acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Edge-wise Computation for Efficient Equivariant Interatomic Potentials

Abstract

Modern equivariant machine learning interatomic potentials are becoming increasingly computationally demanding as their architectures grow more sophisticated. In these architectures, edge-wise operations often account for a substantial portion of the overall computational cost. However, increasing edge-wise complexity is not necessarily the most efficient way to improve model accuracy. Keeping edge-wise computation simple enables effective use of modern fused kernels, substantially improving computational throughput. We therefore introduce EquFlash-S, an equivariant architecture that combines lightweight edge-wise message passing with expressive nonlinear processing in a separate residual node-wise feed-forward block. Its low per-edge cost enables scaling to larger model capacities for higher accuracy while retaining substantially higher throughput than more complex edge-wise designs. EquFlash-S achieves competitive performance across diverse atomistic benchmarks, including Matbench Discovery, the MDR Phonon Benchmark, and OC20. In particular, on Matbench Discovery, EquFlash-S achieves a Combined Performance Score of 0.907 while delivering 2.7–5.7 higher throughput than state-of-the-art models with comparable accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning Interatomic Potentials, Equivariant Neural Networks, Graph Neural Networks

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_U5uUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the trade-off between edge-wise computational complexity, accuracy, and efficiency in equivariant interatomic potentials and introduces EquFlash-S. The model combines simple Clebsch–Gordan tensor-product (CGTP) message passing with nonlinear processing in a separate residual node-wise feed-forward block. It also uses existing fused kernels to reduce computational and memory costs. The authors evaluate the design through architectural ablations and benchmarks including Matbench Discovery, MDR Phonon, OC20, and SPICE-MACE-OFF, demonstrating its prediction accuracy and computational efficiency.

Critical strengths

The model offers a practically useful balance of accuracy and efficiency. EquFlash-S achieves competitive prediction accuracy across several materials and molecular benchmarks while delivering high computational throughput. These results demonstrate the practical value of combining simple edge-wise message passing with expressive node-wise nonlinear processing.

Critical weaknesses
  1. EquFlash-S effectively combines established components: simple CGTP message passing, residual node-wise FFNs, symmetric contractions, and fused kernels. However, the experiments mainly validate this combination, providing limited new insight beyond existing design principles.

  2. Although the ablations include variants inspired by existing models, they mainly use the EquFlash-S framework. Whether the findings on block organization and compute allocation generalize to other CGTP- and SO(2)-based architectures remains unclear. Benchmark comparisons alone do not establish this transferability, limiting their value as broader design guidance.

Questions for authors
  1. Beyond the kernel-optimization ablation in Table 9, could the authors report FLOPs alongside runtimes for the architectural variants in Table 2 at matched atom and edge counts? How much of the advantage comes from fewer operations versus better fusion, reduced memory traffic, and hardware utilization?

  2. Table 2 varies edge-wise complexity while keeping the node-wise FFN fixed. Does the benefit of richer edge-wise processing depend on the FFN’s capacity, and would varying both change the accuracy–efficiency trade-off?

  3. How do Table 2’s gains break down across energy, force, and stress errors? Does the ranking also hold for phonons or thermal conductivity?

  4. Table 7 shows strong results on PubChem and Solvated Amino Acids but weaker results than similarly sized eSEN models on DES370K and Dipeptides. Does the best balance between edge-wise and node-wise processing depend on the task? Understanding when to scale a simple model and when to use richer edge-wise processing would make the findings more useful for architecture selection.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (U5uU) (1/5)

We thank the reviewer for the detailed and constructive feedback. We have conducted additional experiments to further investigate the computational costs, the effect of node-wise FFN capacity, and the generalizability of our architectural findings. The results are summarized below.

W1.:

EquFlash-S effectively combines established components: simple CGTP message passing, residual node-wise FFNs, symmetric contractions, and fused kernels. However, the experiments mainly validate this combination, providing limited new insight beyond existing design principles.

Response:

Thank you for the comment. Please refer to our General Comment on novelty and contributions above.



W2.:

Although the ablations include variants inspired by existing models, they mainly use the EquFlash-S framework. Whether the findings on block organization and compute allocation generalize to other CGTP- and SO(2)-based architectures remains unclear. Benchmark comparisons alone do not establish this transferability, limiting their value as broader design guidance.

