acceptodds
Under review as a conference paper at ICLR 2027

Equivariant Deep Learning in Direct-Product Space for Electronic-Structure Hamiltonian Prediction

10 views

Abstract

Deep learning provides a promising route to efficient prediction of electronic-structure Hamiltonians, enabling high-throughput screening across large collections of materials. However, most existing equivariant architectures propagate features in direct-sum spaces of irreducible spherical tensors, whereas the target Hamiltonian is naturally an operator block in a direct-product representation over local orbital and spin bases. This representation mismatch fragments contiguous Hamiltonian blocks, obscures direct manipulation of orbital couplings, and organizes features into irregular angular-momentum components that require numerous transformations and are poorly aligned with efficient hardware execution and memory management. To address both the representation mismatch and computational bottleneck, we propose DP-Learner, an equivariant learning framework that performs message passing directly on Hamiltonian-like operator blocks in the direct-product space. Theoretically, we propose Lie-generator-based directional embedding and invariantly weighted left–right products and contractions that preserve equivariance to , , and spatial inversions, without explicit Clebsch–Gordan decomposition or inverse reconstruction. Methodologically, we instantiate this principle as a Transformer architecture, in which Hamiltonian-like feature matrices or tensors evolve through equivariant operators in the direct-product space, while non-linear invariant scalar routing controls structured gating over matrix units, directional propagation, attention, and environment-embedded context interactions. We further construct MP-HAM-110K, a large-scale Hamiltonian dataset comprising more than 110,000 material structures covering 68 elements with both Hamiltonian and wavefunction data. Experiments on MP-HAM-110K and Materials-HAM-SOC demonstrate that DP-Learner dramatically improves Hamiltonian prediction accuracy and inference efficiency over the state-of-the-art method, with up to a % reduction in prediction error, an speedup in neural inference, and a % reduction in peak GPU memory consumption.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Quantum Physics, AI for Materials, Equivariant Deep Learning, Electronic-Structure Hamiltonian Prediction

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_1aaQOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DP-Learner, an equivariant neural network for predicting electronic-structure Hamiltonians in direct-product space. Starting from an initial Hamiltonian, the model retains atom-pair matrix blocks and learns corrections using direction operators, neighborhood matrices, and environment-dependent scalar weights, while preserving rotation and spatial-inversion equivariance. The authors also introduce MP-HAM-110K, a dataset covering 68 elements. On two datasets with and without spin–orbit coupling (SOC), DP-Learner reduces Hamiltonian prediction errors, inference time, and GPU memory use compared with NextHAM. Ablation studies support the usefulness of the main components, and band-structure comparisons across several test systems show agreement with DFT references.

Critical strengths
  1. The model learns directly on Hamiltonian matrices while ensuring that predictions transform correctly when the material is rotated. This provides a promising approach to Hamiltonian prediction.

  2. On both datasets, the model is more accurate than NextHAM while running faster and using less GPU memory. The ablation studies show that the main components improve prediction accuracy, and the band-structure comparisons across several materials show close agreement with DFT calculations.

Critical weaknesses

Besides equivariance, Hermiticity is an important physical property of a Hamiltonian. The paper and the released code do not clearly explain how the final predictions are guaranteed to satisfy Hermiticity. They also do not report the extent of any deviations or their effect on band-structure calculations. This leaves the physical consistency of the predictions insufficiently verified.

Questions for authors
  1. How is Hermiticity ensured in the final predicted Hamiltonians? Is Hermitian symmetrization applied before calculating band structures? Is the same step also applied when calculating Gauge MAE?

  2. Appendix K explains the use of four NextHAM submodels and one DP-Learner model. Beyond this difference, is there a runtime breakdown showing which operations account for the speedup, particularly the benefit of direct matrix operations?

  3. NextHAM was retrained on the non-SOC dataset. Could the authors provide more details about its training settings, model selection, and convergence, so that the accuracy comparison can be assessed more fully?

  4. The predicted band structures agree well with the DFT references, but direct comparisons with NextHAM are limited. Are quantitative band-energy or band-gap error comparisons with NextHAM available for a broader test subset?

5 comments

We thank the reviewer for the careful reading and for recognizing the value of learning directly on Hamiltonian matrices. We appreciate the opportunity to clarify how we assemble the final Hamiltonian and how we evaluate DP-Learner and NextHAM.

1. Hermiticity of the final Hamiltonians (Weakness and Question 1)

We apply Hermitian symmetrization before computing band structures. Following Eq. (160) in Appendix K, we jointly symmetrize each reciprocal pair (a,b,R)(a,b,\mathbf R) and (b,a,−R)(b,a,-\mathbf R) using

Habphys(R)=12[H~ab(R)+H~ba(−R)†].H_{ab}^{\mathrm{phys}}(\mathbf R)=\tfrac12\left[\widetilde H_{ab}(\mathbf R)+\widetilde H_{ba}(-\mathbf R)^\dagger\right].

For an on-site block (a=ba=b and R=0\mathbf R=0), this averages the block with its own adjoint. We apply the same operation to the learned correction. Hermitian reciprocity ensures that the Fourier-assembled Hamiltonian is Hermitian at every k\mathbf k. On all 3,000 Materials-HAM-SOC test structures, the measured maximum and mean reciprocity residuals after symmetrization are zero for both the final Hamiltonian and the learned correction, including on-site blocks, in FP32 and FP64 for both original and transformed inputs.

In our original workflow, we applied this symmetrization before computing downstream physical properties. We have now added it to the inference code so users can obtain Hermitian Hamiltonians at inference. Applying symmetrization before scoring reduces DP-Learner's test Gauge MAE from 0.841 to 0.806 meV on Materials-HAM-SOC and from 1.595 to 1.541 meV on MP-HAM-110K. Nevertheless, NextHAM computes Gauge MAE before Hermitian symmetrization. To compare the two models on equal terms, we report DP-Learner’s original Gauge MAEs of 0.841 and 1.595 meV in Table 1, with both models evaluated at the same stage of inference.

Theorem 3.6, proved in Appendix F.1, establishes that the SOC readout preserves rotation and spatial-inversion equivariance and yields a time-reversal-symmetric prediction when H(0)H^{(0)} has this symmetry. The derivation after Eq. (161) in Appendix K shows that Hermitian assembly preserves these guarantees. They therefore hold jointly for the final Hermitian Hamiltonian.

