acceptodds
Under review as a conference paper at ICLR 2027

EACO: Predicting -Equivariant Quantum Operators via Edge-Aligned Convolution

Abstract

Predicting quantum operator matrices from atomic structures with group symmetries enables large-scale electronic-structure calculations and high-throughput materials screening. Focusing on this task, current edge-aligned convolutions guarantee global equivariance. Still, under reflections or spatial inversion, they cannot force particular operator components to vanish or protect electronic degeneracies. To overcome this issue, we propose Edge-Aligned Convolution (EACO), a message-passing layer that achieves local equivariance with strict locality and high computational efficiency. In particular, EACO assigns each irreducible representation (irrep) of angular degree and inversion parity the sign , indicating how it transforms under a reflection in the edge-aligned local frame and thus distinguishing the irreps with the same angular degree but opposite inversion parities. We prove that local equivariance is necessary and sufficient for global equivariance within this construction, so the model built on EACO preserves the transformation laws of the full group. Compared with SLEM, EACO reduces Hamiltonian MAE by up to 22.1% across five materials, with 17.2%29.6% fewer parameters. It also reduces equivariance errors under reflections and spatial inversion by three to four orders of magnitude. On graphene, EACO recovers the horizontal-mirror () selection rule for flexural phonons, suppressing forbidden electron–phonon coupling while accurately reproducing allowed coupling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Density Functional Theory, Local Graph Neural Network, Equivariant Neural Network

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_o3pkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper points out that edge-aligned SO(2)\mathrm{SO}(2) convolutions such as eSCN and SLEM lower the tensor product cost from O(ℓmax⁡6)\mathcal{O}(\ell_{\max}^6) to O(ℓmax⁡3)\mathcal{O}(\ell_{\max}^3) but cannot tell apart irreps that share the same degree ℓ\ell and differ in inversion parity pp. These models are therefore only SO(3)\mathrm{SO}(3) equivariant. The authors propose EACO, which restricts the kernel in the edge frame to the O(2)\mathrm{O}(2) stabilizer of the edge axis. Each irrep gets a reflection sign τℓ,p=p(−1)ℓ\tau_{\ell,p} = p(-1)^{\ell}, and fixed masks let channels with the same sign couple through I2I_2 and channels with opposite signs couple through J2J_2. Theorem 1 shows that local O(2)\mathrm{O}(2) equivariance is necessary and sufficient for global O(3)\mathrm{O}(3) equivariance in this construction, and Proposition B.3 shows that the masked kernel spans the parity allowed Clebsch-Gordan couplings. The method is built on SLEM and tested on Hamiltonian prediction for Si, GaN, HfO$_2$, graphene and MoS$_2$. Compared with SLEM it lowers the MAE by up to 22.1% with 17.2 to 29.6% fewer parameters, reduces the equivariance error under reflection and inversion by three to four orders of magnitude, and respects the σh\sigma_h selection rule for electron-phonon coupling in graphene.

Critical strengths
  1. The method gives exact O(3)\mathrm{O}(3) equivariance at the cost of an SO(2)\mathrm{SO}(2) convolution. EACO keeps the O(ℓmax⁡3)\mathcal{O}(\ell_{\max}^3) cost and strict locality of SLEM, needs fewer parameters, and reaches relative equivariance errors around 10−710^{-7}. This weighs in its favor because it shows that parity awareness can be added to the most efficient family of Hamiltonian models at no extra cost, and that it improves accuracy over the SO(2)\mathrm{SO}(2) baseline
  2. The graphene experiment connects the symmetry to physics. SLEM leaks coupling that the horizontal mirror forbids (0.129 meV for ZA and 0.324 meV for ZO), while EACO keeps it at numerical zero. This shows that the symmetry matters for real applications and not only for an equivariance metric.
Critical weaknesses
  1. The novelty is limited and the paper does not position itself against related work. The argument behind Theorem 1 is the standard stabilizer restriction from steerable kernel theory. The paper cites Weiler et al. (2018) for the SO(2)\mathrm{SO}(2) case but does not discuss the general constructions that already cover O(2)\mathrm{O}(2) and O(3)\mathrm{O}(3) kernels, such as Weiler & Cesa (2019) and Cesa, Lang & Weiler (2022). Theorem 1 is also close to an O(3)\mathrm{O}(3) version of the paper's own Theorem A.3. In addition, recent work has studied the closely related problem of keeping parity-sensitive interactions in fast tensor products. Xie et al. (ICML 2026) recover the interactions missed by the Gaunt tensor product using vector spherical harmonics, and Liang et al. (arXiv 2607.01408) recover antisymmetric paths and parity-odd components using spin-weighted spherical harmonics. These works target Gaunt type products rather than edge-aligned convolutions, but they address the same question of keeping the correct parity behavior at low cost. For these reasons I see the main contribution as extending parity awareness to edge-aligned SO(2)\mathrm{SO}(2) convolutions and showing that it helps Hamiltonian prediction, rather than the theory itself. I would ask the authors to discuss these works and adjust their claims.

  2. It is unclear where the gains over parity-aware baselines come from. DeepH-E3 and MACE-H are already O(3)\mathrm{O}(3) equivariant and also satisfy the mirror selection rule exactly in Table 3. Even so, the error of EACO on the allowed coupling is 2.6 to 3.9 times lower (0.100% compared with 0.260% and 0.280% for ZA, and 0.113% compared with 0.363% and 0.437% for ZO), and its overall MAE in Table 2 is lower on four of the five materials, with MACE-H better on MoS$_2$. Since all three models are parity aware, the parity constraint cannot explain these gains. The paper says they reflect "a more accurate learned Hamiltonian response" but gives no analysis. They could come from the SLEM backbone or from differences in training, such as 4 versus 3 layers, 1,000 versus up to 1,500 epochs, and different losses. The very large on-site errors of DeepH-E3 and MACE-H on GaN and HfO$_2$ (30 to 87 meV, compared with 0.3 to 2 meV for SLEM and EACO) also make me wonder whether these baselines were set up well. Without this analysis it is hard to tell how much of the improvement comes from EACO itself.

  3. The efficiency numbers are given at a degree that is rarely used in practice. The 9.94×9.94\times memory and 16.35×16.35\times time reductions in Section 4.3 (Fig. 5a) are measured at ℓmax⁡=11\ell_{\max} = 11. The paper's own experiments go up to degree 8 for the gg orbitals in HfO2 and use degrees 4 to 6 for the other materials, and ℓmax⁡≤6\ell_{\max} \le 6 is enough for bases up to ff orbitals, which covers most Hamiltonian datasets. Results at ℓmax⁡\ell_{\max} of 4, 6 and 8 would be more representative. The speedup over Clebsch-Gordan products also comes from SLEM, since EACO has the same cost.