Response:

Thank you for highlighting the importance of validating our architectural findings beyond the EquFlash-S framework.

First, as presented in Appendix A.7 (Figure 4), we validated our SO(2) convolution implementation against the original eSEN implementation under matched architectural and training settings. Both implementations use 128 channels, ℓmax⁡=3\ell_{\max}=3, and five interaction layers, and are trained using the same loss function and learning-rate scheduler. Our implementation achieves comparable or lower validation errors across the tested learning rates, supporting its use as a reliable SO(2)-based baseline for the architectural ablations in Section 5.2.

To further investigate the transferability of our interaction-block design, we conducted additional experiments within the MACE framework. To preserve the original MACE architecture as much as possible, we retained its layer-wise readout blocks and element-wise linear self-connections while modifying the organization of the interaction block. Both MACE variants use 128 channels, ℓmax⁡=2\ell_{\max}=2, and five interaction layers, and were trained for 30 epochs.

As shown in the table below, introducing the EquFlash-S-style interaction block improves the energy, force, and stress MAEs from 71.64 to 69.10 meV/atom, 127.03 to 112.40 meV/Å, and 4.64 to 4.30 meV/ų, respectively.

These results suggest that the benefit of the interaction-block organization is not restricted to EquFlash-S and can also be observed within an existing CGTP-based architecture.

While the eSEN comparison validates our SO(2) implementation rather than directly establishing architectural transferability, the additional MACE experiment provides evidence that our interaction-block design can be beneficial beyond the EquFlash-S framework. We acknowledge that broader generalizability of the compute-allocation findings across different architectures remains to be investigated.

Table 7: Comparison of the original MACE interaction block and the EquFlash-S-style interaction block implemented within the MACE framework.

Both models use 128 channels, ℓmax⁡=2\ell_{\max}=2, and five interaction layers, and were trained for 30 epochs. Energy, force, and stress MAEs are reported in meV/atom, meV/Å, and meV/ų, respectively. Lower values indicate better performance.

ArchitectureE/Atom ↓Force ↓Stress ↓
MACE Baseline71.64127.034.64
MACE with EquFlash-S interaction block69.10112.404.30

We have incorporated the results into Appendix A.6 (p. 20).

AuthorsOpenReview ↗

Response to Reviewer 3 (U5uU) (2/5)

Q1.:

Beyond the kernel-optimization ablation in Table 9, could the authors report FLOPs alongside runtimes for the architectural variants in Table 2 at matched atom and edge counts? How much of the advantage comes from fewer operations versus better fusion, reduced memory traffic, and hardware utilization?

Response:

Thank you for the suggestion. We have measured the theoretical FLOPs (forward pass only) of the architectural variants in Table 2 of the manuscript, assuming an average of 45 edges per atom, consistent with the throughput measurements. The results are summarized in Table 1 below. Table 2 additionally reports the theoretical FLOPs per interaction block as a function of the numbers of nodes (NN) and edges (EE).

As shown in Table 1, adding edge-wise Gate and a second CGTP (Row 1 vs. Row 5) increases the computational cost from 0.15 to 1.08 GFLOPs/atom, while reducing training throughput from 19.7k to 2.2k atoms/sec, under the same model configuration.

We further measured the FLOPs and execution time of individual convolution components for the baseline CGTP convolution (XjX_j, CGTP, without Gate or Conv2) and the rich edge-wise CGTP convolution (XjX_j, CGTP, with Gate and Conv2), as shown in Table 3.

The rich edge-wise convolution requires 9.6× more FLOPs (4.605 vs. 0.479 MFLOPs/edge) and 4.8× longer execution time(24.530 vs. 5.105 ms per 1,000 atoms) than the baseline convolution. In particular, the Linear operation preceding the gate activation accounts for approximately 79% of the FLOPs and 54% of the execution time of the rich edge-wise convolution.

Although the Linear operation achieves higher computational throughput in TFLOP/s than the tensor products, its large operation count makes it the dominant runtime contributor. This indicates that reducing the number of edge-wise operations is a major contributor to the computational advantage of the baseline design, rather than simply achieving higher hardware utilization per operation.

Table 1: Comparison of CGTP- and SO(2)-based message-passing layers under architectural modifications and model scaling.

