative force RAL and updated Table 2 with longer training results, which decreased the force and energy errors significantly. We speculated that the larger error (compared to direct force head) is due to that the conservative force differentiates the routing weights in BandMoe, which adds a force term inside each transition window that a direct force head never sees. This issue could be simple addressed with two experts that have an overlapping transition window, which would solve the problem, but beyond the scope of this work.

In Table 1, why does the baseline with (Line 319) exhibit worse accuracy than that with (Line 318)? Based on my experience, larger values may require different optimal training hyperparameters than smaller ones. Moreover, models with more trainable parameters tend to be more prone to training instability on OMol25. Therefore, I suggest that the authors perform additional hyperparameter tuning to ensure a fair comparison.

We thank the Reviewer for being vigilant. We investigated further and the results turned out to be an unnoticed training error. We re-run the L-ablation on several more L-max values (Table 2) and now we confidently confirm that all L-values are correctly reported. These analysis also helped us to pinpoint a sweet spot where RAL with L-max = 6 has lower E/F MAE than the dense L-max=4 models while having similar inference throughput.

See Weakness 2. Can the authors provide guarantees regarding the smoothness of Band-MoE?

I thank the authors for their response and additional experiments. After reviewing the rebuttal and revised manuscript, I still have two concerns:

  1. Unlike UMA's MoLE[1], Band-MoE routes experts based on interatomic distances, which may lead to expert switching during MD simulations. Although the routing is C2C^2-continuous, the force terms and higher-order derivatives introduced by expert switching are not sufficiently controlled. The authors also acknowledge that the method can fail in some applications (e.g., the release of C60, Line 1078).

  2. At the same Lmax⁡L_{\max}, RAL achieves lower accuracy than the dense baseline. Matching the baseline accuracy requires increasing Lmax⁡L_{\max}, which reduces the speedup. For example, in Table 2, RAL with Lmax⁡=5L_{\max}=5 achieves only a 1.20× speedup.

Based on these concerns, I am inclined to maintain my current rating.

[1] Brandon M. Wood et.al UMA: A Family of Universal Models for Atoms. NeurIPS2025

AuthorsOpenReview ↗

Reply on Official Comment by Reviewer ntwT

We thank the Reviewer for their follow-up. What other experiments would convince the Reviewer to consider increasing the score?

On stability: The release of C60 was with the direct-force head model (similarly to other literature models). None of our new NVE simulations fail on the conservative models. These new NVE simulation use all MD22 molecules, including the C60, and the trajectory did not diverge on the three seeds. These results show, that the model is sufficiently controlled and they are stable.

On accuracy at fixed Lmax⁡L_{\max}: At Lmax⁡=4L_{\max}=4, the force accuracy difference remains within the single-run noise threshold (2.6%) across all four evaluated settings. More importantly, at matched accuracy, RAL achieves a 1.20–1.29× speedup, with configurations that are both faster and more accurate than their dense counterparts. We acknowledge that the speedup at matched accuracy is smaller than at fixed Lmax⁡L_{\max} and have clarified this in the revised text. The 31–45% peak-memory reduction remains consistent across all operating points.

Dear Reviewer, as the paper discussion period will end soon, we wanted to follow up. We believe our rebuttal responses address all of your original and follow-up questions.

If our responses have addressed your concerns, we would appreciate it if you would consider raising your score.

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

This paper presents Band-MoE, a novel approach for radial cutoff graph construction in machine-learned interatomic potentials. Band-MoE uses radial gate structures to learn appropriate cutoff radii setting that adapts to different angular resolutions. Benchmark experiments on molecular and crystal benchmarks show good performance and efficiency optimization of the Band-MoE.