  4. Reproducibility could be improved. Appendix D.1 gives the optimizer, learning rate, number of epochs, batch size and channel counts. It does not report seeds, the number of runs, training hardware or training time. The baselines are only described as using their "native" settings. A full configuration table for all methods and released code would make the results easier to verify.

Questions for authors
  1. How does EACO relate, both conceptually and empirically, to Xie et al., Liang et al., and the general O(2)\mathrm{O}(2) and O(3)\mathrm{O}(3) steerable kernel constructions? What does the edge-aligned O(2)\mathrm{O}(2) approach offer beyond them?

  2. Why does EACO outperform DeepH-E3 and MACE-H in Tables 2 and 3, given that both are also O(3)\mathrm{O}(3) equivariant? For the MoS$_2$ in Table 2, why the EACO is not following the same trend with other materials to give the best result? Is there any specific reason for MoS$_2$? Were the baselines (DeepH-E3 and MACE-H) tuned for these datasets, and why are their on-site errors on GaN and HfO$_2$ so large?

  3. What are the memory and time reductions at ℓmax⁡\ell_{\max} of 4, 6 and 8?

  4. How many runs are behind each reported number, and will the full training configurations for EACO and the baselines be released?

5 comments
AuthorsOpenReview ↗

Response to Reviewer o3pk (Part 1)

Thank you for your detailed and constructive review. We have revised our paper based on your comments, with new text in blue. Below, we answer your questions one by one.

W1 & Q1: Relation to steerable kernels, Xie et al., and Liang et al.

A1: Theorem 1 is the O(3) extension of the standard reference-direction construction (Theorem A.3), as stated in App. A (Lines 1091–1092). However, this construction only gives the condition for equivariance, not an operator that satisfies it. Our theoretical contribution is the complete solution of this condition inside the SO(2) convolution (Prop. 1 and Prop. B.3). It brings complete O(3)-equivariant convolution to the O(L3)O(L^3) cost of SO(2) convolution (L=ℓmax⁡L=\ell_{\max}), and the parity-aware edge-aligned convolution you identify as our main contribution is exactly its result.

1. Beyond steerable kernel theory. Weiler & Cesa [1] and Cesa et al. [2] characterize the O(2) and O(3) kernel spaces. In message passing, the O(3) space is realized by CG tensor products at O(L6)O(L^6) cost, or O(L5)O(L^5) with sparse coefficients; our theory gives its edge-aligned form at O(L3)O(L^3). An SO(2) convolution mixes each ±m\pm m pair by aI2+bJ2aI_2+bJ_2, where I2I_2 carries the CG paths with an even sum of input, filter, and output degrees ℓi+ℓf+ℓo\ell_i+\ell_f+\ell_o and J2J_2 the odd ones, and O(3) keeps a path iff (−1)ℓi+ℓf+ℓo=τinτout(-1)^{\ell_i+\ell_f+\ell_o}=\tau_{\rm in}\tau_{\rm out} with τ=p(−1)ℓ\tau=p(-1)^\ell. Hence, equal-τ\tau channels couple only through I2I_2 and opposite-τ\tau channels only through J2J_2 (Prop. 1), and these kernels span exactly the parity-allowed CG couplings (Prop. B.3). Attaching pp to SLEM channels without this solution either keeps forbidden couplings or loses allowed ones.

2. Different from Xie et al. [3] and Liang et al. [4] in two points. First, they complete Gaunt-type tensor products, which miss allowed parity-odd paths; EACO corrects edge-aligned convolutions, which admit forbidden paths. Second, their complete products cost O(L4log⁡2L)O(L^4\log^2 L), while EACO costs O(L3)O(L^3).

OperatorParity paths keptAsymptotic angular cost
CG tensor productallowedO(L6)O(L^6)
Gaunt TPallowed, even-sum onlyO(L3)O(L^3)
VSTP (Xie et al.), SpinGTP (Liang et al.)allowedO(L4log⁡2L)O(L^4\log^2 L)
SO(2) conv. (eSCN, SLEM)allowed + forbiddenO(L3)O(L^3)
EACOallowedO(L3)\mathbf{O(L^3)}

Both works note that these bounds bring no speedup at the degrees used in equivariant networks (Xie et al., Sec. 7; Liang et al., Sec. 5).

3. Empirically, EACO keeps SLEM's efficiency at L=4,6,8L=4,6,8, while CG and SpinGTP cost several times more for the same kernel space (See our response in A3). The released VSTP code provides only the pointwise product, not the complete product that recovers all CG paths, so we do not time it.

Following your suggestion, we have added this discussion to Sec. 2, the path-parity analysis and the table to App. B.5, and the contribution statement above to Sec. 1.

[1] M. Weiler and G. Cesa. General E(2)-Equivariant Steerable CNNs. NeurIPS, 2019.

[2] G. Cesa, L. Lang, and M. Weiler. A Program to Build E(N)-Equivariant Steerable CNNs. ICLR, 2022.

[3] Y. Xie et al. Asymptotically Fast Clebsch–Gordan Tensor Products with Vector Spherical Harmonics. ICML, 2026.

[4] C. Liang et al. Spin-Weighted Spherical Harmonics Enable Complete and Scalable E(3)-Equivariant Networks. arXiv:2607.01408, 2026.

AuthorsOpenReview ↗

Response to Reviewer o3pk (Part 2)

W2 & Q2: Sources of the gains over DeepH-E3 and MACE-H, MoS$_2$, and the baseline setup

A2: Tables 2 and 3 already separate the two sources. SLEM and EACO both beat DeepH-E3 and MACE-H on Si and graphene and both lose to MACE-H on MoS$_2$, and SLEM's allowed-EPC errors (0.22%) are already below the baselines' (0.26–0.44%), so the gains over the O(3) baselines come from the edge-aligned backbone. EACO's contribution is to make this backbone exactly O(3)-equivariant at the same O(L3)O(L^3) cost, which no SO(2) model provides.

1. EACO itself is measured against SLEM. The two share the backbone, data, loss, optimizer, and 1,000 epochs, and differ only in the edge kernel. EACO lowers the reflection and inversion errors by 3–4 orders of magnitude, keeps the mirror-forbidden coupling at numerical zero (SLEM leaks 0.129/0.324 meV), and uses 17–30% fewer parameters while matching or improving SLEM's accuracy, with clear gains on Si and HfO$_2$.

2. The training differences favor the baselines. They use their own losses and schedules (Table 9) for 1,500 epochs, 1.5×1.5\times ours. SLEM and EACO are strictly local, so their 4 layers see one cutoff, while the 3 baseline layers see three.