(C,ℓmax⁡,L)(C,\ell_{\max},L) denotes the model configuration, where CC is the number of channels, ℓmax⁡\ell_{\max} is the maximum angular degree, and LL is the number of interaction blocks. GFLOPs/atom reports the theoretical FLOPs per atom for a forward pass, assuming an average of 45 edges per atom. Throughput (atoms/sec) were measured during training.

IndexInputConv1GateConv2(C,ℓmax⁡,L)(C,\ell_{\max},L)GFLOPs/atom ↓Throughput ↑
1XjX_jCGTP––(128,3,5)0.1519.7k
2XjX_jCGTP––(128,3,10)0.307.9k
3XjX_jCGTP––(128,4,10)0.612.7k
4XjX_jCGTP✓–(128,3,5)0.973.6k
5XjX_jCGTP✓CGTP(128,3,5)1.082.2k
6Xj∥XiX_j \Vert X_iCGTP✓CGTP(128,3,5)2.000.9k
7Xj∥XiX_j \Vert X_iSO(2)––(128,3,5)1.212.7k
8Xj∥XiX_j \Vert X_iSO(2)––(128,3,10)2.411.3k
9Xj∥XiX_j \Vert X_iSO(2)––(128,4,10)4.800.7k
10Xj∥XiX_j \Vert X_iSO(2)✓–(128,3,5)1.382.2k
11Xj∥XiX_j \Vert X_iSO(2)✓SO(2)(128,3,5)1.911.8k

Table 2: Theoretical computational cost (MFLOPs) of the message-passing architectures.

NN and EE denote the numbers of nodes and edges, respectively. The computational cost includes the message-passing and node-wise feed-forward operations within a single interaction block, for one forward pass.

Edge InputConv1GateConv2C=128,ℓmax⁡=3C=128,\ell_{\max}=3C=128,ℓmax⁡=4C=128,\ell_{\max}=4
XjX_jCGTP––8.52N+0.48E8.52N+0.48E16.71N+1.00E16.71N+1.00E
XjX_jCGTP✓–8.52N+4.14E8.52N+4.14E16.71N+9.23E16.71N+9.23E
XjX_jCGTP✓CGTP8.52N+4.61E8.52N+4.61E16.71N+10.22E16.71N+10.22E
Xj∥XiX_j \Vert X_iCGTP✓CGTP8.52N+8.71E8.52N+8.71E16.71N+19.40E16.71N+19.40E

Table 3: Component-wise runtime and computational throughput of the baseline and rich edge-wise CGTP convolutions.

Time is measured in ms per 1,000 atoms, and FLOPs are reported in MFLOPs per edge. Achieved computational throughput is reported in TFLOP/s, assuming an average of 45 edges per atom.

ComponentBaseline Time (ms)Baseline MFLOPs/edgeBaseline TFLOP/sRich Edge-wise Time (ms)Rich Edge-wise MFLOPs/edgeRich Edge-wise TFLOP/s
Radial NN1.0640.38616.331.9810.76317.33
TP14.0410.0931.043.5390.0911.16
Linear–––13.1573.63712.44
Gate–––2.1610.0230.48
TP2–––3.6920.0911.11
Total5.1050.4794.2224.5304.6058.45

We have added a GFLOPs/atom column to Table 2 (p. 6) and included a detailed FLOPs and runtime analysis in Appendix A.10 (p. 23).

AuthorsOpenReview ↗

Response to Reviewer 3 (U5uU) (3/5)

Q2.:

Table 2 varies edge-wise complexity while keeping the node-wise FFN fixed. Does the benefit of richer edge-wise processing depend on the FFN’s capacity, and would varying both change the accuracy–efficiency trade-off?

Response:

Thank you for this insightful question. To investigate the interaction between edge-wise complexity and node-wise FFN capacity, we conducted additional experiments with three representative architectures (Rows 1, 3, and 6 in Table 2 of the manuscript).

Specifically, we varied the node-wise FFN design by replacing Gate with SymmetricContraction (SymCon) and increasing the intermediate FFN width from 1×1\times to 4×4\times the channel dimension.

