acceptodds
Under review as a conference paper at ICLR 2027

EPOCH: Explicit Projection on Concentric Shells for Multi-Order Molecular Geometry

Abstract

For many molecular properties, accurate 3D prediction depends on geometric information across multiple angular scales, from coarse global shape to fine directional detail around local environments. Modern equivariant networks expose this structure through spherical harmonics up to a chosen maximum order , but tensor-product cost grows as per interaction, pushing practical models to a low-order operating point even when chemistry suggests finer angular content matters. We argue that raising is not the only access path to angular structure: finite directional sampling can empirically preserve a useful mid-frequency angular band beyond the practical low-order operating point, at cost linear in the number of probes rather than polynomial in equivariant order. We instantiate this view as EPOCH (Explicit Projection on Concentric Shells), a shared geometric descriptor obtained by sampling a shellwise zonal-kernel response field, precomputed once from coordinates and inserted into scalar, angular, coordinate-equivariant, and tensor-equivariant backbones through lightweight backbone-aware fusion. On QM9, comparisons across SchNet, DimeNet++, EGNN, NequIP, and the strictly local equivariant Allegro show that this shared descriptor helps within heterogeneous backbone families, with the clearest gains concentrated on angular-sensitive electronic targets. Parameter-matched HOMO controls show that preserving response-to-direction correspondence improves over capacity-matched variants, while also clarifying that the effect is target-dependent. Order ablations on a tensor-equivariant backbone further show that EPOCH continues to help as native equivariant order increases, and that augmenting a low-order backbone with EPOCH can often match or remain competitive with a higher-order baseline at a fraction of its parameter and compute cost. Cross-dataset, cross-task evidence on MD17 further shows that the same design improves force prediction. Together, these results indicate that explicit spherical sampling can provide a portable, target-dependent complement to tensor-product equivariance rather than a replacement for it.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, Molecular Representation Learning, AI for Chemistry, Geometric Deep Learning, Equivariant Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_o5rZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce EPOCH, a method for integrating higher-order geometric information into equivariant neural networks. The authors evaluate the effectiveness of EPOCH across a range of molecular property prediction tasks. Empirically, EPOCH improves the expressiveness of different equivariant neural network backbones for such tasks, while allowing less expressive models to achieve results competitive to those of more expressive models via EPOCH's directional sampling mechanism.

Critical strengths
  1. The authors identify a key bottleneck in atomistic representation learning (i.e., efficiently obtaining expressive geometric embeddings) and propose EPOCH to address this bottleneck.
  2. EPOCH works well across a range of equivariant neural network backbones.
  3. EPOCH injects important information for molecular property prediction (and likely other tasks) in a way that reduces the computing requirements needed to achieve high levels of prediction accuracy with a lower computing budget.
Critical weaknesses
  1. The authors do not discuss [1], from what I can tell, which limits the proposed impact of this work, given that new equivariant methods (e.g., EquiformerV2 [2]) have already begun to incorporate eSCN's efficient SO(2) convolutions for higher-order geometric representation learning [1].
  2. The authors do not discuss the inference-time implications of EPOCH, that is, how much using EPOCH slows down a given equivariant method during high-throughput inference of new molecules' properties. It would be good to include this kind of result in this paper.
  3. The authors do not discuss whether or how EPOCH may (negatively) impact the equivariance errors that equivariant methods currently achieve. For instance, does EPOCH introduce additional optimization difficulties or expressiveness constraints beyond improved property prediction accuracy?
  4. In Table 1 of the manuscript, I'd like to see how EPOCH + method X compares to method X alone, where X is a state-of-the-art method for QM9's molecular properties, such as X=EquiformerV2 (as this would test whether EPOCH's additional higher-order directional sampling improves over EquiformerV2's efficient eSCN higher-order convolutions [1, 2]).

References:

[1] Passaro, S., & Zitnick, C. L. (2023, July). Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. In International Conference on Machine Learning (pp. 27420-27438). PMLR.

[2] Liao, Y. L., Wood, B., Das, A., & Smidt, T. (2024, May). EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations. In International Conference on Learning Representations (Vol. 2024, pp. 39282-39309).

Questions for authors
  1. Would the authors please correct the typo descriptor-side side channel that I found on lines 143-144?
  2. Would the authors please rewrite certain phrases in the main text, such as should therefore be read (Line 277) and an MD22 host–guest experiment (Line 468), as these sound very much like awkward LLM-generated phrases?
  3. Less of a question and more of a note: the authors' anonymous GitHub link (provided in the PDF) doesn't work when I tested it.
Reviewer_Gg5gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EPOCH, a lightweight geometric descriptor module for introducing higher-order angular information, without using complex high-order tensor interactions. The experiments demonstrate performance improvements across multiple models: SchNet, DimeNet++, EGNN, NequIP, Allegro.

Critical strengths
  1. The experiments demonstrate performance improvements across multiple models: SchNet, DimeNet++, EGNN, NequIP, Allegro.
  2. Paper is clear, and Figures are well designed for readers to capture the core ideas.
Critical weaknesses
  1. The paper does not sufficiently demonstrate that the proposed method can outperform existing geometric descriptors or simpler geometric features, the experiments primarily compare models only with and without EPOCH. Although Table 5 compares SOAP, this comparison is limited. I believe this kind of comparison should be as important as Table 1.
  2. This paper claim the EPOCH can capture multi-order angular information, while the experiments are limited to demonstrate that. First, the QM9 experiments involving only scalar prediction targets. Second, the MD17 prediction experiments are conducted only with SchNet, where forces are computed via negative gradients.
Questions for authors
  1. Authors used original MD17 dataset rather than the revised MD17 (rMD17) dataset. However, the original MD17 dataset is known to contain numerical inaccuracies. Could authors re-evaluate the results on the revised version?
  2. Could authors provide a comprehensive comparison with other geometric descriptors/features (see W1) like SOAP and other descripters including simpler distance or angle-based features?
  3. Could authors give explanations on the effects of EPOCH, or give evidence that EPOCH captures informative higher-order angular information (see W2)? Like, what if the descriptor is restricted to some spherical harmonic degrees?
  4. The evaluation mainly relies on relatively early backbone architectures. Could the authors provide experiments results on some stronger modern architectures, like mace, gotennet, and the like?
Reviewer_nHK1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose EPOCH, which is a fixed molecular descriptor. It enters SchNet, EGNN, DimeNet++, NequIP, and Allegro through backbone-specific scalar adapters. The authors argue that directional sampling exposes a mid-order angular band beyond low-order equivariant features.

Critical strengths
  1. The descriptor is inexpensive and attaches to five architecturally different backbones without redesigning their core interaction rules.
  2. SchNet force error on MD17 decreases on all ten molecules, with gradients taken through the full descriptor pipeline.
Critical weaknesses
  • The higher-order NequIP baselines seem undertrained or poorly tuned, since the plain first-order model already matches or beats most of them without any help from EPOCH.

  • The claimed mid-order band does not hold up, since the kernel steadily down-weights higher angular degrees. most of the angular content sits at degree four or below, within the NequIP orders already tested. On a shell a Gaussian-smeared SOAP density (Bartok et al., On representing chemical environments) has the same angular form. Thus, the descriptor is close to a molecule-centered SOAP density and its novelty is limited.

Questions for authors
  • could you report seed variance, rerun SchNet and EGNN under your own setup, and retrain the higher-order NequIP baselines so they at least match the first-order model?

  • the principal-axis frame is unstable for symmetric molecules like benzene. How often does it flip along the MD17 trajectories? does the predicted energy jump when it does?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.