Numerical equivariance of the final SOC Hamiltonian. We use Materials-HAM-SOC as a representative SOC benchmark to test the complete prediction pipeline on 3,000 test structures. For each structure, we combine spatial inversion with a Haar-uniform rotation in SO(3)\mathrm{SO}(3) and apply the transformation consistently to the atomic positions, periodic lattice vectors, atom-pair displacements, and H(0)H^{(0)}. We compare the prediction from the transformed input with the correspondingly transformed prediction from the original input. Both paths include the SOC readout and Hermitian symmetrization. We use the same FP32-trained checkpoint and the same transformations for the FP32 and FP64 evaluations.

We compute absolute errors as the moduli of the elementwise differences and pool them over all physical orbital–spin entries in the test set. We compute relative errors only where the magnitude of the transformed reference entry exceeds 1 eV, since dividing by near-zero entries can give large ratios even for small absolute errors. We apply this threshold separately to the learned correction and the final Hamiltonian. The model forward passes use FP32 or FP64, as indicated in the table. For each precision, we construct the reference outputs in FP64 by applying the same combined rotation and spatial inversion to the Hamiltonian and learned correction predicted from the original input.

PrecisionQuantityMax. absolute error (eV)Mean absolute error (eV)Max. relative error (ref. magnitude > 1 eV)Mean relative error (ref. magnitude > 1 eV)
FP32Final Hamiltonian5.896×10−65.896\times10^{-6}2.520×10−62.520\times10^{-6}3.004×10−63.004\times10^{-6}4.885×10−74.885\times10^{-7}
FP32Learned correction3.435×10−63.435\times10^{-6}6.759×10−76.759\times10^{-7}1.732×10−61.732\times10^{-6}2.856×10−72.856\times10^{-7}
FP64Final Hamiltonian4.758×10−154.758\times10^{-15}6.031×10−176.031\times10^{-17}1.156×10−151.156\times10^{-15}3.424×10−173.424\times10^{-17}
FP64Learned correction2.638×10−152.638\times10^{-15}2.441×10−172.441\times10^{-17}9.044×10−169.044\times10^{-16}1.434×10−171.434\times10^{-17}

For both outputs, the maximum absolute equivariance error under combined rotation and spatial inversion is below 6×10−66\times10^{-6} eV in FP32 and falls below 5×10−155\times10^{-15} eV in FP64. The error magnitudes and their decrease at higher precision are consistent with floating-point roundoff and support the spatial equivariance of the complete prediction pipeline.

2. Inference latency and component-wise costs (Question 2)

On Materials-HAM-SOC, Table 1 reports a mean inference-pipeline latency of 3.014 s per structure for NextHAM, with one distance-specific submodel on each of four GPUs, and 0.269 s for DP-Learner on one GPU. These timings include input transfer, the forward pass, output conversion, and saving to disk. We also time the model forward call under these configurations. The mean host wall-clock latency per structure is 2.837 s for NextHAM's joint four-GPU forward call and 0.170 s for DP-Learner.

For memory, we measure PyTorch's peak reserved memory on each device and average it over test structures. The four NextHAM GPUs have mean peaks of 5.845, 4.851, 4.851, and 4.851 GiB, respectively, summing to the 20.398 GiB reported in Table 1. The first GPU also gathers the outputs from the other three devices. DP-Learner has a mean peak of 0.989 GiB on its single GPU.

We profile DP-Learner's FP32 forward pass on one NVIDIA A800 80 GB GPU with the inputs already on the device. CUDA events measure the forward interval on the model stream. The host wall-clock measurement also captures CPU execution, CUDA kernel submission, and any waits during the forward call. The step numbers in the table refer to Algorithm 1 of the revised paper.

Operation or measurementCorresponding computationMean time (ms/structure)Share of CUDA-event timeTrainable parameters
Fixed spin–orbital layout permutationsReorder spin and orbital indices at input and at the correction output (Appendix K).0.1790.21%0
Invariant descriptors and feature assemblyCompute radial features and subblock Frobenius norms, then assemble the invariant inputs (steps 2–3, 7, and 14).10.00711.52%0
Invariant MLPs, excluding attentionInitialize and update invariant node and edge features, and predict subblock gates, directional weights, context coefficients, and the radial envelope (steps 3–4, 8, 12, 15, 18, and 20).22.42325.81%21,035,481
Directional operator constructionCombine the bond direction with the rotation generators to construct VaV_a and VbV_b (step 9, Eq. (110)).10.51412.10%0
Dense matrix productsApply the two-sided directional product and the left, right, and two-sided context products (steps 10 and 16, Eqs. (113) and (124)).14.24016.39%0
Attention and aggregationPredict attention logits, project values, and aggregate incoming edge features (steps 8 and 11–12, Eqs. (114)–(116) and (118)–(119)).14.69816.92%5,409,864
Gating and residual updatesApply the subblock and edge gates, scale the residuals with LayerScale, and accumulate the updates (steps 6, 9–10, 12–13, and 16–18).11.23112.93%36
Other forward work and timing overheadRemaining time within the profiled forward interval.3.5804.12%0
Full forward interval on the CUDA streamModel forward through the orbital–spin correction output.86.875100%26,445,381
Forward-call latency on the hostWall-clock time of the model call through the correction output.170.264——

The dense products used for directional and context updates account for 16.39% of the CUDA-event interval. DP-Learner keeps orbital–spin blocks in the same basis across layers, so these products act directly on the blocks (Appendix H.2). The invariant MLPs, excluding attention, contain 79.5% of the trainable parameters and account for 25.81% of the CUDA-event interval. They operate on rotation-invariant scalar features and learn nonlinear gates and coefficients from the electronic environment. This allocation lets DP-Learner devote most of its trainable capacity to learning how each environment should modify the Hamiltonian, then apply those updates efficiently through batched matrix products in the physical orbital–spin basis.

On the testing set of Materials-HAM-SOC, we separately profiled NextHAM's projection module, edge_lin, the layer that maps NextHAM’s final edge features to the irreducible coefficients of the Hamiltonian correction. We found that this module alone takes 386.755 ms per structure per submodel by CUDA-event timing. DP-Learner’s entire forward pass on one A800 takes 86.875 ms by CUDA-event timing and 170.264 ms by host wall-clock timing. NextHAM’s output projection alone is much longer than DP-Learner’s full forward pass.