Critical strengths
  • This paper presents a novel way to learn good radius cutoff graphs under different angle resolutions for machine-learned interatomic potentials. It is an excellent and insightful contribution to broad AI for science community.
  • Benchmark experiments show that the proposed approach can achieve competing performance-cost tradeoff than baselines.
Critical weaknesses
  • A critical weakness is that there lacks discussion or experiment showing the generalization of Band-MoE idea to other model backbones. Only UMA is used in this paper, but it would be great if authors could discuss if Band-MoE could be a pure plug-in that is usable to any models, and do experiments to show the broad advantages of Band-MoE over different models.
Questions for authors

No other questions, see "Critical Weaknesses" section.

3 comments

We thank the Reviewer for the valuable comment about transferability of our approach.

A critical weakness is that there lacks discussion or experiment showing the generalization of the Band-MoE idea to other model backbones. Only UMA is used in this paper, but it would be great if authors could discuss if Band-MoE could be a pure plug-in that is usable to any models, and do experiments to show the broad advantages of Band-MoE over different models.

We’ve integrated RAL with Band-MoE into NequIP and EquFlashV2, trained on its own config and framework with two different lambdas. For the NequIP L model (L-max=3), RAL-based model outperformed vanilla NequIP by 1.34x-1.48x on batch sizes of 1k to 16k, and its peak memory consumption is 31% less than NequIP. Force and Energy accuracies are +0.47% and +0.85% respectively.

For EquFlashV2 (L-max=4, conservative force head, default settings), the RAL-based model is 1.54x faster at a batch size of 12k atoms, and its peak memory consumption is 45.6% less. Force and energy accuracies are +2.9% and +3.84% respectively.

We further added several other ablation studies and performed more detailed analysis on the behavior of RAL with new NVE MD simulations.

We strongly believe that by showing these new results, the manuscript is substantially better and presents stronger claims than before.

Reviewer_kDhVOpenReview ↗

Follow-up Response

I appreciate authors' efforts in rebuttal. My concerns have been addressed so I will keep my rate.

We thank the Reviewer for the response and we’re glad the new results addressed the main concern about generalization. Is there anything else that would make you consider raising your score?

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

The authors introduce an range-adaptive L_{max}, RAL, layer which each message passing block within an MLIP learns a specific gate radius for each nonzero degree L. The radial gate becomes zero beyond the gate radius and thus

Critical strengths

The paper introduces a plethora of architecture components for MLIPs: RAL, Band-MoE, and reflection-symmetric restriction. The components are novel, physically grounded, and show slight parameter efficiency improvements on carefully selected ablations.

Critical weaknesses

The speedup of the RAL blocks is interesting. I feel the most interesting claim lies in Figure 2 -- showing that the RAL improvement over SOTA models such as UMA and eSEN. However, it's not clear what inference settings were used, especially for UMA. In order for the most fair comparisons to be made, the authors should use the most up to date inference settings provided for each model. Using native PyTorch instead of specifically designed kernels, for example, defeat the purpose of the comparison. Ultimately, practitioners care about raw overall speed over speedup compared to weak baselines. If the authors could also note the inference settings/kernels used, that would help better contextualize the speedup.

If the RAL message passing block is simply masking, where does the additional computational efficiency come from? The gradient signal still gets passed back and computed but is simply multiplied by zeros -- are you doing any explicit computation pruning somehow? This was not clear to me.

Why did you choose the quintic smoothstep function instead of the widely-used polynomial envelope function for ensuring 0 first and second order gradients at the radial gate cutoff? This seems like a non-standard design decision that wasn't mentioned much in the text.

Can you show some long-time energy stability under NVE MD simulation to show the overall conservative nature of your trained MLIPs? I noticed the NVT simulation in Section H; however, the velocity-focused corrections in NVT can hide energy instabilities and general non-conservative potential energy surfaces. Perhaps a high temperature (2000-3000K) simulation at 100 ps, similar to how the eSEN work showed issues with energy stability, would be sufficient (figure 1 of [1]).

