- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents a method to do O(3)-Equivariant convolutions in a more space and time efficient manner by utilizing edge alignment and multiple O(2) frames to present a method that is O(3) equivariant while keeping the speedups of eSCN. They incorporate this new method of O(3) equivariance along with time reversal and magnetic moment calculation to create a more effective MLIP, mTACE. They show through validation that mTACE is able to outperform other state of the art MLIPs in prediction of molecules in CrN and Fe-DeepSpin.
Critical strengths
The paper provides clear and thorough proof as well as empirical tests that there method has a low amount of equivariance error when it comes to augmenting the structure of data. The mix of theoretical and empirical results give confidence to a reader that their method is an effective way to deal with equivariance. This allowed me to be confident that they were actually able to achieve equivariance in a reasonable setting.
Critical weaknesses
Their paper is lacking an analysis of the performance in terms of wall clock time of how their method compare to a full clebsch-gordan tensor product method of O(3) equivariance. While theoretically they show a better performance in terms of speed and space usage. It could be different in reality where the use of multiple O(2) frames may cause higher slowdowns due to a larger constant time operation. I would like to see some evidence that this method has the benefits it says it does. While this is missing, it is not as significant as we can see that their performance on molecular dynamics simulations is already quite an improvement.
Questions for authors
- Something that may be interesting to see for future work is how the choice of order effects the performance of mTACE.
- How does the wall clock time compare between O(3) equivariant models and the several O(2) frames mTACE uses? What about space usage on a GPU?
3 comments
Q2: We measure gather → convolution → scatter: e3nn PyTorch CGTP, EQX PyTorch Linear–Gate–Linear, and OpenEquivariance fused CUDA CGTP.
Protocol. FP32, TF32 off, one RTX PRO 5000 72 GB GPU, with PyTorch 2.15.0, e3nn 0.6.0, OEQ 0.8.1. The sweep uses 16 channels, 1000 nodes, 40000 directed edges, , and all local orders. Nodes contain both spatial parities, while edge harmonics have natural parity. All allowed CGTP paths through output degree retain independent weights and separate outputs.
F/B/DB denote GPU forward, first backward and double-backward times. Wall columns include prerequisites: F alone, F plus full first backward (FB), or F plus the angular-input gradient and its VJP (FDB). The latter does not compute every first gradient, so FDB is not the sum of the three isolated columns. Double backward differentiates the angular-input gradient with respect to inputs and parameters, excluding geometry construction.
Medians follow two warmups and seven samples.
| Method | GPU F / B / DB (ms) ↓ | Wall F / FB / FDB (ms) ↓ | Peak F / FB / FDB (GB) ↓ | |
|---|---|---|---|---|
| 1 | e3nn | 1.52 / 2.28 / 4.28 | 1.54 / 4.11 / 8.36 | 0.163 / 0.222 / 0.316 |
| 1 | EQX O(2) | 1.40 / 2.25 / 7.86 | 1.42 / 4.01 / 13.34 | 0.102 / 0.226 / 0.419 |
| 1 | OEQ | 0.12 / 0.25 / 0.59 | 0.14 / 0.41 / 1.11 | 0.031 / 0.058 / 0.086 |
| 2 | e3nn | 4.39 / 12.54 / 13.74 | 4.42 / 17.28 / 22.69 | 0.659 / 1.059 / 1.488 |
| 2 | EQX O(2) | 1.94 / 2.53 / 9.99 | 1.96 / 4.81 / 16.54 | 0.245 / 0.525 / 0.924 |
| 2 | OEQ | 0.27 / 0.56 / 1.55 | 0.29 / 0.89 / 2.44 | 0.095 / 0.176 / 0.262 |
| 3 | e3nn | 12.83 / 62.88 / 56.90 | 12.86 / 76.28 / 84.94 | 1.911 / 3.108 / 4.378 |
| 3 | EQX O(2) | 3.79 / 5.20 / 12.47 | 3.82 / 9.09 / 20.26 | 0.452 / 0.951 / 1.643 |
| 3 | OEQ | 0.75 / 1.58 / 4.89 | 0.77 / 2.38 / 7.27 | 0.224 / 0.403 / 0.602 |
| 4 | e3nn | 32.72 / 214.04 / 191.91 | 32.76 / 247.25 / 269.84 | 4.592 / 7.352 / 10.271 |
| 4 | EQX O(2) | 6.29 / 9.12 / 21.57 | 6.32 / 15.48 / 34.90 | 0.727 / 1.510 / 2.609 |
| 4 | OEQ | 1.83 / 4.43 / 13.59 | 1.85 / 6.32 / 19.96 | 0.442 / 0.782 / 1.171 |
| 5 | e3nn | 73.68 / 586.99 / 528.58 | 73.73 / 660.37 / 712.25 | 9.462 / 14.840 / 20.711 |
| 5 | EQX O(2) | 9.29 / 13.94 / 32.82 | 9.32 / 23.29 / 52.97 | 1.071 / 2.209 / 3.820 |
| 5 | OEQ | 4.09 / 29.87 / 32.76 | 4.12 / 34.24 / 67.11 | 0.782 / 1.361 / 2.039 |
| 6 | e3nn | 150.98 / 1394.17 / 1264.22 | 151.04 / 1539.66 / 1647.28 | 17.498 / 27.133 / 37.767 |
| 6 | EQX O(2) | 12.85 / 19.82 / 46.57 | 12.88 / 32.76 / 75.07 | 1.489 / 3.056 / 5.279 |
| 6 | OEQ | 10.20 / 72.82 / 90.63 | 10.26 / 83.23 / 173.73 | 1.283 / 2.194 / 3.287 |
At , additional local operations introduce overhead, making double backward slower and more memory-intensive than e3nn. From , is faster and uses less memory than e3nn in all measured stages, with benefits increasing with . OEQ remains faster in forward and uses less memory, but EQX PyTorch has lower FB/FDB wall times at .
Together, these benchmarks complement the theoretical complexity analysis by showing how practical runtime and memory costs depend on angular degree, differentiation stage, and implementation. EquivariantX provides a reusable framework of symmetry-consistent operators and a foundation for further kernel optimization. For the complementary Wigner-(6j) route, matched full-model tests of W6j-mMACE with identical model weights demonstrate numerical agreement in energies and atomic and magnetic forces, alongside runtime and memory improvements (Reviewer 2, W2/Q2; Appendix I, Tables 4–5).
I thank the authors for their response. I am quite pleased with the updates and new information they provided.
We thank the reviewer for the positive assessment of our work, and for suggesting further analyses of model performance.
Q1: Magnetic models are computationally expensive because they must represent both structural and magnetic environments. The main-text mTACE models therefore use an angular cutoff of two throughout, Lmax=ℓmax=mmax=Lmag=2, as a practical accuracy-cost choice.
To examine how fitting accuracy changes with degree, we trained single-layer non-SOC CrN models with 64 channels, correlation order ν=3, complete O(3) and standard local O(2) Linear–Gate–Linear interactions. The configured node, spatial, local-order and magnetic cutoffs increase together, Lmax=ℓmax=mmax=Lmag=L, for L=1,2,3,4,5. The test results are:
Increasing degree from 1 to 3 substantially reduces all four errors. At L=4, energy and stress improve further, atomic-force accuracy remains almost the same, and magnetic-force RMSE increases. At L=5, all four errors reach their lowest values in this sweep. Higher angular resolution therefore improves the fitted model in this sweep, but not every target improves monotonically.