3. NextHAM retraining and model selection on MP-HAM-110K (Question 3)

We retrain NextHAM from scratch on MP-HAM-110K using the main branch of its official implementation (https://github.com/DavidYin94/NextHAM/tree/main). For the common non-SOC orbital basis 4s2p2d1f4s2p2d1f, the input and output irreps both use the full irreducible decomposition of the 729-dimensional space Vorb⊗Vorb∗\mathcal V_{\mathrm{orb}}\otimes\mathcal V_{\mathrm{orb}}^\ast, where Vorb=4V0e⊕2V1o⊕2V2e⊕V3o\mathcal V_{\mathrm{orb}}=4V_0^e\oplus2V_1^o\oplus2V_2^e\oplus V_3^o. We retain the original hidden representations, network depth, internal modules, and four distance-specific submodels.

NextHAM and DP-Learner use the same data split, three seeds (2025, 2026, and 2027), and Gauge MAE evaluation. Each run uses 100 epochs with Adam. The learning rate increases linearly from 10−610^{-6} to 5×10−45\times10^{-4} over the first five epochs, then follows cosine decay to 10−510^{-5} over the remaining 95 epochs. For each run, we select the checkpoint with the lowest Gauge MAE in the validation set. The selected NextHAM checkpoints come from epochs 78, 76, and 81 in seed order. The corresponding Gauge MAEs in the testing set are 2.396, 2.389, and 2.397 meV, with a sample standard deviation of 0.0044 meV. The three-seed means are 2.394 meV for NextHAM and 1.595 meV for DP-Learner, a 33.4% reduction under the matched protocol.

4. Comparison of band-energy errors near the Fermi level (Question 4)

On the testing set of Materials-HAM-SOC, we compare the near-Fermi band energies from H(0)H^{(0)}, NextHAM, and DP-Learner with self-consistent ABACUS DFT references. Following NextHAM, for the two learned models we use the network outputs and DFT labels at atom-pair distances below 6 Å and use H(0)H^{(0)} for both predictions and references at distances of 6 Å or more. The H(0)H^{(0)} baseline uses the self-consistent H(T)H^{(T)} as its reference. For the model comparison, we apply Hermitian symmetrization and solve the generalized eigenvalue problem in PyATB using the same overlap matrix on a common 10×10×1010\times10\times10 mesh of 1,000 kk points. At each kk point, we sort the full spectra and pair bands by index. We subtract the same DFT Fermi energy from both spectra and keep bands whose reference energies then lie within [−10,10][-10,10] eV. For each structure and prediction, we report the MAE before and after an additional rigid alignment. The alignment adds a single shift to the predicted energies, given by the median of the paired reference-minus-prediction differences in the selected window. We report the mean of the per-structure MAEs in eV and the percentage of structures in each error interval.

Energy alignmentHamiltonianMean MAE (eV) ↓[0,0.1)[0,0.1) eV[0.1,0.5)[0.1,0.5) eV[0.5,1)[0.5,1) eV[1,∞)[1,\infty) eV
Before additional rigid alignmentH(0)H^{(0)}0.523626.30%28.96%31.56%13.16%
Before additional rigid alignmentNextHAM0.387918.03%50.66%28.26%3.03%
Before additional rigid alignmentDP-Learner0.107774.96%21.73%2.23%1.06%
After additional rigid alignmentH(0)H^{(0)}0.483227.40%29.26%30.30%13.03%
After additional rigid alignmentNextHAM0.124163.46%33.10%2.96%0.46%
After additional rigid alignmentDP-Learner0.088579.06%18.56%1.96%0.40%

With the DFT Fermi energy as the common energy zero, DP-Learner reduces the mean near-Fermi band-energy MAE by 72.272.2% relative to NextHAM. After an additional per-structure, median-based shift of the predicted band energies, the reduction is 28.728.7%. The fraction of structures with an MAE below 0.10.1 eV rises from 18.0318.03% to 74.9674.96% with the common Fermi reference alone, and from 63.4663.46% to 79.0679.06% after the additional shift.

The improvement indicates that DP-Learner better reproduces the relative positions of near-Fermi bands across the sampled Brillouin zone. The accuracy gains in real-space Hamiltonian prediction thus carry through to the reciprocal-space spectrum, which provides the band edges, Fermi-level crossings, and energy differences needed to analyze electronic and optical properties. The larger fraction of structures with an MAE below 0.10.1 eV highlights the potential of DP-Learner for high-throughput materials screening. For example, its predicted spectra could help prioritize direct-gap semiconductors with gaps in a target range for optoelectronic applications, or identify candidate semimetals with near-Fermi band crossings for detailed study.

Reviewer_3sCJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DP-Learner, a model that predicts electronic structure Hamiltonians by learning corrections to an initial Hamiltonian (H(0)H^{(0)}). It employs Lie-generator-based directional operators, an invariant gating mechanism, and context-dependent matrix contraction operations to perform computations directly on Hamiltonian-form features, thereby avoiding explicit transformations between irreducible representations. Furthermore, the authors have constructed the MP-HAM-110K dataset, which comprises 113,800 periodic structures spanning 68 elements. Experimental results on this dataset, as well as the Materials-HAM-SOC dataset, demonstrate that DP-Learner offers significant improvements in both prediction accuracy and inference efficiency compared to NextHAM.

Critical strengths

This architecture offers an effective way to combine equivariance with efficient dense matrix computation. Retaining operator-valued features during the message-passing process is a highly valuable design choice; the performance gains achieved by this architecture, compared to NextHAM, are significant.

MP-HAM-110K represents a potentially valuable resource. Furthermore, five ablation studies conducted on this architecture—evaluated across two datasets using three different random seeds—provide compelling evidence for the contributions of its individual components.

Critical weaknesses

The comparison with existing methods is insufficient. Among the competing models evaluated, only NextHAM is included. Other relevant approaches—such as equivariant prediction models (e.g., HamGNN), fixed-point methods (e.g., DEQHNet and HamEvo), and flow-based Hamiltonian generation methods (e.g., QHFlow)—serve as essential benchmarks for assessing the representation and update mechanisms proposed in this paper. To establish the advantages of the proposed method in terms of accuracy and efficiency, it is necessary to conduct comparisons against these representative approaches under unified input conditions and evaluation settings. While the internal ablation studies demonstrate the effectiveness of individual components within DP-Learner, they fail to establish the model's competitiveness relative to other architectures—a critical piece of evidence missing from the support for the paper's core argument.