3. On GaN and HfO$_2$, the gains over the baselines come from the angular degrees these materials require.

  • The baselines cannot afford these degrees. At the required degrees and parities (ℓ=6\ell=6 for GaN, ℓ=8\ell=8 for HfO$_2$), DeepH-E3 runs out of memory on an 80-GB A100 for both materials, and MACE-H needs over 1 TB of host memory to build for HfO$_2$ and around 10 days for a single run on GaN.
  • Their large on-site errors come from their official angular degree, not from tuning. Both baselines use hidden degree ℓ=4\ell=4 with natural parity. Ga on-site blocks need degrees up to 6 with both parities and Hf blocks up to 8, as the MACE-H paper [5] itself states. Setting only these label components to zero gives on-site errors of 37.0 meV for Ga and 102.3 meV for Hf, 95–98% of the baselines' Ga and Hf errors (38.1–38.8 and 104.1 meV). Because the GaN and HfO$_2$ rows of Table 2 mixed this degree limit with model quality, we have moved these baseline results to App. D.4 and clarified this in Sec. 4.1.
  • This further highlights EACO's advantage: it makes O(3)-equivariant training at these degrees practical. EACO trains at ℓ=6\ell=6 and ℓ=8\ell=8 at O(L3)O(L^3) cost. It attains 0.126 and 0.326 meV on GaN and HfO$_2$, 17% below SLEM on HfO$_2$, and keeps the equivariance error below 2×10−52\times10^{-5} meV, while SLEM's is 0.10–0.24 meV.

4. On MoS$_2$, EACO stays comparable to MACE-H with far fewer parameters. MACE-H has 7.3×7.3\times EACO's parameters (2.79M vs. 0.38M). However, EACO's overall MAE is only 0.032 meV higher (0.441 vs. 0.409 meV), its on-site MAE is 3.0×3.0\times lower (0.327 vs. 0.966 meV), and it outperforms DeepH-E3 (0.576 meV) with 3.5×3.5\times fewer parameters.

We have added this decomposition to Sec. 4.1, and the cost and on-site analyses to App. D.4.

[5] C. Qian et al. Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing. npj Comput. Mater. 12, 169, 2026.

AuthorsOpenReview ↗

Response to Reviewer o3pk (Part 3)

W3 & Q3: Efficiency at ℓmax⁡=4,6,8\ell_{\max}=4,6,8

A3: The speedup indeed comes from the edge-aligned SO(2) structure. However, SO(2) convolutions are not O(3)-equivariant, so an O(3)-equivariant model had to pay the cost of CG tensor products or complete fast products. EACO brings exact O(3) equivariance to the cost of SO(2) convolution. The table compares the four operators on the same workload as Fig. 5(a): 4,096 edges, 16 channels per degree and parity, forward and backward on one idle A100. EACO, CG, and SpinGTP (official code) use the same CG paths with independent 16×1616\times16 weights, except that SpinGTP lacks 4 paths at L≥6L\ge 6 where its spin-weighted Gaunt factor vanishes, so its numbers are lower bounds. SLEM is the SO(2) reference.

LLEACOSLEMCGSpinGTP
418.0 / 16513.9 / 14194.4 / 676145.9 / 1,796
631.0 / 32625.9 / 279269.3 / 2,353783.3 / 7,197
848.9 / 65342.6 / 653584.2 / 6,0723,135.9 / 21,878

(Time in ms / peak memory in MiB; medians of three runs.)

1. EACO is on par with SLEM (slightly slower and larger at small LL, as grouping channels by parity adds index operations, although EACO needs only half of SLEM's multiplications).

2. CG and SpinGTP cost far more for the same kernel space. CG takes 5.2/8.7/12.0×5.2/8.7/12.0\times more time and 4.1/7.2/9.3×4.1/7.2/9.3\times more memory than EACO at L=4,6,8L=4,6,8, and SpinGTP takes 8/25/64×8/25/64\times more time and 11/22/34×11/22/34\times more memory.

3. SpinGTP is slower than CG at these degrees. Its O(L4log⁡2L)O(L^4\log^2 L) bound needs fast spherical transforms, which bring no speedup at these degrees (Liang et al., Sec. 5). For a fully connected product, its code instead contracts precomputed coupling coefficients into a dense kernel for every edge, and its spin-1 features are complex, with about twice the real dimension.

These degrees are those of our experiments: L=4L=4 for Si, MoS$_2$, and graphene, and L=6L=6 and L=8L=8 for GaN and HfO$_2$, where DeepH-E3 runs out of memory and MACE-H needs over 1 TB of host memory or around 10 days per run (See our response in A2).

Following your suggestion, we have replaced Fig. 5(a) with ℓmax⁡=1\ell_{\max}=1–88, added these results to Sec. 4.3, and added the setup with the table to App. D.3.

W4 & Q4: Runs, configurations, and code

A4: Each number in Tables 2 and 3 comes from a single run with a fixed seed shared by the four methods. We have added a second run of SLEM and EACO with seed 42 (App. D.5). EACO's exact O(3) equivariance and 17–30% fewer parameters hold in every run by construction, and its accuracy gains on Si and HfO$_2$ hold in both runs, by 12–22%. On GaN, graphene, and MoS$_2$, the two methods are on par. Following your suggestion, Table 9 in App. D.1 now lists the full training configurations of all methods, including the seeds and hardware, and Table 12 in App. D.3 reports the training time of every run. We will release our code together with these configurations.

In summary, we hope the above responses can resolve your concerns and help you re-evaluate our work. We would appreciate it if you consider raising your score based on our reply. Thanks in advance for your consideration.

Reviewer_kG3wOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper extends PAssaro and Zitnick's fast edge aligned CG-product trick from SO(3)-equivariance to O(3)-equivariance. The specific domain that the authors look at is learning Hamiltonian matrices for molecular systems (or crystals) as opposed to running DFT. Since the underlying physics in this case is indeed O(3)-equivariant, O(3) equivariant layers are in principle the correct architecture to use and, as expected, they do improve performance by some amount.

The subtler question is how much of an improvement we really expect from this change given that O(3)/SO(3) is just a two-element set, so in general O(3)-equivariance is easy to "fake" just by augmenting each training data point with its mirror image. Granted, this wastes some amount of compute and O(3)-equivariance would only be satisfied up to the model's approximation error.

