acceptodds
Under review as a conference paper at ICLR 2027

Through the Looking-Glass: Efficient Parity-Complete Learning via Local Frames

Abstract

Making the full range of symmetry-allowed tensor representations computationally accessible is a central challenge in equivariant atomistic learning. Clebsch-Gordan and Cartesian tensor products (CGTPs/CTPs) provide expressive constructions but become increasingly expensive as angular degree and the number of coupling paths grow, whereas existing approaches lack a unified treatment of natural- and unnatural-parity representations. Here, We introduce a general local framework that combines complete spatial-parity representations with efficient angular computation. We begin by developing a generalized Wigner- convolution that exactly recouples interactions involving additional node representations, replacing the first edge-level tensor-product intermediate with reusable node-level features. To address the angular coupling costs that remain in global tensor products, we then formulate computation in edge-aligned frames. By explicitly accounting for both rotations and reflections, we derive representation maps and a closed set of linear, tensor-product, and gating operations that guarantee global equivariance across both spatial parities. Incorporating time-reversal labels extends this framework to -equivariant learning. The resulting local convolutions retain angular scaling, compared with for dense CGTP. We demonstrate the framework through magnetic TACE (mTACE), with distinct interaction branches respecting coupled spatial–spin rotations with spin-orbit coupling (SOC) and independent rotations without it. For the collinear CrN and noncollinear Fe benchmarks, mTACE achieves substantially lower atomic- and magnetic-force errors than the reference models. We further provide EquivariantX, a library of local representation maps and equivariant operators that makes parity-complete local computation available as reusable building blocks for equivariant architectures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine Learning Interatomic Potential, Equivariant GNNs, AI 4 Science, Atomistic Simulations

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_4MWKOpenReview ↗
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
  1. Something that may be interesting to see for future work is how the choice of order ℓ\ell effects the performance of mTACE.
  2. 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

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=2L_{\max}=\ell_{\max}=m_{\max}=L_{\rm mag}=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\nu=3, complete O(3) and standard local O(2)O(2) Linear–Gate–Linear interactions. The configured node, spatial, local-order and magnetic cutoffs increase together, Lmax⁡=ℓmax⁡=mmax⁡=Lmag=LL_{\max}=\ell_{\max}=m_{\max}=L_{\rm mag}=L, for L=1,2,3,4,5L=1,2,3,4,5. The test results are:

LLParametersE RMSE (meV/atom) ↓F RMSE (meV/Å) ↓S RMSE (meV/ų) ↓1D magnetic F RMSE (meV/μB) ↓
1189,4781.3453.773.9721.91
2308,5820.8426.221.3313.00
3476,8380.4217.851.0110.94
4710,6300.3417.630.9711.88
51,018,1500.2615.150.6610.48

Increasing degree from 1 to 3 substantially reduces all four errors. At L=4L=4, energy and stress improve further, atomic-force accuracy remains almost the same, and magnetic-force RMSE increases. At L=5L=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.

Q2: We measure gather → convolution → scatter: e3nn PyTorch O(3)O(3) CGTP, EQX PyTorch O(2)O(2) 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, L=1,…,6L=1,\ldots,6, and all local orders. Nodes contain both spatial parities, while edge harmonics have natural parity. All allowed CGTP paths through output degree LL 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.