The scope of the evaluation is too narrow, being confined solely to the selected material datasets. Both datasets employed in the study focus on the prediction of periodic materials given an initial Hamiltonian H(0)H^{(0)}; however, the proposed equivariant matrix operations do not inherently rely on periodicity. Existing molecular datasets—such as QH9, MD17 configurations with Hamiltonian annotations, and drug-like molecular datasets like nablaDFT and ∇2\nabla^2 DFT—could serve as supplementary benchmarks. Notably, methods such as DEQHNet and QHFlow have previously been evaluated on the QH9 and MD17 datasets. The MP-HAM-110K dataset, despite its specific scale and elemental coverage, cannot substitute for comprehensive testing across diverse molecular structures, conformational variations, and system sizes. Without such supplementary evaluations, the applicability and robustness of the proposed framework cannot be adequately demonstrated.

Questions for authors

Could the authors compare their proposed model against representative equivariant Hamiltonian models, fixed-point-based Hamiltonian models, or flow-based Hamiltonian models, while ensuring consistent input and supervision settings? Once the influence of major architectural differences is accounted for, what performance advantages does the model still retain?

Beyond the need to prepare compatible H(0)H^{(0)} inputs and consistent basis function conventions, are there any other factors hindering evaluation on molecular Hamiltonian benchmarks? Could the results on these benchmarks demonstrate the broader applicability of the proposed framework?

7 comments

We thank the reviewer for the careful assessment and for asking us to broaden the empirical comparison. We have reported the results of our method on a molecular benchmark with several representative Hamiltonian predictors in the revised paper. Below, we explain the new results, the coverage of the original benchmarks, and how these experiments assess the proposed representation-learning framework.

1. Research scope: broad materials evaluation, complemented by molecular experiments.

Given the chemical and structural complexity of atomistic systems, a thorough study of either crystalline materials or molecules is a substantial and scientifically reasonable scope for a single Hamiltonian-learning paper. Each domain encompasses extensive variation in composition, geometry, and electronic structure, with its own data-generation and evaluation requirements. Published Hamiltonian-learning studies provide numerous precedents for this focused approach:

  • Materials-focused studies: DeepH [14], DeepH-E3 [1], the DeepH-2-based universal materials model [13], and NextHAM [2] center their investigations on crystalline materials.

  • Molecular studies: PhiSNet [15], QHNet [3], WANet [6], SPHNet [7], QHNetV2 [8], QHFlow [4], and QHFlow2 [9] conduct their reported evaluations on molecular systems.

Our original evaluation was designed to test a unified Hamiltonian representation learning framework across chemically diverse periodic materials in this study. Together, the two datasets contain 130,800 structure records across their training, validation, and test splits. Each dataset covers 68 elements, and the experiments include both SOC and non-SOC Hamiltonians. This provides a broad test within the materials domain. The molecular experiments further examine whether the same representation principle extends to finite systems. We had been developing this extension for a follow-up study and have now incorporated the completed QH9-Stable-ID results into Appendix M in response to the reviewer's suggestion.

QH9-Stable-ID contains 130,831 molecular geometries, with 104,664/13,083/13,084 molecules in the training/validation/test splits [5]. The molecular implementation retains DP-Learner's direct-product feature propagation, matrix contractions, gating, and attention. We replace lattice-indexed crystal blocks with molecular atom-pair blocks and adapt the orbital layout to the molecular basis. These experiments confirm that periodicity is not a restriction of the framework.

The physical initialization also extends naturally from crystals to molecules. Following QHFlow2, we use PySCF's MINAO initial density to construct H(0)H^{(0)} without SCF iterations. This matrix already contains the one-electron terms and density-dependent mean-field interactions in the molecule's atomic-orbital basis. It therefore supplies physically informed orbital couplings for learning the correction to the converged Hamiltonian, preserving the role of H(0)H^{(0)} in our materials model.

For the QHFlow2 comparison, we use the same data split, B3LYP/def2-SVP targets, MINAO construction of H(0)H^{(0)}, complete atom-pair graph, corruption distribution, training objective, optimizer, and learning-rate schedule. We compare the same molecular metrics with one- and three-step Euler sampling. Training uses 260,000 updates and a global batch size of 32 on two A800 GPUs. The DP-Learner checkpoint is selected by the full-validation Hamiltonian MAE with three-step sampling, as specified in Appendix M. By design, DP-Learner uses the complete H(0)H^{(0)} as matrix features, including off-site blocks; QHFlow2 directly conditions on on-site physical blocks and adds the complete H(0)H^{(0)} back during residual reconstruction. Appendix M describes these input interfaces explicitly.

Across its one- and three-step settings, DP-Learner achieves the lowest value on all five error metrics among the compared models, with 42.8% fewer trainable parameters than QHFlow2-M. It also matches the highest occupied-orbital similarity at the displayed precision. Table 6 of the revised paper is reproduced below. Errors are in μHa\mu\mathrm{Ha}, and ScS_c is in percent. Steps gives the number of Euler sampling steps. DP-Learner is evaluated with one and three steps using the same checkpoint, with one network evaluation per step. Bold values include ties after rounding.

ModelParam.StepsH↓H\downarrowϵocc↓\epsilon_{\mathrm{occ}}\downarrowSc↑S_c\uparrowHOMO$\downarrow$LUMO$\downarrow$GAP$\downarrow$
QHNet (w/o init)20M—77.7963.594.81546.318257.317822.6
QHNet20M—46.0459.698.3470.62781.22967.9
WANet29M—80.0833.696.9———
SPHNet30M—45.5334.397.8———
QHNetV2 (w/o init)——31.5417.998.6———
QHFlow28M—23.0119.799.5179.5438.0553.9
QHFlow2-S12.37M310.784.299.792.8222.8276.4
QHFlow2-S12.37M110.785.199.794.1214.9267.4
QHFlow2-O(3)O(3)-S12.37M39.168.499.877.3193.6232.9
QHFlow2-O(3)O(3)-S12.37M19.168.499.878.6190.3227.3
QHFlow2-M44.25M37.151.099.854.1105.0129.0
QHFlow2-M44.25M17.151.199.855.1101.3126.3
DP-Learner25.30M36.447.499.849.8111.5118.0
DP-Learner25.30M16.447.699.850.7100.2117.6

The earlier QHNet, WANet, SPHNet, and QHFlow values are the reference results collected in the QHFlow2 paper [3, 4, 6, 7, 9]; QHNetV2 is taken from its own publication [8]. QHFlow2 values use the official updated benchmark release, v0.2.0-rc.2, dated October 28, 2026 [10]. Parameter counts follow the manuscript's convention: reported counts for earlier baselines and trainable counts for QHFlow2 and DP-Learner.