Critical strengths
  • The basic idea is clean, it is adequately described (modulo the general tendency to overcomplicate things a little in this field) and the recipe it proposes is practical.
  • I agree with the general philosophy that when we know the underlying symmetry of a problem our algorithms should respect it rather than trying to fudge it by e.g. data augmentation.
  • Having said the above, in this particular case, augmentation would just amount to adding the mirror image of each data point, so the bar is a little higher for showing an advantage. At a simplistic level, the proposed O(3)-equivariant layers amount to masking interactions that would violate mirror symmetry, which is good, because it suggests that regardless of the ultimate effect on accuracy, being equivariant to O(3) rather than SO(3) might save compute.
  • The graphene case is a neat example where exact mirror symmetry becomes physically important.
  • The EACO-vs-SLEM comparison is controlled in architecture and recipe.
Critical weaknesses

The following are potential ways that the paper could be improved but I wouldn't say that any of them are critical weaknesses:

  • Comparison to previously reported the results on the same systems are a bit murky. The original SLEM paper reports results on the same physical systems using 4.5M model and a 0.7M. Instead of comparing to these published results, the current paper re-runs a version of SLEM which is somewhere in between. This is good from the point of view that the re-run SLEM follows the same recipe as EACO so they are easy to compare head-to-head. On the other hand, it is not clear what the relationship is to the previously published SLEM results is: the re-run SLEM is better than the 0.7M on some systems but worse on others? So which one should really be the true baseline? And why are the results from the 4.5M SLEM not listed in the table at all? Given that it is a much larger model it would not be fair to compare it one-to-one to EACO but I feel the numbers should still be reported.

  • The original DeepH-E3 paper reports results on the same systems, but once again the present paper only reports their own retrained results, which are sometimes better sometimes worse than the those in the original. It would be good to get some clarity on why this is and what version of these models is the most fair to compare to EACO.

  • In the case of DeepH-E3 and MACE-H there is a huge difference between on-site and off-site errors on GaN and HfO_2, potentially suggesting a setup problem with the model making it harder to believe the results. Please discuss.

  • If possible, please also add SLEM-with-agumentation results to Table 2.

  • Something that is not specific to just this paper is that the O(\ell^6) vs O(\ell^3) comparison of the full CG-product vs the edge-aligned one seems a bit unfair. Firstly, by using specialized kernels that take advantage of the m=m_1+m_2 contraint, the \ell^6 constraint immediately drops to \ell^5 and several packages explicitly do this. Second, edge convolution is really a very special case that only applies to multiplying by a zonal vector, like the spherical harmonics. It is much less expressive operation than a full CG-product, so the real question is whether a network that only ever multiplies O(3)-equivariant activations with spherical harmonics sufficiently expressive for the given task.

  • Computational speed is almost perfectly matched with SLEM, further suggesting that EACO is more of a tweak on an existing models than a truly novel architectures. (This is more of a comment than a a criticism.)

  • The wording "reduces Hamiltonian MAE by up to 22.1% across five materials, with 17.2%–29.6% fewer parameters" seems to be overstating the authors' case a little. "reduces MAE by 0.5–22% (median about 14%) relative to a retrained SLEM, with 17–30% fewer parameters" might be more accurate. Again this is not a criticism of the method, just a reflection of the fact that SO(3) and O(3) are so closely related that enforcing reflection will only buy us so much.

  • The intro says "leaving reflections and spatial inversion unconstrained". These are the same thing, right? If you add reflections to SO(3) you automatically get inversion and vice versa.

  • The wording "atomic features as irreducible representations" is imprecise. The irreducible representations are either operator valued functions on the group or subspaces of functions on the group or its quotient spaces . It would be more accurate to say "atomic features are expressed in the bases of the group's irreducible representations".

  • There is a recent preprint title "Through the Looking-Glass: Efficient Parity-Complete Learning via Local O(2) Frames" that seems to suggest the same construction. There are no quantum chemistry experiments and the preprint is at best concurrent with the present paper, so it does not detract from EACO, but I feel it should be cited prominently.

Questions for authors

See above for points that the authors may wish to clarify. My rating is not contingent on any specific unresolved question.

1 comment
AuthorsOpenReview ↗

Response to Reviewer kG3w

Thank you very much for your appreciation of our work. We are also grateful that you still took the time to offer many sincere suggestions, although you considered none of them critical. We have revised our paper based on your comments, with new text in blue. Below, we answer your comments one by one.

W1 & W2: Published results of SLEM and DeepH-E3

A1: The published results come from models of other sizes and training recipes. Table 2 trains all methods on the same splits with the same checkpoint rule, and SLEM shares EACO's recipe and width, so their comparison isolates the edge kernel. This SLEM is a competitive baseline: it is more accurate than the published 0.7M and 1.7M SLEM models on Si, graphene, and GaN (Zhouyin et al. [1]). The 4.5M models are 88–13×13\times the size of EACO and test capacity rather than the kernel.

[1] Z. Zhouyin et al. Learning Local Equivariant Representations for Quantum Operators. ICLR, 2025.

W3: On-site errors of DeepH-E3 and MACE-H on GaN and HfO$_2$

A2: These errors come from the angular degree ℓ=4\ell=4 of the two baselines, which cannot represent the on-site blocks of Ga and Hf. These blocks need degrees up to 6 and 8, respectively, as the MACE-H paper [2] also notes, and zeroing only these label components already gives 95–98% of the baselines' Ga and Hf on-site errors. At these degrees, DeepH-E3 runs out of memory on an 80-GB A100, as Zhouyin et al. also report for HfO$_2$, and MACE-H needs over 1 TB of host memory for HfO$_2$ or around 10 days per run on GaN. We have therefore moved these results to App. D.4, and Table 2 compares SLEM and EACO on these materials, where EACO trains at degrees 6 and 8 at O(ℓ3)O(\ell^3) cost.

[2] C. Qian et al. Equivariant Electronic Hamiltonian Prediction with Many-Body Message Passing. npj Comput. Mater. 12, 169, 2026.

W4: SLEM with augmentation

A3: Following your suggestion, we report SLEM with inversion augmentation on Si in the format of Table 2 (Table 18 in App. D.8). We keep it outside Table 2, which compares architectures on the same data, because augmentation is a training strategy for any SO(3) model. As you anticipated, augmentation enforces O(3) equivariance only up to the approximation error, and this error matters. It lowers SLEM's MAE, but the improper residuals stay about three orders of magnitude above those of EACO, and on graphene they still open the mirror-forbidden channel in the resistivity of App. D.7.

W5 & W6: Cost and expressiveness of edge-aligned convolutions

A4: Sparse CG coefficients indeed reduce a full CG product to O(ℓ5)O(\ell^5), and we have added this count to App. B.8. Edge alignment applies to the product of a feature with the spherical harmonics of the edge direction, which is how messages are built in Eq. (3), and for this product EACO loses no coupling. Prop. B.3 shows that its kernels span exactly the parity-allowed CG couplings, at O(ℓ3)O(\ell^3) cost. Products of two learned features are a different operation, listed as future work in Sec. 5. Matching SLEM's speed is our design goal: as you note, EACO masks the mirror-violating interactions, and Prop. 1 shows that these masks are the complete solution, so EACO removes 17–30% of the parameters at SLEM's cost.

W7, W8 & W9: Wording

A5: Following your suggestion, we now state that, with 17.2–29.6% fewer parameters than SLEM, EACO matches or improves its Hamiltonian accuracy across five materials, with clear gains on Si and HfO$_2$ (Abstract, Sec. 1, Sec. 4.1). A second run confirms the gains on Si and HfO$_2$ (App. D.5). Yes, with rotations, reflections and inversion generate each other; Sec. 1 now reads "leaving reflections, and hence inversion, unconstrained" and "express atomic features in the bases of irreducible representations".