LLMethodGPU F / B / DB (ms) ↓Wall F / FB / FDB (ms) ↓Peak F / FB / FDB (GB) ↓
1e3nn1.52 / 2.28 / 4.281.54 / 4.11 / 8.360.163 / 0.222 / 0.316
1EQX O(2)1.40 / 2.25 / 7.861.42 / 4.01 / 13.340.102 / 0.226 / 0.419
1OEQ0.12 / 0.25 / 0.590.14 / 0.41 / 1.110.031 / 0.058 / 0.086
2e3nn4.39 / 12.54 / 13.744.42 / 17.28 / 22.690.659 / 1.059 / 1.488
2EQX O(2)1.94 / 2.53 / 9.991.96 / 4.81 / 16.540.245 / 0.525 / 0.924
2OEQ0.27 / 0.56 / 1.550.29 / 0.89 / 2.440.095 / 0.176 / 0.262
3e3nn12.83 / 62.88 / 56.9012.86 / 76.28 / 84.941.911 / 3.108 / 4.378
3EQX O(2)3.79 / 5.20 / 12.473.82 / 9.09 / 20.260.452 / 0.951 / 1.643
3OEQ0.75 / 1.58 / 4.890.77 / 2.38 / 7.270.224 / 0.403 / 0.602
4e3nn32.72 / 214.04 / 191.9132.76 / 247.25 / 269.844.592 / 7.352 / 10.271
4EQX O(2)6.29 / 9.12 / 21.576.32 / 15.48 / 34.900.727 / 1.510 / 2.609
4OEQ1.83 / 4.43 / 13.591.85 / 6.32 / 19.960.442 / 0.782 / 1.171
5e3nn73.68 / 586.99 / 528.5873.73 / 660.37 / 712.259.462 / 14.840 / 20.711
5EQX O(2)9.29 / 13.94 / 32.829.32 / 23.29 / 52.971.071 / 2.209 / 3.820
5OEQ4.09 / 29.87 / 32.764.12 / 34.24 / 67.110.782 / 1.361 / 2.039
6e3nn150.98 / 1394.17 / 1264.22151.04 / 1539.66 / 1647.2817.498 / 27.133 / 37.767
6EQX O(2)12.85 / 19.82 / 46.5712.88 / 32.76 / 75.071.489 / 3.056 / 5.279
6OEQ10.20 / 72.82 / 90.6310.26 / 83.23 / 173.731.283 / 2.194 / 3.287

At L=1L=1, additional local operations introduce overhead, making double backward slower and more memory-intensive than e3nn. From L≥2L\geq2, O(2)O(2) is faster and uses less memory than e3nn in all measured stages, with benefits increasing with LL. OEQ remains faster in forward and uses less memory, but EQX PyTorch has lower FB/FDB wall times at L=5,6L=5,6.

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.

Reviewer_8GKzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes two ways to handle spatial parity and time reversal in equivariant networks. One applies Wigner-6j recoupling before edge gathering, and the other uses local O(2)O(2) and O(3)×Z2TO(3)\times\mathbb{Z}_2^T operators with O(L3)O(L^3) angular cost. The authors build mTACE in EquivariantX and test it on collinear CrN and noncollinear Fe, though every run uses the local O(2)O(2) block and leaves Wigner-6j untested.

Critical strengths

The local representation gets a tricky symmetry detail right. Eq. (8) sets the m=0m=0 reflection character to q=p(−1)ℓq=p(-1)^\ell, and Eq. (7) places positive order components from different parities into one reflection convention. Sections 4.4 and 4.5 then mix polar, axial, time-even, and time-odd features without treating every local scalar as invariant.

That said, the raw errors in Table 2 are low. On CrN, atomic and magnetic force RMSEs reach 6.85 meV/Å and 3.34 meV/μB\mu_B, beating mMACE at 24 and 19. On non-SOC Fe, they hit 5.19 meV/Å and 1.05 meV/μB\mu_B versus 13.55 and 6.94 for HotPP-spin, and Figure 4 separates built-in invariance from augmentation while letting SOC energies vary under spin rotations.