Line 250 says "For OMol25, we train on the first 400k structures of its 4M training split, the 4M split and the 102M split, and evaluate on its 2.76M validation structures". What is the second 4M split you trained on?

It is not clear what the second, third, fourth, fifth, and sixth group of rows are indicating. The first and last groups seem to be dense vs. RAL comparisons on the OMol25 and OMat24 datasets respectively. However, the others are not so clear. This should be cleaned up and negatively affects the presentation and understanding of the paper. It also seems as though some rows are copied between different groups of rows.

Smaller points:

  • Line 82 links to Figure 1b, but Figure1b is not related to the primary claims of the sentence (25.9% fewer parameters while increasing energy/force accuracy compared to some baseline)
  • Line 262 and 263 should have author names within parentheses for citations. If you are using latex, that would be \citep{} instead of \citet{}

[1] https://arxiv.org/html/2502.12147v1

Questions for authors

See above.

4 comments

We thank the Reviewer for the comments and questions.

The speedup of the RAL blocks is interesting. I feel the most interesting claim lies in Figure 2 -- showing that the RAL improvement over SOTA models such as UMA and eSEN. However, it's not clear what inference settings were used, especially for UMA. In order for the most fair comparisons to be made, the authors should use the most up to date inference settings provided for each model. ... [reduced to spare characters] ... If the authors could also note the inference settings/kernels used, that would help better contextualize the speedup.

We agree that the inference settings were not detailed enough. We added a detailed description on how each model was compared to each other. In short: we did not use custom kernels and all models were used with their baseline settings. Furthermore, we showed on Nequip (L and XL) and EquFlashV2 that our range adaptive approach can be adapted to other architectures.

If the RAL message passing block is simply masking, where does the additional computational efficiency come from? The gradient signal still gets passed back and computed but is simply multiplied by zeros -- are you doing any explicit computation pruning somehow?

We have clarified these concerns in Section 3.4. In short, masking is used only in Stage A to learn the gate radii and provides no speed. In Stage B edges are grouped by their maximum retained degree kek_e and each group is evaluated with the degree-kek_e operator, which uses smaller slices of the same weights. Higher-degree features are therefore never computed, reducing the O(2) mixing cost in both forward and backward passes.

Why did you choose the quintic smoothstep function instead of the widely-used polynomial envelope function for ensuring 0 first and second order gradients at the radial gate cutoff?

Our radial gate is in fact a polynomial envelope as it satisfies the definition. Also, it equals the DimeNet envelope (Gasteiger et al., 2020) with p=3p=3, u(x)=1−10x3+15x4−6x5u(x) = 1-10x^3+15x^4-6x^5, evaluated at x=(r−(Rl−wl))/wlx=(r-(R_l-w_l))/w_l, i.e. 1−σ5(x)=σ5(1−x)1-\sigma_5(x)=\sigma_5(1-x), which is the quintic smoothstep. We chose p=3p=3 because it’s the lowest-degree member of a family with vanishing first and second derivatives at both ends. The only difference from the standard use over [0,rc][0, r_c] is that we apply the envelope on a window ending at a learnable radius RlR_l, so the gate is one on short edges and exactly zero beyond RlR_l, which Band-MoE requires to skip computation. We now state this equivalence and cite DimeNet in Section 3.3

Can you show some long-time energy stability under NVE MD simulation to show the overall conservative nature of your trained MLIPs? I noticed the NVT simulation in Section H; however, the velocity-focused corrections in NVT can hide energy instabilities and general non-conservative potential energy surfaces. Perhaps a high temperature (2000-3000K) simulation at 100 ps, similar to how the eSEN work showed issues with energy stability, would be sufficient (figure 1 of [1]).

We performed 78 new NVE MD (following the eSEN examples) simulations on the MD22 and the OMol25 datasets’ molecules and added the results to the manuscript. In short: the conservative RAL+Band-MoE shows the same trajectories as the dense baseline models and none of the conservative trajectories failed under 100ps at 2000K temperature. We believe that these results reflect that our RAL approach does not introduce instabilities to the models.