W10: The Looking-Glass preprint

A6: Thank you for pointing out this preprint (Xu & Xie [3]), which we now cite in Sec. 2. It builds local O(2) frames for magnetic interatomic potentials with time-reversal symmetry, whereas EACO targets quantum operators: it gives all O(2)-equivariant edge kernels in closed form, proves that they span exactly the parity-allowed CG couplings, and is tested on Hamiltonians and electron–phonon selection rules.

[3] Z. Xu and W. Xie. Through the Looking-Glass: Efficient Parity-Complete Learning via Local O(2) Frames. arXiv:2608.16592, 2026.

Thank you again for your appreciation and support of our work, and for your valuable suggestions, which have made the paper clearer and more complete.

Reviewer_3qE3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work describes a type of MPNN for predicting electronic structure that is equivariant under O(3) rather than the (previously studied) group SO(3), i.e. it includes the constraints arising from parity. At a technical level this is done by improving upon a previously implemented approach that uses SO(2)-equivariant convolutions around the axis of each molecular bond. The new approach of this work improves on the representation theory of this equivariance, promoting it to O(2), which is shown to result it a full model that is O(3) equivariant. The work shows empirical benefits, improved equivariance error, and a better capture of physics which depends on parity structure (e.g. various matrix elements that should vanish in Section 4.2 actually do vanish).

Critical strengths

This is a physically very sensible thing to do -- a correct description of the physics should involve treating parity correctly, and it is worthwhile to improve upon the work called SLEM (i.e. Zhouyin et. al 2025) in this way. There are empirical benefits shown in Table 2, but perhaps more importantly Section 4.2 shows that there are genuine physical effects that are not captured without this equivariance hard-coded in.

Critical weaknesses
  1. An overall weakness is that this is a relatively simple and straightforward improvement over SLEM.
  2. The empirical benefits shown in Table 2 may need a little bit of explanation, as some of them do seem kind of small and it would be good to have some idea of the uncertainty in the numbers.
Questions for authors
  1. Question: what is the source of the equivariance error of O(10^{-5}) etc. in Table 2? Am I right that this should be strictly zero mathematically, and that it is coming purely from machine precision? Could the authors then explain this 10^{-5} from first principles (i.e. the precision of the datatype used etc.)
  2. Question: some of the benefits shown in Table 2 are kind of small (e.g. 0.126 vs 0.132 etc. for the overall MAE in GaN, etc.). Is there an understanding of the uncertainty in these numbers (e.g. the seed-to-seed variation) so that we can understand whether to take this benefit seriously?
  3. Re: comparison with the literature -- its curious that p3 of the SLEM paper by Zhouyin et. al actually claims that their model is already O(3) equivariant, i.e. "Here we consider O(3) group, then x, y can be composed..." -- am I right that the current authors believe that this is actually a notational slip and that work is only SO(3) equivariant, with its O(3) extension being done here?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 3qE3 (Part 1)

Thanks for your positive and constructive comments. We have revised our paper based on your comments, adding more explanations and experiments, with new text in blue. Below, we answer your questions one by one.

W1: Contribution beyond SLEM

A1: The final rule is indeed simple: each pair of channels keeps either I2I_2 or J2J_2, chosen by their reflection signs τ=p(−1)ℓ\tau=p(-1)^\ell. However, this simplicity is the result of the analysis, not its starting point. EACO is the complete solution of the O(2) constraint inside the SO(2) convolution, so it brings exact O(3) equivariance at the cost of SO(2) convolution.

1. The rule is derived, not chosen. Theorem 1 reduces global O(3) equivariance to local O(2) equivariance. Prop. 1 shows that these masks give all O(2)-equivariant kernels, and Prop. B.3 shows that they span exactly the parity-allowed Clebsch–Gordan couplings. Attaching parity labels to SLEM channels without this solution either keeps forbidden couplings or loses allowed ones.

2. The missing constraint was not obvious. SLEM already labels its features with O(3) irreps, yet its convolution enforces only SO(3) (See our response in A4).

3. The constraint matters physically and keeps the complexity of SLEM. It removes the mirror-forbidden electron–phonon coupling in graphene, which SLEM leaks at 0.129 and 0.324 meV (Sec. 4.2), and it uses 17–30% fewer parameters with the same O(ℓmax⁡3)O(\ell_{\max}^3) angular complexity.

W2 & Q2: Uncertainty of the accuracy gains

A2: Some differences in Table 2 are indeed small, such as 0.126 vs. 0.132 meV on GaN. However, the main effects of EACO are exact symmetry and fewer parameters at the same O(ℓmax⁡3)O(\ell_{\max}^3) complexity, and they follow from the construction rather than from a particular training run.

1. A second run separates the robust gains. We have trained SLEM and EACO again with another seed (App. D.5). EACO is more accurate on Si and HfO$_2$ in both runs, by 12–22%. On GaN, graphene, and MoS$_2$, the two methods are on par; on GaN, the difference is −4.8%-4.8\% and +2.6%+2.6\% in the two runs.

2. We describe the accuracy accordingly. With 17–30% fewer parameters, EACO matches or improves SLEM's accuracy, with clear gains on Si and HfO$_2$ (Abstract, Sec. 1, Sec. 4.1).

3. The symmetry gains are exact. EACO's improper residual is float32 roundoff (See our response in A3), and its mirror-forbidden coupling is zero, whereas SLEM's improper residuals are 10−410^{-4}–10−310^{-3}.

AuthorsOpenReview ↗

Response to Reviewer 3qE3 (Part 2)

Q1: Source of the equivariance errors in Table 2

A3: Yes. The residuals of EACO are zero in exact arithmetic and come only from float32 roundoff. In float64, EACO's residuals under inversion and mirror reflections fall to the float64 roundoff level, whereas SLEM's stay unchanged.

1. First-principles estimate. Float32 has unit roundoff u=2−24≈6.0×10−8u=2^{-24}\approx6.0\times10^{-8}. Under inversion and mirror reflections, the maximum relative residual of EACO is 0.80.8–1.5×10−71.5\times10^{-7}, i.e., one to three times uu. The Hamiltonian matrix elements are of order 0.1 eV, so this relative error gives absolute residuals ΔimpH\Delta_{\mathrm{imp}}H of order 10−510^{-5} meV, as in Table 2.