Critical weaknesses
  1. Table 2 is not controlled, so I wasn't convinced the gains come from the local operators. Appendix H admits test splits may differ across studies, and the mMTP row lists training errors on all 2,423 structures from Section III.1 and Table 1 of Kotykhov et al. beside test metrics. Model sizes also differ: Table 3 compares a 1-layer mMACE on CrN (64×0e64\times0e features, magnetic degree 1) against a 6-layer mTACE at degree 2 that adds TACE product expansion and RRA without ablations or parameter counts.
  2. No experiments back up the speed or memory claims. Appendix F gives an asymptotic count for the local block while ignoring Eq. (16) edge features and TACE product expansion. Table 3 also leaves out runtime and memory use, and never scales past degree 2. Meanwhile, the Wigner-6j block is never run at all, as noted after Eq. (19), with no check on gradients or equivalence against two-stage tensor products and no memory benchmark.
  3. Key reproducibility and symmetry details are missing. There is no code link for EquivariantX, and Table 3 omits the magnetic radial basis size and hidden irreps, plus the split indices. More importantly, I couldn't see how the Eq. (19) RRA block stays equivariant. As the cited TECE paper shows in Section 3.8 and Eqs. (34)-(35), a nonzero phase in Re⁡[eiθ(r)Q‾K]\operatorname{Re}[e^{i\theta(r)}\overline{Q}K] breaks O(2)O(2) equivariance because reflection conjugates z=Q‾Kz=\overline{Q}K and flips sin⁡θ Im⁡z\sin\theta\,\operatorname{Im}z.
  4. The benchmarks barely test spin-orbit coupling, which alongside the unverified baselines pushes me toward a weak reject. CrN SOC relies on augmentation without time-reversal labels, and on Fe in Section 6, SOC barely moves force RMSEs from 5.19 to 5.10 and 1.05 to 1.06. Related work also misses E2Former-V2 in Section 4.3 and Appendix B, which already combined Wigner-6j recoupling with local SO(2)SO(2) ops and parity reindexing.
Questions for authors
  1. Can you rerun baselines on shared splits and mark Table 2 mMTP as training error, while stating if CrN magnetic forces are scalar or 3D?
  2. What are runtime and peak memory for mTACE versus fused CGTP kernels, and can you show equivalence and memory tests for Wigner-6j?
  3. What phase and parity rules in Eq. (19) keep RRA attention logits reflection-even and time-even across time sectors?
  4. With a fixed TACE backbone, what do explicit parity rules add in an ablation, and do orientation-dependent energies match DFT on a SOC benchmark?
6 comments

We thank the reviewer for the careful assessment of our symmetry framework and the detailed comments, which have helped us strengthen the validation and clarify the scope of our conclusions.

W1/Q1: Table 2 now explicitly labels the mMTP CrN results as training errors on all 2,423 structures, whereas the mTACE and mMACE results are test-set errors. All CrN magnetic-force errors are now explicitly marked as 1D in Table 2.

Regarding rerunning deeper baselines, mMACE introduces magnetic regular solid harmonics in both its interaction and product blocks. Their magnitude scales as ∥m∥ℓ\|\mathbf m\|^\ell, and repeated coupling through the many-body expansion can amplify feature variance and gradients. In our CrN training trials, mMACE exhibited unstable feature variance and exploding gradients at four or more message-passing layers. Increasing the depth of the published one-layer model is therefore not a straightforward controlled comparison. For Fe-DeepSpin, both mTACE and HotPP-spin use six layers.

To address the CrN depth directly, we trained an additional non-SOC mTACE with the same scalar node hidden representation 64×0e64\times0e as the published mMACE, one message-passing layer, spatial cutoff degree ℓmax⁡=3\ell_{\max}=3, magnetic degree Lmag=1L_{\rm mag}=1 and correlation order ν=3\nu=3. We do not usa data augmentation here.

ModelE RMSE (meV/atom) ↓F RMSE (meV/Å) ↓S RMSE (meV/ų) ↓Magnetic F RMSE (meV/μB) ↓
mMACE, one layer (published)1.21241.5619
mTACE, one layer (new)0.3917.531.0510.52

The complementary degree sweep is reported in Reviewer 3, Q1.

W2/Q2: Reviewer 3, Q2 compares standard/fused CGTP and local O(2)O(2) convolutions, excluding ACE and magnetic feature construction, since these operations are computationally inexpensive compared to convolution. Furthermore, Since these modules are transferable across different model architectures, we prioritize comparing the convolution operations. For mTACE, we construct the i,ji,j pairs on edges, leaving the product operation unchanged. In contrast, mMACE introduces an additional tensor product between node features and node magnetic solid harmonics.

Appendix I/Tables 4–5 compare MACE v0.3.17 with W6j-mMACE, preserving identical weights, products, radial networks and readouts. Settings: 64 channels, node Lmax⁡=1L_{\max}=1, edge ℓmax⁡=3\ell_{\max}=3, magnetic degree 3, two layers, correlation order 3. BCC Fe supercells run on an RTX PRO 5000 72 GB GPU, TF32 off. Time and MAE use the same 15 energy/atomic-force/magnetic-force calls after five warmups, excluding neighbor finding and model construction.

Mean ms/peak allocated GB. Gains: e3nn/CUDA.