The results show that the relative benefit of richer edge-wise processing varies with the node-wise FFN design and capacity. With Gate (1×) and SymCon (1×), the rich edge-wise CGTP architecture (Row 6) achieves lower validation losses than the scaled CGTP architecture (Row 3): 0.199 vs. 0.206 and 0.193 vs. 0.197, respectively.

However, with the more expressive SymCon (4×) FFN, the scaled CGTP architecture achieves a lower validation loss (0.188 vs. 0.192) while maintaining 3× higher training throughput (2.7k vs. 0.9k atoms/sec).

These findings suggest that scaling a simpler edge-wise architecture can provide a more favorable accuracy–efficiency trade-off when combined with a more expressive node-wise FFN.

Table 4: Effect of node-wise FFN design and capacity on the accuracy–efficiency trade-off of three representative CGTP architectures (Rows 1, 3, and 6 in Table 2).

The width multiplier (1×1\times or 4×4\times) specifies the intermediate FFN feature width relative to the channel dimension CC. Validation loss and training throughput (atoms/sec) are reported.

ArchitectureGate (1×) Loss ↓Gate (1×) Throughput ↑SymCon (1×) Loss ↓SymCon (1×) Throughput ↑SymCon (4×) Loss ↓SymCon (4×) Throughput ↑
Baseline CGTP (Row 1)0.23121.8k0.22022.2k0.21219.7k
Scaled CGTP (Row 3)0.2062.8k0.1972.8k0.1882.7k
Edge-wise CGTP (Row 6)0.1991.0k0.1931.0k0.1920.9k

We have incorporated the relevant discussion and table into Section 5.2, Table 3 (pp. 7–8).

Reviewer_c8fQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce EquFlash-S, a family of 3 MLIPs (EquFlash-S-small, EquFlash-S-medium, and EquFlash-S-large) achieving 2.7-5.7x higher throughput in terms of atoms/s/step of inference compared to other models.

Critical strengths

The authors introduce a novel architecture as well as perform a detailed architectural analysis in section 5 regarding the validation loss relationship with the design of interaction blocks and message passing layers. The study of replacing memory and compute-expensive edgewise computations with node-centered or equivariant operations for MLIPs is a noble and important pursuit.

Critical weaknesses

My primary concern lies with the baselines. I believe Matbench Discovery is quite dated and has been effectively saturated for a while. I'm surprised to not see comparisons against the currently accepted SOTA models such as UMA [1] , the UPET line of models [2], and Orbv3/OrbMol v2 [3] -- I noticed these works were all cited but not benchmarked against? Why is that?

I also think some comparisons on the OMol25 dataset, currently the de facto flagship molecule dataset, would help contextualize the work. I do realize this is very challenging to do given the time constraints of the rebuttal period, but at the minimum, some comparison with the smaller 4M train set would be helpful.

[1] https://arxiv.org/html/2506.23971v2 [2] https://github.com/lab-cosmo/upet [3] https://www.orbitalindustries.com/news/blog/introducing-orbmol-v2

Questions for authors

See above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (c8fQ) (1/2)

We thank the reviewer for the constructive feedback and valuable suggestions. We have expanded our baseline comparisons in the revised manuscript and conducted additional experiments on OMol25 to address the concerns.

Q1.:

My primary concern lies with the baselines. I believe Matbench Discovery is quite dated and has been effectively saturated for a while. I'm surprised to not see comparisons against the currently accepted SOTA models such as UMA [1], the UPET line of models [2], and Orbv3/OrbMol v2 [3] -- I noticed these works were all cited but not benchmarked against? Why is that?

Response:

Thank you for pointing out the missing comparisons with recent state-of-the-art models. We agree that including these baselines is important for better contextualizing the performance of EquFlash-S.

We have expanded the comparisons in Table 3 (Matbench Discovery) by adding Orb-V3, UMA-m-1.1, UMA-s-1.2, and PET-OAM-XL (UPET family) in the non-compliant setting. With these additional baselines, EquFlash-S-xlarge achieves the highest CPS of 0.910, compared to 0.861 for Orb-V3, 0.889 for UMA-m-1.1, 0.893 for UMA-s-1.2, and 0.898 for PET-OAM-XL.

We have also added UMA-m-1.1, UMA-s-1.2, and PET-OAM-XL to Table 5 (MDR Phonon Benchmark). EquFlash-S-large achieves the lowest errors across all four evaluated phonon properties among the compared models.