2. Float64 check. We re-evaluate the same checkpoints in float64 (App. D.6). The maximum relative residual under inversion and the three mirrors is:

SiMoS$_2$GrapheneGaNHfO$_2$
EACO, float328.9×10−88.9\times10^{-8}1.4×10−71.4\times10^{-7}1.5×10−71.5\times10^{-7}1.3×10−71.3\times10^{-7}7.9×10−87.9\times10^{-8}
EACO, float642.6×10−16\mathbf{2.6\times10^{-16}}3.0×10−16\mathbf{3.0\times10^{-16}}3.2×10−16\mathbf{3.2\times10^{-16}}2.6×10−16\mathbf{2.6\times10^{-16}}1.7×10−16\mathbf{1.7\times10^{-16}}
SLEM, float323.0×10−43.0\times10^{-4}8.1×10−48.1\times10^{-4}4.9×10−44.9\times10^{-4}6.9×10−46.9\times10^{-4}7.1×10−47.1\times10^{-4}
SLEM, float643.0×10−43.0\times10^{-4}8.1×10−48.1\times10^{-4}4.9×10−44.9\times10^{-4}6.9×10−46.9\times10^{-4}7.1×10−47.1\times10^{-4}

EACO falls to a few times the float64 unit roundoff (1.1×10−161.1\times10^{-16}), so its residual in Table 2 is purely numerical. SLEM does not change with precision, so its residual is a genuine violation of O(3) equivariance.

Q3: SLEM and O(3) equivariance

A4: Yes. SLEM labels its features with O(3) irreps and parity (Zhouyin et al. [1], Sec. 2), but its SO(2) convolution learns both the I2I_2 and the J2J_2 coefficients for every channel pair, regardless of parity. SLEM is therefore exactly SO(3)-equivariant but not O(3)-equivariant, and EACO adds the missing reflection constraint. The numbers confirm this: SLEM's improper residuals are 10−410^{-4}–10−310^{-3} in Table 2 and do not decrease in float64 (See our response in A3), while its residuals under proper rotations stay at the float32 roundoff, as for EACO. We have added this clarification to App. A.5.

[1] Z. Zhouyin et al. Learning Local Equivariant Representations for Quantum Operators. ICLR, 2025.

In summary, we hope the above response resolves your concerns and makes you more confident in raising your score in the following discussion and decision phases. Thanks in advance for your consideration.

Thank you very much for addressing my points. I am happy with the answers to most of the questions, in particular regarding the scaling of the errors with machine precision, the situating in the previous literature. I also appreciate the extra run that was done to estimate the uncertainty in the comparison with SLEM.

My biggest weakness however was really that this -- though a very reasonable thing to do -- is perhaps not a conceptually very novel approach. The authors response to this has not really changed my mind on this, so under these circumstances I believe that a weak accept is appropriate and I will maintain my score.

AuthorsOpenReview ↗

Thank you for your Comment

Thank you very much for your time and for your positive assessment. Your feedback has been very valuable in improving our paper.

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

This paper proposes Edge-Aligned O(2) Convolution (EACO) for equivariant prediction of quantum operator matrices. Edge-aligned SO(2) convolutions constrain rotations around an aligned edge, but this constraint alone does not guarantee equivariance under reflections or inversion. EACO adds the corresponding local reflection constraints while retaining the computational structure of edge-aligned operations.

The construction assigns each O(3) representation a local reflection sign determined by its angular degree and inversion parity. This sign determines the permitted coefficients in the local convolution. Theorem 1 establishes that, within the stated edge-aligned kernel construction, local O(2) equivariance is necessary and sufficient for global O(3) equivariance. The appendices characterize the complete permitted kernel space, establish its correspondence with parity-allowed Clebsch–Gordan couplings, and prove equivariance of the complete network.

The model follows SLEM's strictly localized architecture and replaces its angular operations with EACO. Experiments evaluate Hamiltonian prediction on five materials, equivariance under improper transformations, mirror-symmetry selection rules for electron–phonon coupling in graphene, inversion augmentation, and angular-operator efficiency.

Overall, I lean toward acceptance. The method addresses a precise limitation of edge-aligned convolutions, the theoretical development is explicit, and the implementation preserves their computational advantages. My main reservation is that the application studies remain preliminary relative to the broader physical motivation.

Critical strengths

Strength 1: The paper provides a clear and complete construction for incorporating reflection constraints into efficient edge-aligned convolutions.

The distinction between SO(3) and O(3) equivariance matters for operator prediction because reflections and inversion can impose exact selection rules. The paper identifies the missing local constraint and gives an explicit parameterization that enforces it.

The theoretical contribution goes beyond showing that a particular mask is sufficient. Proposition 1 characterizes the full space of permitted local linear kernels, including the zero-order block and admissible biases. Proposition B.3 shows that this construction retains the complete space of parity-allowed CG kernels through an invertible change of basis. The reduction in parameters therefore removes symmetry-incompatible coefficients without removing permitted angular couplings.

The complete-network analysis is also useful. Appendices B.6 and B.7 explain how initialization, nonlinearities, invariant pair features, normalization, residual updates, and the final readout preserve the required transformation laws. Together, these results make the proposed layer well specified and potentially reusable.

Strength 2: The experiments support the intended symmetry and efficiency properties, with a relatively controlled comparison against SLEM.

SLEM and EACO share the architecture, channel widths, loss, and training duration, with the local convolution as the stated difference. This provides a useful comparison for assessing the proposed constraint. Table 2 reports lower overall Hamiltonian MAE than SLEM on all five materials, fewer parameters, and substantially smaller residuals under improper transformations.

The graphene study also usefully distinguishes forbidden and allowed couplings. Suppressing a forbidden coupling checks symmetry preservation, whereas accurately predicting an allowed coupling requires a useful learned Hamiltonian response. EACO performs well on both measures.

The angular-operator benchmark is documented in detail, including hardware, precision, warm-up, synchronization, and memory measurement. It includes frame construction and feature rotations, supporting the claim that the additional reflection constraint retains the low computational cost of edge-aligned operations.

Critical weaknesses

Weakness 1: The application studies remain preliminary and do not yet establish the practical physical benefit of the proposed symmetry constraint.

The paper provides a clear theoretical construction, but the application evidence is less developed. The five-material benchmark primarily measures Hamiltonian matrix-element errors. These results demonstrate predictive performance, but provide limited evidence about how the improvements affect physically important observables.

The graphene experiment is a useful illustration of the mirror selection rule. However, the vanishing forbidden coupling follows from the imposed symmetry and is also achieved by the existing O(3)-equivariant baselines. It therefore validates the implementation more directly than it establishes a distinctive application advantage.

The allowed-coupling results are more informative, but are obtained in a restricted shared-reference setting. Appendix D.2 holds DFT electronic states, energies, overlap derivatives, and phonon data fixed, replacing only the Hamiltonian derivative with the model prediction. This is a reasonable way to isolate the learned response, but leaves open how the improvement translates to a complete prediction workflow. The study also considers one initial electronic wavevector and two out-of-plane phonon modes.