Compared with one-step QHFlow2-M, one-step DP-Learner reduces Hamiltonian MAE from 7.1 to 6.4, occupied-orbital energy MAE from 51.1 to 47.6, and gap MAE from 126.3 to 117.6, corresponding to improvements of approximately 9.9%, 6.8%, and 6.9%. One-step DP-Learner also improves all five error metrics over three-step QHFlow2-M. Relative to its one-step setting, three-step DP-Learner slightly reduces the occupied-orbital energy and HOMO errors, whereas one step gives lower LUMO and gap errors; Hamiltonian MAE and ScS_c are unchanged at the displayed precision. The S and M variants bracket our model's parameter count; QHFlow2-L has approximately 188M parameters, about 7.4 times the size of our model. We focus this comparison on S and M and leave a matched scaling study with the larger variants to future work.

These results also speak to the representation-learning question underlying our work. A Hamiltonian block couples the orbital spaces of two atoms. DP-Learner preserves these two orbital indices in its hidden features and learns changes in their couplings through left, right, and bilateral matrix actions. Directional operators and neighboring matrix features supply geometric and electronic context, while invariant networks determine how that context controls the updates. The lower matrix errors are accompanied by lower occupied-state and frontier-orbital energy errors, supporting the effectiveness of this representation in a molecular setting. Appendix M now develops this analysis alongside the complementary evidence from the crystal ablations.

2. The original materials evaluation already covers a broad range of chemistry and electronic-structure settings.

The two benchmark names summarize substantial diversity within each collection. Materials-HAM-SOC contains 17,000 structures, and MP-HAM-110K contains 113,800 structures; both cover 68 elements. The former tests complex orbital–spin Hamiltonians with SOC, while the latter tests spin-independent Hamiltonians across unit cells containing 1–169 atoms and up to nine element types. MP-HAM-110K uses element-dependent orbital bases reaching 4s2p2d1f4s2p2d1f. Together, these experiments examine the framework across different compositions, structural sizes, orbital spaces, and spin treatments. The revised presentation makes these dimensions of coverage explicit so that the breadth of the evaluation is visible beyond the number of dataset names. We chose the two materials benchmarks to test how well DP-Learner predicts Hamiltonians across crystals with diverse compositions and structures. Across 68 elements, a materials model must accommodate substantially more heterogeneous orbital bases and electronic environments, with spin-dependent couplings adding a further challenge in the SOC setting.

To make comparisons across this range of crystal chemistries reproducible, researchers need access to the same physical inputs and reference data. Few public materials benchmarks provide a physically informed H(0)H^{(0)}, a converged H(T)H^{(T)}, overlap matrices, and eigenstate information for the same structures. Materials-HAM-SOC is one such benchmark. MP-HAM-110K brings these quantities together for a broader range of materials, including crystals with more atoms per unit cell, thereby supporting Hamiltonian prediction and the joint real- and reciprocal-space supervision used in this work at a larger scale. By releasing the dataset alongside our code and trained models, we provide a shared benchmark and a reproducible baseline for future studies. Researchers can compare new methods with DP-Learner on the same structures, using the same physical inputs and reference targets.

3. NextHAM provides a strong and directly relevant principal baseline for the crystal experiments.

We selected NextHAM as the baseline to compare because it is a strong materials Hamiltonian predictor that also uses physically informed H(0)H^{(0)} inputs and residual learning. Its published comparisons establish a demanding reference: on Materials-HAM-SOC, Table 5 of NextHAM reports Gauge MAE of 1.417 meV versus 12.605 meV for DeepH-E3 [2]. Table 7 additionally reports improved accuracy on graphene and MoS$_2$ using a reduced NextHAM variant adapted to those benchmarks.

Another recent study provides further evidence on a different materials collection. Table 2 of Evaluating Predicted Densities, Hamiltonians, and Density Matrices as Periodic SCF Initializers [11] evaluates Hamiltonian predictors on 1,997 test crystals from the ρ\rhoHD-43K collection. After reconstructing the predicted Hamiltonians, the study computed MAE at the positions where the converged DFT reference was nonzero, using the same positions for all models. Under this evaluation, the NextHAM implementation achieved approximately 93.8% and 94.3% lower MAE than the HamGNN and DeepH-E3 implementations, respectively [1, 11, 12]. This comparison provides additional support for selecting NextHAM as our principal materials baseline.

Our comparison tests whether propagating Hamiltonian-like features directly in the direct-product space improves on the state-of-the-art performance. Under the materials evaluation protocol in our paper, DP-Learner reduces Gauge MAE from 1.417 to 0.841 meV on Materials-HAM-SOC and from 2.394 to 1.595 meV on MP-HAM-110K, improvements of 40.6% and 33.4%, respectively. The new molecular table broadens the accuracy comparison to representative regression and flow-based predictors, including the QHFlow and QHFlow2 families.

Reviewer_6nkGOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DP-Learner, which performs equivariant message passing while preserving the matrix structure of Hamiltonian blocks. It takes the zeroth-step Hamiltonian (H(0)H^{(0)}) as input and learns a delta-learning to the converged Hamiltonian. Intermediate features remain matrices or orbital–spin tensors in the AO basis rather than being decomposed into irreducible tensors. The model combines Lie-generator directional embeddings, atom-wise context aggregated from neighboring atom-pair features, and invariantly weighted matrix products, with conditions ensuring rotation and inversion equivariance. The paper also introduces MP-HAM-110K, containing 113,800 periodic structures across 68 elements, and evaluates accuracy, inference time, GPU memory, and band structures on this dataset and Materials-HAM-SOC.

Critical strengths

Accuracy and inference efficiency improve together. Incorporating directional and environmental information into Hamiltonian-shaped matrix features is an interesting design. Table 1 reports Gauge MAE reductions of 40.6% and 33.4% relative to NextHAM on Materials-HAM-SOC and MP-HAM-110K, respectively, alongside approximately 10–11.2× faster neural inference. Further analysis is needed to identify the sources of these gains, but the simultaneous improvements across two chemically diverse benchmarks are encouraging.

