- 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
-
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.
-
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
-
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?
-
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?
-
How do Table 2’s gains break down across energy, force, and stress errors? Does the ranking also hold for phonons or thermal conductivity?
-
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
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 () and edges ().
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 (, CGTP, without Gate or Conv2) and the rich edge-wise CGTP convolution (, 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.
denotes the model configuration, where is the number of channels, is the maximum angular degree, and 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.
| Index | Input | Conv1 | Gate | Conv2 | GFLOPs/atom ↓ | Throughput ↑ | |
|---|---|---|---|---|---|---|---|
| 1 | CGTP | – | – | (128,3,5) | 0.15 | 19.7k | |
| 2 | CGTP | – | – | (128,3,10) | 0.30 | 7.9k | |
| 3 | CGTP | – | – | (128,4,10) | 0.61 | 2.7k | |
| 4 | CGTP | ✓ | – | (128,3,5) | 0.97 | 3.6k | |
| 5 | CGTP | ✓ | CGTP | (128,3,5) | 1.08 | 2.2k | |
| 6 | CGTP | ✓ | CGTP | (128,3,5) | 2.00 | 0.9k | |
| 7 | SO(2) | – | – | (128,3,5) | 1.21 | 2.7k | |
| 8 | SO(2) | – | – | (128,3,10) | 2.41 | 1.3k | |
| 9 | SO(2) | – | – | (128,4,10) | 4.80 | 0.7k | |
| 10 | SO(2) | ✓ | – | (128,3,5) | 1.38 | 2.2k | |
| 11 | SO(2) | ✓ | SO(2) | (128,3,5) | 1.91 | 1.8k |
Table 2: Theoretical computational cost (MFLOPs) of the message-passing architectures.
and 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 Input | Conv1 | Gate | Conv2 | ||
|---|---|---|---|---|---|
| CGTP | – | – | |||
| CGTP | ✓ | – | |||
| CGTP | ✓ | CGTP | |||
| CGTP | ✓ | CGTP |
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.
| Component | Baseline Time (ms) | Baseline MFLOPs/edge | Baseline TFLOP/s | Rich Edge-wise Time (ms) | Rich Edge-wise MFLOPs/edge | Rich Edge-wise TFLOP/s |
|---|---|---|---|---|---|---|
| Radial NN | 1.064 | 0.386 | 16.33 | 1.981 | 0.763 | 17.33 |
| TP1 | 4.041 | 0.093 | 1.04 | 3.539 | 0.091 | 1.16 |
| Linear | – | – | – | 13.157 | 3.637 | 12.44 |
| Gate | – | – | – | 2.161 | 0.023 | 0.48 |
| TP2 | – | – | – | 3.692 | 0.091 | 1.11 |
| Total | 5.105 | 0.479 | 4.22 | 24.530 | 4.605 | 8.45 |
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).
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 to 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 ( or ) specifies the intermediate FFN feature width relative to the channel dimension . Validation loss and training throughput (atoms/sec) are reported.
| Architecture | Gate (1×) Loss ↓ | Gate (1×) Throughput ↑ | SymCon (1×) Loss ↓ | SymCon (1×) Throughput ↑ | SymCon (4×) Loss ↓ | SymCon (4×) Throughput ↑ |
|---|---|---|---|---|---|---|
| Baseline CGTP (Row 1) | 0.231 | 21.8k | 0.220 | 22.2k | 0.212 | 19.7k |
| Scaled CGTP (Row 3) | 0.206 | 2.8k | 0.197 | 2.8k | 0.188 | 2.7k |
| Edge-wise CGTP (Row 6) | 0.199 | 1.0k | 0.193 | 1.0k | 0.192 | 0.9k |
We have incorporated the relevant discussion and table into Section 5.2, Table 3 (pp. 7–8).
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, 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, 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, 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.
We have incorporated the results into Appendix A.6 (p. 20).