Line 250 says "For OMol25, we train on the first 400k structures of its 4M training split, the 4M split and the 102M split, and evaluate on its 2.76M validation structures". What is the second 4M split you trained on?

Both mentions of “4M” refer to the same official OMol25 4M split: we used its first 400k structures for preliminary experiments and ablations, and the full 4M split for the main results. We have rewritten the sentence in Section 4 to make it clear.

It is not clear what the second, third, fourth, fifth, and sixth group of rows are indicating. The first and last groups seem to be dense vs. RAL comparisons on the OMol25 and OMat24 datasets respectively. However, the others are not so clear. This should be cleaned up and negatively affects the presentation and understanding of the paper. It also seems as though some rows are copied between different groups of rows.

The table mixed several comparisons without labels, and the dense baseline was repeated across groups. We have restructured Table 1 into labeled panels. The Table now much simpler to follow.

Line 82 links to Figure 1b, but Figure1b is not related to the primary claims of the sentence (25.9% fewer parameters while increasing energy/force accuracy compared to some baseline).

Removing the W2 matrix in the eSCN-like SO(2) convolution and enforcing O(2) convolution indeed results in parameter reduction and small accuracy and speed gain detailed in the Results sections and Table 1.

Dear Reviewer, as the paper discussion period will end soon, we wanted to follow up. We believe our rebuttal responses address all of your concerns and questions.

If our responses have indeed addressed your concerns, we would appreciate it if you would consider raising the score.

Dear Authors,

Apologies for the delays!

I took a look a the manuscript and feel that all but one of my concerns have been resolved. However, this may be my largest concern:

We agree that the inference settings were not detailed enough. We added a detailed description on how each model was compared to each other. In short: we did not use custom kernels and all models were used with their baseline settings. Furthermore, we showed on Nequip (L and XL) and EquFlashV2 that our range adaptive approach can be adapted to other architectures.

I believe that in order to make the most fair inference comparisons possible, the authors should benchmark their fastest methods against the fastest common baseline methods. I do not believe it is fair to measure a method claiming speedup over baselines when those baselines are not running at their commonly accepted fastest settings. Until those comparisons are performed, I cannot raise my score.

We thank the Reviewer for the follow-up. We believe that there might be a misunderstanding, because our RAL-method is generally applicable to similar architectures and we showed the applicability of RAL directly on an additional 2 architectures.

In short, our original speed comparison was fair to all models. To test each model's with custom kernels (cuEquivariance, OpenEquivariance or custom CUDA), would require to write these kernels to AllScAIP, eSEN, and to our model as well; and also to retrain all these models from scratch, which is beyond the scope of this work.

Nonetheless, we added Appendix P to the manuscript, where we measured the inference speed with updated settings (fastest mixed bench).

We kindly ask the Reviewer to evaluate our work from the speed-gain perspective and not from raw-speed comparisons.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all reviewers for their valuable feedback and suggestions. We are pleased to hear that the reviews are all positively responded to our work calling it “novel, and physically grounded” (CWky); “an excellent and insightful contribution to broad AI for science community” (kDhV); “easy to understand” and “meaningful research” (ntwT).

We have addressed all reviewer comments and conducted substantial additional experiments, including more comprehensive ablations and molecular dynamics simulations.

A brief overview of the most significant changes:

  • Stable conservative NVE MD simulations for 100ps at 2000K for 13 molecules at different seeds (CWky, ntwT).
  • Showed that RAL+BandMoE works on other architectures, such as Nequip and EquFlashV2 with similar performance gain as before (kDhV).
  • Comprehensive ablation on L-max and parameters search for the conservative RAL (CWky, ntwT).

We believe that after the rebuttal, our manuscript is significantly improved in both the quality and the representation of the results.

Our edits are highlighted in blue in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.