In addition, we provide comparisons with recent molecular MLIPs, including OrbMol-v2, on OMol25 in our response to Q2 below.

AuthorsOpenReview ↗

Response to Reviewer 2 (c8fQ) (2/2)

Q2.:

I also think some comparisons on the OMol25 dataset, currently the de facto flagship molecule dataset, would help contextualize the work. I do realize this is very challenging to do given the time constraints of the rebuttal period, but at the minimum, some comparison with the smaller 4M train set would be helpful.

Response:

Thank you for the suggestion. To further evaluate EquFlash-S on a recent large-scale molecular benchmark, we conducted additional experiments on OMol25-0 using direct force prediction. We trained EquFlash-S for 80 epochs on the 4M subset and 10 epochs on the full OMol25-0 dataset. The results are summarized in the tables below.

For OrbMol-v2, as we could not find publicly reported results on the OMol-0 validation set, we evaluated the model ourselves.

OMol25-0 (4M subset). Comparing models trained with direct force prediction, EquFlash-S-xxlarge achieves a total energy MAE comparable to eSEN-md-d. (1.81 vs. 1.78 kcal/mol), while achieving a lower total force MAE (0.14 vs. 0.16 kcal/mol/Å). Notably, EquFlash-S-xxlarge achieves 2.8× higher inference throughput than eSEN-md-d. (16.46 vs. 5.81 katoms/s), demonstrating a favorable accuracy–efficiency trade-off.

For throughput measurements, we selected a single atomic structure from the dataset and evaluated inference throughput on an NVIDIA H100 GPU across batch sizes of 8, 16, 32, 64, 128, 256, 512, and 1024. We report the maximum throughput achieved across these batch sizes for each model. Table 1: OMol25-0 (4M) out-of-distribution composition validation results and peak inference throughput.

Energy (E) and force (F) MAEs are reported in kcal/mol and kcal/mol/Å, respectively. Peak inference throughput is reported in thousands of atoms per second (katoms/s), measured on an NVIDIA H100 GPU using a single atomic structure sampled from the dataset, with batch sizes ranging from 8 to 1024. The maximum throughput across batch sizes is reported for each model. Total metrics are evaluated on the full validation set. The best result for each metric is highlighted in bold.

ModelParam. (M)Biomol. E ↓Biomol. F ↓Electrolytes E ↓Electrolytes F ↓Metal E ↓Metal F ↓Organics E ↓Organics F ↓Total E ↓Total F ↓Throughput ↑
eSEN-sm-d.6.332.940.193.110.294.450.931.380.472.990.30102.09
eSEN-sm-cons.6.332.900.142.710.263.610.820.970.392.650.2670.09
eSEN-md-d.50.671.620.081.810.153.290.630.770.211.780.165.81
EquFlash-S-xlarge33.871.750.082.190.163.450.630.750.212.050.1627.86
EquFlash-S-xxlarge57.231.550.071.920.143.210.570.640.181.810.1416.46

OMol25-0 (full dataset). Due to the limited time available during the rebuttal period, we trained only EquFlash-S-xlarge on the full OMol25-0 dataset. EquFlash-S-xlarge achieves a total force MAE of 0.11 kcal/mol/Å, matching that of eSEN-md-d. Although its total energy MAE is higher (1.48 vs. 1.15 kcal/mol), the results demonstrate competitive force prediction accuracy on this large-scale molecular benchmark.

Table 2: OMol25-0 (full dataset) out-of-distribution composition validation results.

Energy (E) and force (F) MAEs are reported in kcal/mol and kcal/mol/Å, respectively. Total metrics are evaluated on the full validation set. The best result for each metric is highlighted in bold.

ModelBiomol. EBiomol. FElectrolytes EElectrolytes FMetal EMetal FOrganics EOrganics FTotal ETotal F
eSEN-sm-d.2.240.152.030.223.350.760.790.322.060.23
eSEN-sm-cons.2.000.111.690.193.060.670.530.261.770.19
eSEN-md-d.1.170.061.120.102.250.460.380.131.150.11
MACE-OMol-L-06.120.154.620.263.750.731.310.354.560.25
UMA-S-1.1 (OMol)1.550.131.390.222.760.710.560.311.450.22
UMA-M-1.1 (OMol)1.360.081.050.132.160.510.340.151.140.14
OrbMol-v20.760.120.900.202.810.700.490.320.980.21
EquFlash-S-xlarge1.390.061.540.112.430.460.420.121.480.11

Overall, these additional experiments demonstrate that EquFlash-S remains competitive with recent molecular MLIPs on OMol25, achieving comparable or lower force errors while offering a favorable accuracy–efficiency trade-off on the 4M subset.