The dataset and architectural ablations are useful. MP-HAM-110K provides H(0)H^{(0)}, converged Hamiltonians, overlap matrices, and information for reciprocal-space supervision across diverse periodic structures. Table 4 includes an H(0)H^{(0)}-only baseline and five architectural ablations evaluated across three seeds, helping clarify the roles of directional embeddings, equivariant context, and invariant coefficients within the model.

Critical weaknesses

The additional contribution of the matrix-product design and its distinction from prior work are not sufficiently established

The paper emphasizes the representational and computational benefits of preserving Hamiltonian matrix structure (Introduction, lines 108–121; Section 6, lines 415–417). However, Table 1 does not isolate which design choices produce these gains. Total parameter counts (Table 1, lines 378–389) and asymptotic analyses (Appendix H, lines 1851–1874, 1905–1930) are provided, but component-wise runtime and parameter breakdowns are needed for basis transformations, tensor/matrix products, invariant MLPs, and attention. In particular, fixed transformation costs should be distinguished from learnable operations, and the comparison should account for optimized transformation implementations.

Model configuration is another confounding factor. NextHAM uses four distance-specific submodels, whereas DP-Learner uses one model across all distance intervals (Appendix K, lines 2184–2197). This is a practical comparison of complete predictors, but does not separate representation from model partitioning. The authors report increased overfitting with distance-specific DP-Learner models (Appendix K, lines 2188–2189), but quantitative results and training details are needed. Comparisons should control for model size and partitioning under the same data splits, and comparable training budgets. Internal ablations establish the usefulness of components within DP-Learner, rather than the advantage of matrix-product updates over existing architectures.

Relevant prior work is also missing. E3x [1] provides matrix-based equivariant interactions; Xie et al. [2] analyze runtime–expressivity trade-offs, including matrix tensor products; and BOA [3] introduces basis-overlap-based equivariant message passing. These works should be discussed to clarify what is added by persistent AO-basis matrices and the directional/context updates. Fixed matrix-product constructions can restrict independently adjustable couplings [2]. Although this does not establish an expressivity limitation for DP-Learner as a whole, equivariance alone does not establish comparable expressivity at lower cost.

QHFlow [4], QHFlow2 [5], QHNetV2 [6], HELM [7], and MALOQ [8] are also not cited, despite relevant developments in learning strategies, efficient architectures, elemental and basis-set coverage, and optimized kernels. Their relationship to this work should be explained, distinguishing directly comparable settings from those requiring adaptations for different inputs, basis sets, periodicity, or SOC.

The presentation makes the method and its contribution difficult to follow.

As I understand it, the main idea is to keep features in the form of AO-basis matrices throughout the network, whereas existing irreps-based Hamiltonian models decompose the matrices into irreducible components and update those features. However, this difference—and how the matrix updates preserve symmetry—is explained in a more complicated way than necessary. Despite the extensive mathematical detail, it remains difficult to see what changes in the actual computation. The figures do not make these differences much clearer, and Section 4 repeatedly refers readers to the appendices for definitions and update equations (lines 304–323, 346–357).

I strongly recommend starting with a simple comparison of the feature representations and update operations. This should explain what each method passes between layers, how the features interact, and why the updates preserve the required symmetries. A concise description of one forward pass, supported by pseudocode and a clearer architecture diagram, would help connect the mathematical formulation to the actual model. The main text and appendices should be organized around this explanation, with less repetition and a clearer separation between the main design and supporting details.

Questions for authors

Questions

Component-wise computational cost. Could the authors compare component-wise training/inference runtimes and trainable parameter counts with NextHAM, including the absolute and relative costs of basis transformations? How much of the difference remains with MALOQ’s Triton backend [8, 9], or an equivalent implementation, under matched basis, precision, and hardware settings? For training, please also consider HELM-style supervision directly on preprocessed irrep targets [7], avoiding Hamiltonian reconstruction for that loss. How do training time, memory, and accuracy compare under this setting, and what changes to the original objective are required? For on-the-fly use, please additionally report stage-wise total runtime including H(0)H^{(0)} generation and, if possible, band evaluation, beyond the currently reported neural inference time (Appendix K, lines 2424–2429).

A fair comparison with NextHAM. How much do NextHAM’s distance-specific submodels versus DP-Learner’s unified model affect Table 1? Please compare under the same data splits with reasonable training budgets. If a unified NextHAM model is impractical, could DP-Learner be evaluated with NextHAM’s distance partitioning and training settings? Quantifying the reported overfitting (Appendix K, lines 2188–2189) would help separate architectural benefits from joint training across distance intervals.

Computational cost versus expressivity. Xie et al. [2] analyze runtime–expressivity trade-offs for particular matrix tensor-product constructions. How does the proposed update compare with a general equivariant tensor product in its freedom to adjust couplings independently? Does their analysis apply here, and if so, do invariant coefficients and context updates compensate for those limitations?

Symmetry verification of the complete model. Could the authors report numerical equivariance tests under random rotations and spatial inversion, consistently transforming the structure, periodic lattice, and H(0)H^{(0)}? Please include final output post-processing and report absolute/relative errors for both the predicted Hamiltonian and learned correction, together with numerical precision. And please specify how the Hermiticity is guaranteed.

Quantitative physical observables. Beyond band plots, could the authors compare test-set band-gap errors and band-energy errors near the Fermi energy against NextHAM and H(0)H^{(0)}? Averages and error distributions would help establish how Hamiltonian MAE improvements translate into physical predictions.


Additional Suggestions

Evaluation on Hamiltonian-label benchmarks based on QH9 or MD17, with matched H(0)H^{(0)}, labels, functional, and basis settings, would help establish whether the accuracy and computational advantages extend to molecular systems.


References

[1] Unke and Maennel. E3x: E(3)-Equivariant Deep Learning Made Easy, 2024.

[2] Xie et al. The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products, 2025.

[3] Klockow et al. A Function-Centric Graph Neural Network Approach for Predicting Electron Densities, ICLR 2026. (BOA)

[4] Kim et al. High-order Equivariant Flow Matching for Density Functional Theory Hamiltonian Prediction, NeurIPS 2025. (QHFlow)

[5] Kim et al. Machine Learning Hamiltonians are Accurate Energy-Force Predictors, ICML 2026. (QHFlow2)

[6] Yu et al. Efficient Prediction of SO(3)-Equivariant Hamiltonian Matrices via SO(2) Local Frames, ICML 2026. (QHNetV2)

[7] Kaniselvan et al. Learning from the electronic structure of molecules across the periodic table, ICLR 2026. (HELM)

[8] Kaniselvan et al. Massively Accelerated Learning of Operators for Quantum Transport, 2026. (MALOQ)

[9] MALOQ public implementation. Triton basis-transform backend, commit e946a47.