My main reservation is therefore the gap between the physical motivation and the depth of application validation. A more demanding case in which symmetry-breaking prediction errors measurably affect a physical observable would strengthen the paper considerably. Such a comparison should include both SO(3)- and O(3)-equivariant baselines, so that the benefits of exact symmetry can be distinguished from the accuracy and efficiency advantages of EACO's implementation.

I view this as a limitation of the demonstrated application scope rather than a reason to reject the methodological contribution.

Weakness 2: The evaluation protocol and training variability need clarification, particularly for the smaller accuracy gains.

The bulk datasets are described as containing 150 training and 30 validation structures, without a separate test split. Checkpoints are selected using validation loss, and Figure 2 explicitly reports validation-set results. Graphene and MoS2 have separate test splits, but the evaluation split used for each Table 2 result should be stated explicitly.

The paper also does not report multiple training seeds or uncertainty estimates for the prediction results. The three independent runs in the operator benchmark measure benchmark variability rather than variability from training. This distinction matters for small differences: on MoS2, overall MAE changes from 0.432 meV for SLEM to 0.430 meV for EACO, while MACE-H achieves 0.408 meV. EACO's on-site errors are also higher than SLEM's on both MoS2 and graphene.

The inversion-augmentation result deserves similarly careful interpretation. On Si, augmentation reduces SLEM's MAE from 0.069 to 0.056 meV, close to EACO's 0.054 meV. This experiment clearly shows that the tested augmentation protocol does not enforce exact equivariance, but the remaining accuracy advantage needs variability estimates.

These qualifications do not weaken the theoretical guarantee. They do limit how broadly the accuracy improvements should be described.

Questions for authors

Clarification requests (answers could change my assessment).

  1. Which evaluation split is used for each result in Table 2? For the bulk datasets, are the reported errors measured on the validation structures used for checkpoint selection? Which monolayer results use the held-out test sets?

  2. Can you report repeated training runs for SLEM and EACO, especially for MoS2 and the comparison with inversion-augmented SLEM on Si? Are the smaller accuracy differences reproducible across seeds?

  3. How sensitive are the allowed EPC errors to the finite-difference displacement amplitudes? Please also clarify the relationship between the EPC reference structures, their displaced variants, and the training data.

  4. Beyond the graphene selection-rule example, which application would most clearly demonstrate a practical advantage over both the SO(3)-equivariant SLEM baseline and the existing O(3)-equivariant baselines? Even a focused additional result connecting Hamiltonian errors to an observable would help establish the method's application value.

Suggestions for revision (not conditions for acceptance).

  1. State the contribution primarily as a complete and efficient reflection-constrained parameterization for edge-aligned operator prediction. Keep the necessity-and-sufficiency claim explicitly tied to the kernel construction covered by Theorem 1.

  2. Distinguish symmetry verification from application validation. The forbidden graphene coupling is a convincing check of the constraint; broader physical benefits require evidence beyond recovering a symmetry-enforced zero. State prominently that the EPC comparison uses shared DFT states, overlap derivatives, and phonon data.

  3. Qualify the accuracy claims by distinguishing overall, on-site, and off-site errors. EACO improves overall MAE over SLEM on all five materials, but does not improve every component and is not the most accurate model on MoS2.

  4. Separate angular-operator efficiency from complete-model performance. End-to-end measurements at representative Table 2 configurations would complement the operator benchmark. Epoch-based convergence comparisons should also be distinguished from elapsed training time.

  5. Provide exact split identifiers, code versions, random seeds, and complete scheduler and checkpoint-selection settings. These details would make the experimental results easier to reproduce.

3 comments
AuthorsOpenReview ↗

Response to Reviewer TXMg (Part 1)

Thanks for your positive and constructive comments. We have revised our paper based on your comments, adding more explanations and experiments, with new text in blue. Below, we answer your questions one by one.

W1 & Q4 & S6: Practical advantage beyond the selection-rule check

A1: The forbidden coupling in Table 3 is indeed a symmetry check, and DeepH-E3 and MACE-H pass it as well. However, the two kinds of baselines differ from EACO in different ways: SLEM has the same architecture but lacks the reflection constraint, while DeepH-E3 and MACE-H are exactly O(3)-equivariant. Following your suggestion, we therefore separate the two advantages: exact symmetry against SO(3) models, and accuracy and cost against O(3) models. We show the first in a transport observable and the second in the full coupling matrix (new App. D.7), using the same DFT data and float64.

1. Exact symmetry changes a transport observable (vs. SLEM). The horizontal mirror forbids one-phonon scattering of π\pi electrons by flexural phonons [1, 2]. We compute the phonon-limited resistivity of doped graphene at 300 K with the default HamEPC workflow [3]. It covers all six branches and uses the complete prediction workflow: each model supplies both HH and ∂H\partial H. SLEM's small mirror error opens the forbidden flexural channel. Flexural phonons are soft (ω∝q2\omega\propto q^2), and their scattering is cut off only by the sample size LL or by strain [4], so this channel grows with LL.

Table R1. Spurious flexural resistivity in % of the DFT resistivity (27.0 Ω\Omega), EF=0.1E_F=0.1 eV, 300 K.

LLSLEMSLEM+augEACODeepH-E3MACE-H
1 μ\mum57156.6×10−176.6\times10^{-17}6.9×10−166.9\times10^{-16}6.8×10−136.8\times10^{-13}
10 μ\mum6341576.3×10−166.3\times10^{-16}7.7×10−157.7\times10^{-15}7.7×10−127.7\times10^{-12}

At EF=0.05E_F=0.05 eV, SLEM reaches 265% and 2850%. Inversion augmentation reduces this channel but does not remove it. For an SO(3) model, the forbidden coupling vanishes only through cancellations, so its size depends on how the derivative is assembled. For O(3)-equivariant models, the selection rule holds term by term, and the channel is zero in every workflow.

2. Accuracy and cost: the full coupling matrix (vs. DeepH-E3 and MACE-H). We compare gmnν(k,q)g_{mn\nu}(\mathbf k,\mathbf q) with DFT over the whole Brillouin zone: a 12×1212\times12 k\mathbf k mesh, 143 q\mathbf q points, all six branches, and the eight lowest bands, from −20-20 to +18+18 eV around the Dirac point.

Table R2. Relative error of the coupling matrix (%) and inference cost of one Hamiltonian for the 98-atom supercell (A100, float32).

SLEMEACODeepH-E3MACE-H
Flexural, shared DFT states0.2240.1020.3190.348
In-plane, shared DFT states0.8600.8481.0681.140
All branches, shared DFT states0.6770.6560.8470.904
All branches, complete workflow, within 3 eV of the Dirac point0.5870.5480.7910.649
Parameters433k358k628k2.35M
Time / memory94 ms / 117 MiB102 ms / 117 MiB83 ms / 259 MiB211 ms / 6,987 MiB

With shared DFT states, EACO's error is 3.13.1–3.4×3.4\times lower on the flexural branches and 23–27% lower on all branches than those of the O(3) baselines, with the fewest parameters and 60×60\times less memory than MACE-H. Our DFT reference reproduces ⟨gK2⟩\langle g^2_{\mathbf K}\rangle (0.100 vs. 0.0994 eV$^2$ [5]), and its acoustic resistivity (21.9 Ω\Omega) is close to the first-principles value of Sohier et al. [6].