We have updated Section 6.4 (OMol25, p. 9) with the OMol25-4M results and Appendix A.17 (p. 29) with the OMol25 full-dataset results.

Apologies for the delays in response!

Overall, I feel as though the additional experiments on OMol as well as the additional comparisons on MatBench Discovery are quite reasonable. EquFlash achieves higher throughput for a fixed energy and force accuracy level. Therefore, I will raise my score.

Reviewer_3W9VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper explores the trade-off of computation cost and accuracy in the design of MLIP using spherical harmonics.

The paper proposes to reduce the cost of computing the spherical tensors when aggregating messages from neighbour nodes, while leaving non-linearity at the node level.

The paper performed ablation studies to decide which operations to use and which to remove at the message passing and at the node level.

The paper presents the performance in terms of accuracy (forces and energy) and throughput (samples per second).

Critical strengths

The paper addresses a relevant topic, creating a fast MLIP but still accurate.

In A.3 COMPARISON OF EQUIVARIANT INTERACTION DESIGNS, the paper shows the connection among other architectures.

Extensive evaluation has been performed on a good number of datasets.

Performance shows improvement in speed.

The model selection/ablation study is interesting.

Critical weaknesses

The paper does not really propose a new architecture but a model selection on others' architectures.

Small improvement in accuracy, but this is not the aim.

Sometimes the loss is reported and not the actual train/test error. The unit are present only in the SPICE dataset.

Not always clear what the mathematical operation behind the name of the function (e.g. Linear) or operator, for example in Eq.(1) and the previous equation.

Questions for authors

Could you write the mathematical operations explicitly?

It would be nice if the authors would present a generic framework, then propose the model selection.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3W9V

We thank the reviewer for the thoughtful review and constructive suggestions. We have revised the manuscript to address the questions and concerns, and attached an updated PDF with the modifications highlighted.

W4.&Q1. :

Not always clear what the mathematical operation behind the name of the function (e.g., Linear) or operator is.

Could you write the mathematical operations explicitly?

Response:

Thank you for the suggestion. We have added brief explanations of the equivariant operations, including Linear, Norm, Gate, and SymmetricContraction, in Section 2 Background (p.3). We have also added their explicit mathematical definitions, along with the path-weighted Clebsch–Gordan tensor product, in Appendix A.1. (p.15)


Q2.:

It would be nice if the authors would present a generic framework, then propose the model selection.

Response:

Following the reviewer's suggestion, we have reformulated the interaction block using a general two-stage framework in Section 4.2(p.4):

\begin{align} \tilde{\mathbf{X}} &= \mathcal{B}_M \left( \mathbf{X}^{n}, \mathcal{M}(\mathcal{N}_M(\mathbf{X}^{n});\mathbf{r}) \right),\\ \mathbf{X}^{n+1} &= \mathcal{B}_F \left( \tilde{\mathbf{X}}, \mathcal{F}(\mathcal{N}_F(\tilde{\mathbf{X}})) \right). \end{align}

Here, BM\mathcal{B}_M and BF\mathcal{B}_F specify the interaction-block organization, M\mathcal{M} and F\mathcal{F} denote message passing and node-wise feed-forward operations, respectively, and NM\mathcal{N}_M and NF\mathcal{N}_F represent optional normalization.

This formulation provides a common framework for systematically examining three architectural design choices: interaction-block organization (Section 5.1), message-passing design (Section 5.2), and node-wise feed-forward design (Appendix A.5). We have revised the corresponding sections to clarify how EquFlash-S is identified through these architectural analyses.


W3.:

Sometimes the loss is reported and not the actual train/test error. The units are present only in the SPICE dataset.