10 comments

We thank the reviewer for the careful reading and for the detailed questions about our method and experiments. The comments help us clarify the design of DP-Learner and strengthen its evaluation. We respond to each point below.

1. Component-wise computational cost (response to Weakness 1, first paragraph, and Question 1)

We thank the reviewer for asking how the computation and trainable parameters are distributed across DP-Learner. We profile FP32 inference on the Materials-HAM-SOC test set using one NVIDIA A800 80 GB GPU, with the inputs already on the device. We time the forward interval on the model stream using CUDA events and report the mean per test structure. We also report host wall-clock timing, which reflects CPU execution, CUDA kernel submission, and any synchronization waits during the forward call. The step numbers in the table refer to Algorithm 1 of the revised paper.

Operation or measurementCorresponding computationMean time (ms/structure)Share of CUDA-event timeTrainable parameters
Fixed spin–orbital layout permutationsReorder spin and orbital indices at input and at the correction output (Appendix K).0.1790.21%0
Invariant descriptors and feature assemblyCompute radial features and subblock Frobenius norms, then assemble the invariant inputs (steps 2–3, 7, and 14).10.00711.52%0
Invariant multilayer perceptrons (MLPs), excluding attentionInitialize and update invariant node and edge features, and predict subblock gates, directional weights, context coefficients, and the radial envelope (steps 3–4, 8, 12, 15, 18, and 20).22.42325.81%21,035,481
Directional operator constructionCombine the bond direction with the rotation generators to construct VaV_a and VbV_b (step 9, Eq. (110)).10.51412.10%0
Dense matrix productsApply the two-sided directional product and the left, right, and two-sided context products (steps 10 and 16, Eqs. (113) and (124)).14.24016.39%0
Attention and aggregationPredict attention logits, project values, and aggregate incoming edge features (steps 8 and 11–12, Eqs. (114)–(116) and (118)–(119)).14.69816.92%5,409,864
Gating and residual updatesApply the subblock and edge gates, scale the residuals with LayerScale, and accumulate the updates (steps 6, 9–10, 12–13, and 16–18).11.23112.93%36
Other forward work and timing overheadRemaining time within the profiled model-forward interval.3.5804.12%0
Forward interval on the CUDA stream (CUDA events)Model forward through the orbital–spin correction output.86.875100%26,445,381
Forward-call latency (host wall clock)Host wall-clock timing of the model forward through its correction output.170.264——

The dense products used for directional and context updates account for 16.39% of the CUDA-event interval. DP-Learner keeps orbital–spin blocks in the same basis across layers, so these products act directly on the blocks (Appendix H.2). The invariant MLPs, excluding attention, contain 79.5% of the trainable parameters and account for 25.81% of the CUDA-event interval. They operate on rotation-invariant scalar features and learn nonlinear gates and coefficients from the electronic environment. This allocation lets DP-Learner devote most of its trainable capacity to learning how each environment should modify the Hamiltonian, then apply those updates efficiently through batched matrix products in the physical orbital–spin basis.

We appreciate the reviewer’s suggestion to profile NextHAM alongside DP-Learner with a component-by-component comparison. However, such a component-by-component comparison is difficult to interpret because the two models build their forward passes around different representations. NextHAM learns in the direct-sum space, where features are grouped into irreducible channels of different dimensions and updated through angular-momentum couplings. DP-Learner learns in the direct-product space, where each atom-pair feature retains its two orbital–spin indices and is updated through matrix operations. The two designs therefore differ in which operators they use, the dimensions of the features those operators act on, and how the operations are ordered and combined. For example, one two-sided matrix product in DP-Learner incorporates bond-direction information by acting on both orbital–spin indices of an atom-pair block. In NextHAM, bond-direction information enters the updates across multiple irreducible channels and coupling paths, and the orbital–spin block is reconstructed at readout. A timing comparison between individual operations would therefore capture different parts of the computation in the two models.

We compare the two models at the level of the full forward pass and the deployed inference pipeline. On Materials-HAM-SOC, Table 1 reports mean per-structure inference-pipeline latencies of 3.014 s for NextHAM on four GPUs and 0.269 s for DP-Learner on one GPU. These timings include input transfer, the network forward pass, conversion to the Hamiltonian output format, and saving to disk. The mean host wall-clock forward-call latencies per structure are 2.837 s for NextHAM and 0.170 s for DP-Learner, with the latter consistent with the value in the table above. Together with the trainable parameter counts and Gauge MAE, these timings compare the cost and accuracy of producing the same physical output on the same test structures.

2. Comparison with NextHAM under the same distance partition (Weakness 1, second paragraph, and Question 2)

We thank the reviewer for asking how distance partitioning affects the comparison in Table 1. With four distance-specific submodels for both methods, DP-Learner reduces Gauge MAE in the testing set relative to NextHAM by 37.0% on Materials-HAM-SOC and 31.6% on MP-HAM-110K.

We follow NextHAM's distance partition [1]. The four submodels predict corrections for atom pairs in the intervals [0,1)[0,1), [1,2)[1,2), [2,4)[2,4), and [4,6)[4,6) Å, with the first interval including the on-site blocks. We assemble a single Hamiltonian from the four sets of corrections and retain H(0)H^{(0)} at distances of 6 Å or more. The four-submodel DP-Learner uses four NVIDIA A800 GPUs for training and inference, with one submodel per GPU. Each DP-Learner submodel has the same parameter count as the corresponding unified model, so the complete predictor has four times as many parameters.

On each benchmark, we use the same data splits as in the Table 1 experiments and evaluate Gauge MAE on each assembled Hamiltonian. We retain the four Hamiltonian-loss weights used in the unified DP-Learner and NextHAM (λR=0.99955\lambda_R=0.99955, λP=0.0002\lambda_{\mathcal P}=0.0002, λQ=0.0001\lambda_{\mathcal Q}=0.0001, and λPQ=0.00015\lambda_{PQ}=0.00015). We train the four-submodel DP-Learner for 100 epochs with Adam, a five-epoch linear warmup from 10−610^{-6} to 5×10−45\times10^{-4}, and cosine decay to 10−510^{-5} over the remaining 95 epochs, following the NextHAM schedule. For the MP-HAM-110K results in Table 1, we report the mean over seeds 2025, 2026, and 2027 using checkpoints selected by Gauge MAE in the validation set.

