acceptodds
Under review as a conference paper at ICLR 2027

E2Former-V3: Atom-Role-Aware Equivariant Interatomic Potentials

Abstract

Machine-learned interatomic potentials enable high-level accuracy for large-scale biological simulation, yet remain costly. In aqueous systems, hydrogens and the water solvent dominate the atom count, while uniform equivariant architectures give every atom the same representation capacity and spend most of their computation on them. We introduce E2Former-V3, an atom-role-aware equivariant potential architecture that represents every atom explicitly and keeps forces conservative. A static chemical role, assigned before the backbone runs, sets each atom’s latent width and how often it is updated. Two configurations follow. E2Former-V3-Standard routes hydrogens through half as many full-width block updates, reaching a force MAE of 8.146 meV/ ˚A, slightly better than E2Former-V2 (Huang et al., 2026), at 1.75–1.84× its throughput. E2Former-V3-Flash routes hydrogens and water oxygens through a narrow stream with fewer updates, trading 3.126 meV/ ˚A higher force error for 4.00–5.85× the throughput. We run molecular dynamics over water, a salt solution and a solvated protein for 20 ps, across three ensembles and three seeds. The runs hold their target temperature, keep chemical structure intact, and reproduce the measured liquid structure to within a few hundredths of an ˚Angstrom. On the solvated protein, both production checkpoints complete the full 1 ns without water fusion, nonbonded clashes or solute collapse.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Equivariant Graph 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_kStQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

E2Former-V3 assigns different representation widths and update frequencies to atoms according to predefined chemical roles. It retains explicit atoms and energy-derived forces, and introduces neighbor-union filtering to reduce unnecessary computation. The paper reports substantial inference speedups, memory savings, and molecular-dynamics validation.

The implementation is useful, but the architectural idea is straightforward. Acceptance would require convincing evidence that the proposed allocation achieves a better accuracy–cost trade-off than simpler alternatives. That distinction is currently unproven.

Critical strengths

The method addresses a relevant bottleneck: solvent atoms dominate the computational cost of explicitly solvated systems.

The reported speed and memory improvements are practically meaningful. Neighbor-union filtering is clearly motivated and supported by an on/off comparison.

The authors report physical failures and limitations alongside successful simulations, providing useful information beyond aggregate prediction errors.

Critical weaknesses
  1. The methodological contribution is incremental. The design combines role-dependent conditional computation, established equivariant channel projections, cross-role attention, and a conventional energy-gradient readout. These components form a reasonable architecture, but the paper provides limited insight into why the particular role partition, widths, and update schedules should be preferred. Preserving equivariance and conservative forces through standard constructions should not itself be counted as a separate methodological advance.

  2. The decisive comparison with simpler models is missing. Appendix I.3 acknowledges the absence of a recipe-matched uniform model and a same-fraction random-routing control. A uniformly narrower or shallower model could potentially achieve a similar trade-off. This concern is particularly important for pure water, where Flash bypasses the high-resolution stack entirely. The experiments therefore demonstrate that the proposed configurations are faster, but do not establish the advantage of chemically informed allocation over ordinary model reduction.

  3. The accuracy–efficiency claims require stronger qualification. The small reported force improvement over V2 uses single-seed results and a published baseline whose metric convention is acknowledged as unspecified. Aggregate accuracy also obscures relevant regressions: Flash’s SPICE water force MAE is 16.28 versus 5.26 meV/Å for V2. The 1 ns protein trajectories demonstrate structural stability for one system, but do not resolve the reported solvent thermodynamic discrepancies or softened short-range repulsion.

Questions for authors

At matched inference latency or memory, does role-aware allocation outperform uniformly narrower and shallower models?

Does chemical routing outperform random or alternative partitions when actual computational cost is matched?

Can the main accuracy comparisons be repeated with identical metrics, energy referencing, and training protocols?

Which solvent observables remain accurate at the reported speedups, and how robust are they across longer, independently replicated simulations?

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

The paper proposes to allocate computation resources differently to different atoms.

Hydrogen and water atoms constitute the majority of the atoms in a simulation.

The paper proposes two architectures: 1) high-resolution includes water's oxygen atoms, 2) low-resolution component that contains hydrogen and water's oxygen atoms.

The architectures also differ in the latent dimensions and the attention operations.

The paper proposes two types of experiments: 1) short trajectories (20ps) and 2) two "long" simulations (1ns)

The experiments in throughput and accuracy show the compromise between the two, allowing simulation of up to 74k atoms in a 40Gb GPU.

Critical strengths
  1. The paper is clear in its objective: allocate computational resources to different atom types
  2. The proposed architecture with different update or feature sizes is a clear change in the base architecture
  3. The experiments consider more realistic systems
  4. The throughput and memory allocation improve with respect to the state of the art
Critical weaknesses
  1. The contribution is limited
  2. The architecture is mostly described in words; therefore, it is not clear, and there is no proper mathematical description
  3. Chapter 3.3 is not clear and uses a function name typical of AI-written text. Similar in 3.2
  4. For Figure 1 and the architecture description (A2H, A2L), the approach is heuristic
Questions for authors

The architecture description is either in words or as a block diagram. This should be expanded.

Reviewer_7RqpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce E2Former-V3, a new model within the E2Former family of MLIPs which dynamically route different atoms through a varying number of block updates in pursuit of efficiency gains -- particularly in water solvated systems. The method achieves a slightly better force MAE compared to its predecessor, E2Former-V2, while at 1.75x-1.84x its throughput.

Critical strengths

The method seems relatively straightforward, is grounded in sensical physical intuition, and shows strong results on the modeling of water.

Critical weaknesses

The paper is clearly written with Claude Opus 5 or similar. The prose is extremely difficult to understand, and I encourage the authors to clean up the writing to be more clear -- particularly section 3.2 which actually introduces the methods...

The most important benchmark results are no doubt in Table 1 -- the energy and force results on the seminal OMol25 dataset. It does not appear that E2Former-V3 displays strong results on this dataset and more modern MLIPs are not benchmarked against the SPICE test set in Table 2. The throughput gains are particularly interesting; however, it's not clear what inference settings were used for models such as UMA. UMA, in particular, has specific-designed triton kernels that are becoming standard in the inference stack. It would not be a fair comparison otherwise.

One question I had was how the short circuiting behavior of hydrogens and oxygens throughout the MLIP can still lead to physically conservative and smooth energy surfaces. Can the authors elaborate on this? This was hard to understand. Furthermore, I noticed NVE energy drift results were reported over 2-20 ps in Figure 3c; however, 20 ps is not a sufficient period of time to empirically establish or dismiss energy drifting concerns. I recommend the authors to show results on 100 ps of high temperature NVE simulation (1000K or higher) to fully stress test the system.

Overall, I believe the very unnatural and difficult-to-digest writing of the work combined with the unclear inference settings in baselines warrants a weak reject.

Questions for authors

See above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.