To answer Q4: electron–phonon transport in 2D materials with a mirror plane is such an application. There, exact mirror symmetry rules out a spurious scattering channel that SO(3) models can open, and EACO provides it with fewer parameters and lower coupling errors than the O(3) baselines.

We have stated in Sec. 4.2 that only the Hamiltonian derivative comes from each model in Table 3, so it isolates the learned response, and added both studies to App. D.7.

[1] J. L. Mañes. Symmetry-Based Approach to Electron–Phonon Interactions in Graphene. Phys. Rev. B 76, 045430, 2007.

[2] E. Mariani and F. von Oppen. Flexural Phonons in Free-Standing Graphene. Phys. Rev. Lett. 100, 076801, 2008.

[3] Y. Zhong et al. Accelerating the Calculation of Electron–Phonon Coupling Strength with Machine Learning. Nat. Comput. Sci. 4, 615–625, 2024.

[4] E. Mariani and F. von Oppen. Temperature-Dependent Resistivity of Suspended Graphene. Phys. Rev. B 82, 195403, 2010.

[5] S. Piscanec et al. Kohn Anomalies and Electron–Phonon Interactions in Graphite. Phys. Rev. Lett. 93, 185503, 2004.

[6] T. Sohier et al. Phonon-Limited Resistivity of Graphene by First-Principles Calculations: Electron–Phonon Interactions, Strain-Induced Gauge Field, and Boltzmann Equation. Phys. Rev. B 90, 125414, 2014.

AuthorsOpenReview ↗

Response to Reviewer TXMg (Part 2)

W2 & Q1: Evaluation splits and checkpoint selection

A2: Table 2 originally reported validation errors for all five materials. We now report test errors for graphene and MoS$_2$ in Table 2 and Fig. 2(b). The bulk datasets of SLEM provide 150 training and 30 held-out structures, and SLEM reports its errors on these 30 structures. We follow this protocol for Si, GaN, and HfO$_2$, and these structures also select the checkpoint (Sec. 4.1).

1. The split does not change the comparison. On graphene and MoS$_2$, the test errors of all four methods are within 2.5% of their validation errors, and the ranking of the methods is unchanged.

2. Checkpoint selection follows the official baselines. DeepH-E3 tests the checkpoint with the lowest validation loss (best_model.pkl), MACE-H keeps the same checkpoint, and DeePTB, the code base of SLEM, keeps its best checkpoint in the same way. We apply this rule to all four methods, and the selected checkpoints lie in the last 7% of training.

W2 & Q2 & S7: Repeated runs and the scope of the accuracy claims

A3: We have added a second run of SLEM and EACO with another seed in App. D.5. EACO is more accurate on Si and HfO$_2$ in both runs, by 12–22%, and the two methods are on par on GaN, graphene, and MoS$_2$.

1. Accuracy claims. Following your suggestion, we now state that, with 17–30% fewer parameters, EACO matches or improves SLEM's accuracy, with clear gains on Si and HfO$_2$ (Abstract, Sec. 1, Sec. 4.1). Sec. 4.1 also notes that, in both runs, EACO improves the on-site errors on the three bulk materials, and that MACE-H is the most accurate model on MoS$_2$.

2. Augmentation on Si. A second run of inversion-augmented SLEM on Si reaches 0.067 meV, compared with 0.056 meV for EACO (Table 18 in App. D.8). Augmentation leaves the improper residual of SLEM at 10−410^{-4}, three orders of magnitude above that of EACO (Fig. 4).

AuthorsOpenReview ↗

Response to Reviewer TXMg (Part 3)

Q3: Displacement amplitude and the EPC structures

A4: 1. The allowed errors do not depend on the amplitude. We recompute Table 3 at each amplitude η\eta.

Table R3. Allowed-coupling error (%) at η=0.005\eta=0.005 / 0.010.01 / 0.020.02 Å and the η2\eta^2-weighted estimate of Table 3.

ZAZO
SLEM0.221 / 0.221 / 0.218 / 0.2190.218 / 0.218 / 0.220 / 0.220
SLEM+aug0.210 / 0.210 / 0.207 / 0.2080.245 / 0.246 / 0.249 / 0.248
DeepH-E30.283 / 0.282 / 0.280 / 0.2800.439 / 0.438 / 0.437 / 0.437
MACE-H0.263 / 0.262 / 0.260 / 0.2600.360 / 0.362 / 0.363 / 0.363
EACO0.101 / 0.100 / 0.099 / 0.1000.110 / 0.111 / 0.114 / 0.113

Across amplitudes, each error changes by at most 0.004 percentage points. The forbidden couplings are also unchanged: 0.129 and 0.324 meV for SLEM, and below 2×10−102\times10^{-10} meV for the three O(3)-equivariant models.

2. The EPC structures are not in the training set. The reference is an ideal flat 7×77\times7 supercell with 98 atoms, and each displaced structure moves one of its two inequivalent atoms by ±η\pm\eta along xx, yy, or zz, which gives 36 displaced structures. The training set contains 270 thermal snapshots of a 6×66\times6 supercell with 72 atoms, and none of them is flat: their out-of-plane RMS displacements of 0.08–0.31 Å are 44–62×62\times the finite-difference amplitudes. The test therefore probes a new supercell and an exactly flat geometry.

We have added Table R3 and this description to App. D.2.

S5, S8, S9: Revision suggestions

A5: 1. Contribution (S5). Sec. 1 now states the contribution as the complete solution of the O(2) constraint inside the edge-aligned convolution, and it ties the necessity-and-sufficiency result to the class of edge-aligned local-frame maps.

2. Complete-model cost (S8). Fig. 5(a) measures the angular operator alone. Table 12 in App. D.3 now reports the complete training time of every run. EACO takes 31–92 h, similar to SLEM (33–90 h), and finishes before DeepH-E3 (56–104 h) and MACE-H (56–133 h) on every material. Fig. 5(b) counts epochs, and Table 12 gives the elapsed time. The inference cost on the 98-atom graphene supercell is in Table R2 (See our response in A1).

3. Reproducibility (S9). Table 9 in App. D.1 lists the full training configurations, including the learning-rate schedules and seeds, and Sec. 4.1 states the splits and the checkpoint rule. We will release our code together with these configurations.

In summary, we hope the above response resolves your concerns and makes you more confident in raising your score in the following discussion and decision phases. Thanks in advance for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.