PrecisionAtomse3nn (ms/GB) ↓W6j Torch (ms/GB) ↓W6j CUDA (ms/GB) ↓Speedup / memory reduction ↑
FP3212876.12 / 1.98170.97 / 3.0825.85 / 0.652.94 / 3.04
FP321024496.91 / 15.681475.79 / 24.5647.46 / 5.1210.47 / 3.06
FP3234561666.61 / 52.85OOM136.63 / 17.2212.20 / 3.07
FP6410241836.11 / 31.352886.20 / 49.15218.70 / 10.228.40 / 3.07

About numerical equivalence, at 1024 atoms, CUDA MAEs versus e3nn are (4.51×10−13,5.68×10−12,2.18×10−11)(E/F/Fmag)(4.51\times10^{-13},5.68\times10^{-12},2.18\times10^{-11})(E/F/F^{mag}) in FP64, respectively in eV/atom, eV/Å and eV/μB\mu_B. Unfused recoupling adds path intermediates and small kernels, explaining the slower Torch result. CUDA reuses node products and fuses gather–contraction–scatter without edge messages. Separate sparse replacements of the two CGTPs retain their edge intermediate and miss this reuse.

W3/Q3: We have uploaded the code for EquivariantX and the corresponding mTACE implementation to OpenReview. Regarding the hyperparameters, we have provided more detailed settings in Table 3 of the paper. As for the dataset splits, we will release the corresponding indices once the paper is made public, as the full lists of dataset indices are too extensive to be included in the paper itself.

RoPE uses complex phases to relate tokens i,ji,j. In RRA, distance already conditions the interaction, and radial bias matters more than a phase scale. mTACE instead uses θ(r)=0\theta(r)=0 (TECE Eq. (35)), or a real radial scale, preserving reflections.

W4/Q4: We have explicitly identified the symmetry of each model in Table 2 and clarify the scope of the SOC comparisons below.

A non-SOC model enforces independent rotations of the spin and lattice degrees of freedom, whereas the SOC formulation allows their relative orientation to affect the energy. The additional symmetry constraint of the non-SOC model can be useful when it matches the reference data, but neither formulation is expected to give lower prediction errors in every setting.

For the non-SOC CrN dataset, training the SOC model with augmentation illustrates approximate enforcement of the additional non-SOC symmetry; it does not constitute a test of physical SOC effects. Fe-DeepSpin was generated with SOC, but its authors report in Section 4.4 that the SOC effects are smaller than the model errors. Following HotPP-spin, we trained both SOC and non-SOC variants. Their similar errors are therefore consistent with the limited SOC signal in this benchmark and should not be interpreted as a quantitative validation of orientation-dependent SOC energies.

E2FormerV2's axis-aligned rearrangement differs fundamentally from our local O(2)O(2) framework. It exactly reformulates CGTP, already derived in eSCN rather than introduced by E2FormerV2. See Table below and Reviewer 1, W3.

OperationeSCNE2FormerV2
Align the harmonic to mf=0m_f=0Eqs. (8)–(11)Eqs. (11)–(12)
Only mi=±mom_i=\pm m_o can contributeProposition 3.1Definition 4.2, Eq. (13)
Even degree sum: diagonal couplingAppendix A.1, Eq. (28)Appendix C.1, Eqs. (29)–(30)
Odd degree sum: signed exchangeAppendix A.1, Eq. (28)Appendix C.1, Eqs. (29)–(30)
Odd degree sum: vanishing m=0m=0 couplingAppendix A.1, Eq. (29)Appendix C.1

I thank the authors for their rebuttal, which addresses some of my concerns.

I have a few follow-up comments. First, in abstract you mention "generalized Wigner 6j convolution" is still a bit confusing in my view when I first read this. By "generalized Wigner 6j convolution," I would expect something like a node-wise factorization extended to regular harmonics as opposed to solid harmonics. As I now understand it, however, the construction applies standard recoupling to an additional node input while retaining an edge-dependent contraction. This makes it more like a "specialized" Wigner 6j convolution whose contribution is principally in its computational organization and implementation. Also, since this operation reduces two edge-level TPs to one edge-level TP, would it be useful to further reduce the remaining edge-level TP to a node-level TP?

I see the most important contribution of the paper as the systematic local (O(2)) treatment of spatial and time parities. I would encourage the authors to revise the abstract and main presentation to highlight this contribution more explicitly and distinguish it from the established ideas of Wigner-6j recoupling and local (SO(2)) computation. The current presentation somewhat obscures where the main technical advance lies.