Response:

Thank you for pointing this out. To complement the validation losses reported in Table 2, we have added a detailed breakdown of energy, force, and stress MAEs for the architectural variants in Appendix A.11 (p.25). We also report phonon-property prediction errors and the thermal conductivity metric κSRME\kappa_{\mathrm{SRME}} for these variants.

In addition, we have clarified the physical units of the evaluation metrics in the MDR Phonon Benchmark (p.9) and MatCalc benchmark (Appendix A.14,p.27), as well as in the newly added Omol25 results (p.10,p.29)


W1.:

The paper does not really propose a new architecture but a model selection on others' architectures.

Response:

Thank you for the comment. We have addressed this concern regarding novelty and contributions in our General Comment above.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment: Novelty and Contributions

We thank the reviewers for their constructive feedback regarding the novelty and informativeness of our work. During the rebuttal period, we also developed and evaluated an additional scaled model, EquFlash-S-xlarge, which achieves a new state-of-the-art Combined Performance Score (CPS) of 0.910 on Matbench Discovery, surpassing the previous best score of 0.908.

We acknowledge that EquFlash-S builds upon existing equivariant operations rather than introducing a fundamentally new equivariant primitive. However, our primary contribution lies in the systematic investigation of architectural design choices for efficient equivariant MLIPs.

In recent years, many high-performing equivariant MLIPs have adopted SO(2)-based convolutions and increasingly expressive edge-wise processing, including additional equivariant transformations, nonlinearities, and attention mechanisms. While these approaches have achieved strong benchmark performance, we revisit whether such complexity is necessary to obtain competitive accuracy and computational efficiency. In particular, our work demonstrates that simpler CGTP-based message passing, when combined with an appropriate interaction-block design and effective model scaling, can achieve competitive or even superior accuracy–efficiency trade-offs.

To investigate this question, we organize representative designs from existing equivariant MLIP architectures within a unified framework consisting of three components: interaction-block organization, message-passing design, and node-wise feed-forward design. Within this framework, we conduct comprehensive architectural studies to understand the accuracy–efficiency trade-offs of different design choices and identify an effective combination, which forms the basis of EquFlash-S.

In response to the reviewers' valuable suggestions, we have further strengthened our architectural analysis with additional experiments. In particular, we investigated the relationship between edge-wise expressivity and node-wise FFN capacity, and conducted detailed FLOPs and component-wise runtime analyses to better understand the computational costs of different message-passing designs. These additional results provide further insights into the trade-offs involved in allocating computation between edge-wise and node-wise operations.

Based on these architectural findings, EquFlash-S achieves competitive or state-of-the-art performance across diverse atomistic benchmarks, spanning bulk materials (Matbench Discovery and MDR Phonon Benchmark), molecular systems (OMol25), and catalytic systems (OC20). In particular, EquFlash-S achieves approximately 2.5–2.6× higher inference throughput than recent state-of-the-art models on Matbench Discovery while maintaining competitive or better accuracy. On the OMol25-0 4M subset, EquFlash-S-xxlarge also achieves competitive energy accuracy and lower force error than eSEN-md-d., with approximately 2.8× higher inference throughput.

Furthermore, we conducted additional experiments within the MACE framework, where adopting the EquFlash-S-style interaction block improves energy, force, and stress prediction accuracy. Although this does not establish universal generalizability, it suggests that some of our architectural findings may also be applicable beyond the EquFlash-S framework.

We therefore hope that the contribution of this work can be appreciated primarily as a systematic architectural study that provides useful empirical findings and design guidance for efficient equivariant MLIPs, rather than as the introduction of a new equivariant operation. We have revised the Introduction to clarify this perspective, and we will further update the revised manuscript to emphasize this framing throughout the paper and incorporate the additional analyses and findings discussed in our responses.

Manuscript Revisions. We have revised the Introduction and Conclusion to better reflect the main contributions of our work and incorporate the additional findings from the rebuttal period. We have also reorganized the manuscript to place greater emphasis on our key results. Specifically, we have added the OMol25 results (Section 6.4, pp. 9–10) and the experiments comparing node-wise FFN designs and scaling strategies (Section 5.2, pp. 7–8) into the main text. Accordingly, the SPICE-MACE-OFF results and the throughput analysis across different edge densities have been moved to the appendix (Appendix A.16, p. 28, and Appendix A.9, p. 22, respectively).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.