acceptodds
Under review as a conference paper at ICLR 2027

BranchIP: Learning Adaptive Equivariant Computation for Interatomic Potentials

Abstract

Equivariant machine learning interatomic potentials (MLIPs) have revolutionized atomistic modeling, but accurate treatment of complex materials and molecular systems demands expensive models. This limits simulation length- and time-scales, with tensor products a key computational bottleneck. The recent emergence of foundation-scale MLIPs further exacerbates this challenge. We present Branch Interatomic Potential (BranchIP), a single-model framework for learned adaptive tensor product computation, trained with a novel distillation loss. In our experiments on two systems of physical interest, a heterogeneous catalysis system and a proton-conducting solid acid electrolyte, BranchIP accelerates MLIPs across model sizes by up to while reducing memory usage by up to . This is achieved while maintaining physical fidelity. Furthermore, the learned adaptive computation provides model interpretability by revealing which interactions demand deeper computation and showing how computational depth relates to chemical complexity and dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning interatomic potentials, adaptive computation, dynamic routing, early exit, equivariant neural networks, tensor products, molecular dynamics, foundation models

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_szg5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a technique for accelerating the inference of equivariant machine learning interatomic potentials (MLIPs). The proposed method is inspired by, though not identical to, early exit techniques in Transformer models. Specifically, a router module placed on the model features before each tensor product layer determines whether to continue the computation for each edge feature, thereby reducing the overall number of operations. To train this router, the authors introduce a knowledge distillation scheme that uses the force discrepancy between the full model (teacher) and the routed model (student) as supervision.

Experiments are conducted on the Allegro architecture, using both a pretrained Allegro foundation model fine-tuned on the target dataset and a model trained from scratch on the target dataset. As a result, the method achieves approximately twofold inference speedup with a modest decrease in energy and force prediction accuracy. The authors additionally present an interpretability analysis, which visually shows which parts of a given material system exit the tensor product computation early and which do not.

Critical strengths
  • The idea of letting the model decide for itself whether to continue the computation is interesting, and the authors effectively bring the concept of early exit into equivariant MLIPs.
  • The authors comprehensively evaluate the method beyond energy and force prediction accuracy, using physically meaningful metrics such as NVE energy conservation and bulk modulus.
Critical weaknesses
  • The evaluation is conducted only on two specialized systems, CO/Pt and CDP, rather than on standard benchmarks, which makes it difficult to intuitively compare the proposed method with existing approaches.
  • Although the paper is generally understandable, there is considerable room for improvement in the writing and figures, as detailed in the next section.
Questions for authors

I have the following questions and suggestions, which I believe could strengthen the work:

  • The proposed method is implemented on Allegro, an edge-centric and strictly local architecture. How would the method behave on atom-centric models such as NequIP?
  • The evaluation is limited to two specialized systems. I suggest additional experiments on standard benchmarks, which could be designed in two settings: (i) evaluating on the data used to pretrain Allegro-OAM-L, and (ii) fine-tuning Allegro-OAM-L on well-known benchmark datasets not included in its pretraining data. In addition, while Figure 5 (a-b) presents routing statistics, these are limited to the specific systems. Could the authors report such statistics on general benchmarks covering diverse materials, as suggested above?
  • It would be informative to include a baseline that distills the full model into a smaller Allegro model with the same number of layers but fewer channels, without routing. Such a comparison would clarify whether adaptive routing offers a better accuracy-efficiency trade-off than conventional distillation into a smaller model.
  • Equivariant MLIPs typically do not stack many layers, which seems to leave limited room for adaptive depth allocation. In this context, the approximately twofold speedup, while meaningful, appears somewhat marginal. Could the authors discuss the practical significance of adaptive depth for equivariant MLIPs?
  • The interpretability analysis is interesting, but its takeaway is unclear. It is not evident whether the routing patterns confirm that the model behaves in a physically reasonable manner, or whether they suggest directions to further improve accuracy or speed. Could the authors elaborate on what insights researchers can draw from this analysis?

In addition, I have the following suggestions regarding presentation:

  • The manuscript is readable overall, but the writing would benefit from substantial polishing.
  • Figure 1 and Figure 2 (a) do not sufficiently convey how the proposed method works. In addition, the two figures appear redundant and could be merged into a single, clearer overview figure.
  • The inverted y-axis in Figure 3 is confusing, even with the explanation in the caption.
  • The font sizes in Figures 3 and 4 should be increased.
  • The colorbar in Figure 3 does not indicate the values at both ends of its range.
  • The 3D plots in Figure 5 (d) hinder readability; 2D plots would be more effective.
  • Line widths, marker sizes, or axis ranges in Figures 7-10 of the Appendix could be adjusted for better readability.

Overall, the idea of adaptive equivariant computation is interesting. However, the contribution remains marginal in its current form, and the presentation requires substantial improvement. I therefore recommend a weak rejection at this stage.


Post-rebuttal update: The additional experiments and revisions in the rebuttal have addressed most of my concerns, and I have updated my score accordingly.

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

The paper proposes BranchIP, which lets an MLIP skip tensor product (TP) computation for edges that no longer need it, in the spirit of early exit networks. Small learned routers placed before TP layers output a gate for each edge, and edges with a zero gate are removed from all later TPs while still receiving scalar updates. The gate is C2C^2 and exactly zero below a threshold, so the sparse evaluation gives the same energy, forces and second derivatives as the dense gated model, and equivariance is preserved. The routers are trained by distillation from a full depth teacher, with targets set by the force error on each edge. Implemented in Allegro, the method is tested on CO on Pt and on CsH$_2$PO$_4$ (CDP), for bespoke models and an Allegro-OAM-L foundation model fine tuned on each system. It reports 1.6×1.6\times to 2.4×2.4\times faster MD and 1.2×1.2\times to 2.6×2.6\times lower memory, and argues that the routing patterns reveal which atomic environments need deeper computation.