Reviewer_rmZ1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper combines e2former's node-based convolution and eSCN's SO2 convolution with additional parity injection. The authors then incorporate these into TACE as mTACE and benchmark over two magnetic datasets, and the results look promising.

Critical strengths

The paper is clean and easy to follow. The idea is a natural extension of e2former, and the result is overall good on the provided benchmarks.

Critical weaknesses
  1. Line 152: "the stabilizer" -> "the stabilizer w.r.t. O(3)"
  2. To demonstrate parity, I highly recommend the simple task of Tetris [1] [2]. If the parity is implemented correctly, the network should be able to classify all these shapes 100%.
  3. It would be clearer to derive the explicit formula using the e2former node-based convolution (eq. 28) with the O(2) frames, and derive the consequent time complexity O(L3)O(L^3) of this operation. The equivariance proof should also be based on this.
  4. The consequence of the method should be a reduction in the time complexity like eSCN, but there is no benchmarking over time. I would like to see the authors include the tensor product operation runtime compared to e2former-based convolution under different LL, batch size, and multiplicities, and the overall network runtime compared to other networks like TACE that the authors used.

[1] Xie, YuQing, et al. "The price of freedom: Exploring expressivity and runtime tradeoffs in equivariant tensor products." arXiv preprint arXiv:2506.13523 (2025).

[2] Liang, Chenxing, et al. "Spin-Weighted Spherical Harmonics Enable Complete and Scalable $\mathrm {E}(3) $-Equivariant Networks." arXiv preprint arXiv:2607.01408 (2026).

Questions for authors

See weaknesses.

2 comments

We thank the reviewer for recognizing the clarity of our presentation and the promising results. We address the concerns below.

W1: We have updated the description in the paper to “The stabilizer of the aligned direction with respect to O(3)O(3) is”, making the ambient group explicit.

W2: Section 5/Figure 3 in the main text now report the suggested 3D Tetris classification.

Following [1,2] and e3nn, four local O(2)O(2) layers predict even and odd scalars. Training uses seven shapes under proper rotations, withholding one enantiomer. Evaluation includes all eight under new rotations and reflections. We scan Lmax⁡=1,2,3,4L_{\max}=1,2,3,4 and mmax⁡=0,…,Lmax⁡m_{\max}=0,\ldots,L_{\max} with three seeds, totaling 42 runs. All 30 runs retaining positive orders achieve 100% accuracy, including the withheld enantiomer, with all seven rounded outputs correct. For this scalar-input architecture, mmax⁡=0m_{\max}=0 cannot distinguish the enantiomers, whereas mmax⁡≥1m_{\max}\geq1 suffices.

W3: Appendix J, “Relation to eSCN and E2FormerV2”, and Table 6 clarify the following distinctions (selection-rule correspondence: Reviewer 2, W4/Q4).

  1. Path-resolved aligned CGTP. eSCN derived the aligned CGTP identity but did not provide a complexity proof for path-resolved aligned CGTP. Its Appendix C establishes O(L3)O(L^3) complexity for the path-combined SO(2)SO(2) convolution, not for aligned CGTP retaining separate output paths. Aligned CGTP and sparse CGTP are mathematically equivalent implementations of the same CG tensor product, preserving independent path weights and outputs. Alignment leaves only mf=0m_f=0 in the edge harmonic, allowing sparse evaluation. Retaining O(L3)O(L^3) separate outputs and densely inverse-rotating each costs O(∑path(2ℓo+1)2)=O(L5)O(\sum_{\mathrm{path}}(2\ell_o+1)^2)=O(L^5), the same asymptotic angular complexity as sparse CGTP. CuEquivarianceand OpenEquivariance appeared later, so the historical speedup over e3nn does not establish an asymptotic advantage for path-resolved aligned CGTP over these sparse implementations. Constructing and applying Wigner-DD matrices also adds equivariance erros.

  2. eSCN SO(2) convolution. Combining paths into order-wise learned linear maps before inverse rotation yields O(L3)O(L^3) angular complexity. It differs from retaining every CGTP output path separately.

Our main contribution is standardizing O(2) theory and operator framework that consistently handles rotations, reflections and time reversal. It accommodates both natural- and unnatural-parity representations and time-even and time-odd features while preserving global (O(3)×Z2TO(3)\times\mathbb{Z}_2^{\mathcal T}) equivariance. EquivariantX implements these representation rules and symmetry-compatible operators in a reusable library, and mTACE applies them to magnetic interatomic potentials.

Wigner-6j6j recoupling is introduced and implemented as Pytorch and CUDA kernel here primarily to accelerate mMACE and is not used in the reported mTACE models. Meanwhile, aligned CGTP's usage combined with Wigner-6j6j has already been accomplished by eSCN and E2formerV2. As discussed in W3/1, sparsifying aligned CGTP does not provide additional computational benefits. Moreover, directly combining local O(2) convolution with Wigner-6j recoupling is not feasible, because Wigner-6j recoupling requires preserving each intermediate path separately, whereas the local SO(2)/O(2) approach merges these paths at an earlier stage. We therefore implement and evaluate these as two complementary routes.

W4: We measured forward/backward/double-backward times and peak memory for e3nn PyTorch CGTP, EquivariantX PyTorch local O(2)O(2) and OpenEquivariance fused CUDA CGTP. The timed operation is gather → convolution → scatter. TrimSO(2) and FlashSO(2) demonstrate the potential of efficient fused local convolutions in Reviewer 3, Q2. However, specialized fused kernels alone do not provide a shared, composable library of SO(2)SO(2)/O(2)O(2) operators with consistent representation and symmetry conventions. This is why we first implement an e3nn-style standard PyTorch library as a common foundation for community reuse and subsequent kernel development. Fused kernels are not the focus of this paper. The broader EQX project's spherical/Cartesian O(3)O(3)/O(2)O(2) kernels belong to separate work.

For Wigner-6j6j acceleration, please see Reviewer 2, W2/Q2 and Appendix I. EquivariantX provides a W6j convolution interface for mMACE, and we compare its inference time and peak GPU memory with the original implementation using unchanged parameters and non-interaction blocks.

Full-model timings also reflect the execution infrastructure. TACE supports AOTI and Sobek-style streaming convolutions that avoid materializing convolution weights, reducing temporary storage and memory traffic. The inspected mMACE and HotPP-spin releases do not provide an equivalent integrated execution path. Although MACE has efficient symmetric contraction, this optimizes the product block rather than eliminating convolution-side costs. TACE can therefore benefit from engineering advantages beyond the choice of equivariant operators. In Matbench Discovery's MD speed benchmark, for example, TACE-OAM-L uses L=5L=5 yet runs at nearly the same speed as MACE-MPA-0.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank the reviewers for their constructive comments. We have revised the manuscript and added experiments and implementation details in response to their concerns. We have also uploaded the EquivariantX code and the corresponding mTACE implementation as part of the Supplementary Materials. The main updates are summarized below.

  1. Clarified the theoretical formulation and relation to previous work. We expanded the discussion of eSCN and E2FormerV2, distinguishing path-resolved aligned tensor products from path-combined local O(2)O(2) convolutions and clarifying their computational complexity. We also clarified that Wigner-6j6j recoupling is implemented and evaluated for accelerating mMACE, whereas the reported mTACE models use local O(2)O(2) operators.

  2. Added explicit parity validation. We included the suggested 3D Tetris classification task across different angular cutoffs and local orders. All configurations retaining positive local orders achieved 100% accuracy, including correct classification of the enantiomer withheld from training.

  3. Added runtime, memory, and numerical-equivalence benchmarks. We measured forward, backward, and double-backward times and peak GPU memory for EquivariantX local O(2)O(2) operators, e3nn tensor products, and OpenEquivariance fused kernels. We also benchmarked Wigner-6j6j-accelerated mMACE with unchanged model weights and verified numerical agreement in energies, atomic forces, and magnetic forces.

  4. Expanded model comparisons and angular-degree analysis. We added a single-layer mTACE comparison with key settings matched to the published mMACE model and an angular-degree sweep from L=1L=1 to 55, reporting parameter counts and prediction errors. We corrected the labelling of mMTP training errors and explicitly identified the CrN magnetic-force errors as one-dimensional.

  5. Clarified symmetry and implementation details. We expanded the hyperparameter descriptions, explicitly identified the SOC and non-SOC model settings, and clarified the reflection-preserving choice used in the RRA block.

We hope these materials can provide a clearer understanding of the proposed framework.

Sincerely,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.