Critical strengths
  1. Applying adaptive depth to equivariant MLIPs while keeping forces exactly conservative is a new direction. The flat zero region of the gate ensures that skipped edges contribute nothing to the energy or its derivatives, and MD with the routed models is stable.

  2. The routing patterns are interesting. The model keeps adsorbed CO over gas phase CO and keeps hydrogen bonded H to P edges longer than covalent ones, which gives some insight into where deeper computation is used.

Critical weaknesses
  1. The practical value over simpler or more general accelerations is not established. The speedups are moderate and come with lower accuracy in most cases. In Table 1, the routed models are worse than their teachers on most metrics, most clearly for the CDP foundation model, where the force MAE rises from 19.0 to 26.6 meV/Å and the energy MAE from 0.339 to 0.441 meV/atom. The bespoke CDP model lands roughly halfway between the one and two layer Allegro models (force MAE 35.7 against 40.3 and 31.5), so it is unclear whether routing does better than simply training a smaller dense model, which is never tested at matched speed. Kernel and algorithm level TP accelerations, such as cuEquivariance, FlashTP, OpenEquivariance and SO(2) or Gaunt based products, apply to any model without retraining and do not change accuracy. The paper calls BranchIP complementary to these, but never shows the two combined, and it does not say whether its baselines use optimized TP kernels at inference. Since only the TP is skipped, its benefit shrinks as TPs become cheaper.

  2. The learned router adds cost and raises robustness concerns that are not studied. Training needs a converged teacher, then 100 to 500 more epochs of distillation per system, plus a sweep over routing tolerances and seeds to pick a model, with tolerances set by hand for each layer and system. No new labelled data is needed, but this is a substantial extra training cost. More importantly, the routers are only evaluated on the same distribution they were trained on. For configurations outside it, such as new adsorbates, defects or higher temperatures, a router may skip computation that is needed, and the accuracy loss would go unnoticed since there is no fallback or uncertainty signal. This matters most for the foundation model case. Because it is fine tuned and routed for one system, the resulting model gives up the generality that is the main reason to use a foundation model.

  3. The router target for an edge is the average force error of its two atoms, so the interpretability results may mostly show where errors are large. A causal test and a check across seeds would support the claims.

  4. The active edge set changes at every MD step. It is unclear how this works with compiled models, CUDA graphs or multi GPU runs, since all benchmarks use one GPU.

Questions for authors
  1. Do the Allegro baselines use optimized TP kernels at inference, and what share of their runtime is spent in TPs?

  2. What is the total training cost of a routed model, including the teacher, the distillation and the tolerance sweep, compared with training the teacher alone? During the usage time of the routed model, how many times does the routed model should be used (in table 1 there are experiments use 2 and 4 routed model)? How that can be decided for different datasets?

  3. Can you do a comparison with smaller dense Allegro models at matched MD speed? What is the results with BranchIP combined with optimized TP kernels?

  4. Need to do tests on configurations outside the training distribution, comparing errors and routing behavior with the teacher.

Reviewer_xAw9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

BranchIP introduces learned, feature-level early termination of tensor-product computation in equivariant interatomic potentials. Its Allegro implementation uses invariant routers and smooth gates with exact-zero regions, allowing sparse inference while preserving the gated model’s energy derivatives and symmetries. Experiments cover CO/Pt catalysis and proton-conducting CsH₂PO₄, using both system-specific and fine-tuned foundation models.

Critical strengths
  1. The method addresses a substantive technical constraint. Dynamic skipping must preserve a smooth potential-energy surface. The combination of smooth exact-zero gates and sparse execution provides a coherent solution, with an explicit equivalence argument for energies, forces and Hessians.

  2. The computational gains are meaningful. The selected configurations achieve approximately 1.6–2.4× higher MD throughput and 1.2–2.6× lower memory usage. Evaluation includes seven student-training seeds and one-layer Allegro controls for the system-specific models.

  3. Physical validation goes beyond prediction errors. Energy profiles, RDFs, NVE trajectories and proton diffusion provide useful evidence. In particular, the system-specific Branchlegro model reproduces a diffusion activation energy of 0.40 eV, within the cited experimental range.

Critical weaknesses
  1. The advantage of learned routing is insufficiently isolated. Comparisons with one-layer Allegro are useful, but compute-matched narrower models, distilled shallow models and fixed or distance-based routing would better establish whether learned adaptive depth is necessary. The proposed distillation objective also lack a focused ablation.

  2. The accuracy cost is understated in some settings. For the CDP foundation model, force MAE increases from 19.0 to 26.6 meV/Å, a 40% degradation. The diffusion activation-energy validation covers the system-specific models, so it does not establish that this foundation-model degradation is physically harmless.

  3. Routing patterns support interpretation, but not causal necessity. Retaining an interaction through deeper layers does not establish its intrinsic chemical complexity or required body order. Counterfactual routing interventions would strengthen these claims. Generality beyond Allegro and the two studied systems also remain untested.

Questions for authors
  1. How does learned routing compare with compute-matched static routing and distilled smaller models?

  2. What is the contribution of the proposed distillation loss relative to simpler accuracy-plus-compute objectives?

  3. Can forced termination or continuation of selected edge classes verify the proposed chemical interpretation?

  4. Please quantify additional training cost and uncertainty in the diffusion activation energies, and clarify robustness under changes in temperature and local environment.

I am in between weak accpet and weak reject, and would consider raise my score if the authors can convince me that the method is also solid when comparing with smaller models and static routing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.