The four-submodel comparison matches the distance partition. The table gives the total parameter count of each predictor.

DatasetModelSubmodelsTotal parameters (M)Gauge MAE in the training set (meV) ↓Gauge MAE in the testing set (meV) ↓
Materials-HAM-SOCNextHAM450.285—1.417
Materials-HAM-SOCDP-Learner (four submodels)4105.7820.6700.893
Materials-HAM-SOCDP-Learner (unified)126.4450.7590.841
MP-HAM-110KNextHAM450.263—2.394
MP-HAM-110KDP-Learner (four submodels)4104.3751.3961.638
MP-HAM-110KDP-Learner (unified)126.0941.5041.595

NextHAM’s ablation study on Materials-HAM-SOC shows that assigning separate submodels to different distance ranges reduces Gauge MAE in the testing set from 1.862 to 1.417 meV [1, Table 4]. On both benchmarks, the unified DP-Learner achieves lower Gauge MAE in the testing set with one quarter of the parameters of its four-submodel counterpart. The four-submodel version fits the training set more closely. The train–test gap grows from 0.082 to 0.223 meV on Materials-HAM-SOC and from 0.091 to 0.242 meV on MP-HAM-110K. The wider training–test gaps suggest that the distance-specific submodels are more prone to overfitting. One possible explanation is that the direct real-space supervision for each submodel comes from Hamiltonian blocks in its assigned distance range. These results support the unified configuration used in Table 1.

Reviewer_B99GOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DP-Learner for predicting electronic-structure Hamiltonians from an initial, non-self-consistent Hamiltonian H(0). Instead of propagating features in a direct sum of irreducible representations, the model maintains Hamiltonian-shaped matrices throughout the network: real 27 × 27 matrices in the non-SOC setting and complex 54 × 54 matrices in the SOC setting. Invariant scalar networks modulate unit-wise gates, Lie-generator-based directional operators, attention, and matrix contractions involving atom-wise context features.

The theoretical contribution establishes the rotation and inversion properties of these operations under explicit assumptions on the representations, basis alignment, and invariant coefficients. The empirical evaluation compares DP-Learner with NextHAM on Materials-HAM-SOC and the proposed MP-HAM-110K dataset. Table 1 reports lower Gauge MAE and lower measured inference time and GPU memory usage on both benchmarks. Appendix L includes five architectural ablations and a reference using H(0) without a learned correction. MP-HAM-110K contains 113,800 periodic structures with initial and self-consistent Hamiltonians, overlap matrices, and information for reciprocal-space supervision.

Critical strengths

Strength 1: The architecture offers a concrete alternative to propagating irreducible features, with explicit analysis of its symmetry properties.

Maintaining Hamiltonian-shaped matrices as the persistent network state is a clear architectural choice. It allows the model to express feature updates through matrix operations while retaining the row–column structure of orbital couplings. The paper develops this idea beyond a high-level proposal: Theorems 3.1–3.4 and Proposition 3.5 specify conditions under which the directional operators, contractions, and aggregations preserve the required transformation laws. Appendix G then connects these results to the layer operations. The explicit assumptions make the scope of the theoretical results easier to assess.

Strength 2: The experiments provide several complementary checks, and the dataset is a useful additional contribution.

Both DP-Learner and NextHAM use H(0) to predict self-consistent corrections, so their prediction tasks are aligned. Table 1 reports improvements on both benchmarks, while Table 4 shows that the full model outperforms the H(0) reference and all five architectural ablations. Appendix K reports three training seeds for the full model and each ablation, with Gauge MAE standard deviations below 0.005 meV. This supports the consistency of the reported architectural comparisons.

MP-HAM-110K also extends the available data for H(0)-conditioned Hamiltonian learning. Appendix J documents the electronic-structure calculation settings, element-dependent basis information, dataset size, and train/validation/test split. Its coverage of 68 elements and structures with varying cell sizes makes it a potentially useful benchmark beyond the proposed architecture.

Critical weaknesses

Weakness 1: The connection between the equivariant internal features and the final physical Hamiltonian needs a more explicit justification.

Appendix G.6 introduces an output module mapping the accumulated feature correction to the predicted Hamiltonian correction, but does not fully specify that mapping mathematically. Appendix K provides an important additional detail for SOC systems: the real parts of the predicted corrections to the two spin-conserving blocks are averaged and added to both blocks, while the remaining components are retained from H(0). Thus, the output processing is described in part; the remaining issue is how the complete readout connects to the transformation laws proved for the internal features.

In particular, preserving time-reversal symmetry does not by itself establish compatibility with the rotation and inversion actions used in the theory. The paper should explicitly show that the implemented output processing, including the SOC restriction and the extraction of the physical orbital blocks from the common layout, preserves the claimed equivariance. This would close the gap between the layer-level results and the statement following Eq. (109) about the final prediction.

I also could not identify an explicit description of how Hermitian reciprocity is enforced between reverse atom-pair blocks and opposite lattice translations, or how on-site blocks are made Hermitian. These conditions do not follow from rotational equivariance alone. This is a request to clarify the implementation and its mathematical guarantees, rather than evidence that the implementation necessarily violates them. The distinction matters because the resulting reciprocal-space Hamiltonian should be Hermitian for the intended physical interpretation of its bands.

Weakness 2: The reported gains need a more precise evaluation specification and clearer separation of architectural effects from implementation choices.

Appendix K provides several important details about Gauge MAE: it removes a common energy-reference offset per structure, uses an evaluation mask, takes the complex modulus for SOC errors, includes all four spin blocks, and averages results over test structures. Appendix I also derives the energy shift used for the joint training objective. However, the evaluation description does not explicitly specify whether the test-time offset is obtained from that same objective or from a separate criterion. The exact evaluation mask, normalization, treatment of reverse-pair duplicates, and handling of Hamiltonian blocks outside the 8 Å graph cutoff also remain unclear. An explicit formula or evaluation pseudocode would make the reported numbers reproducible.

The baseline setup is partly documented: the Materials-HAM-SOC comparison uses released NextHAM checkpoints, whereas NextHAM is adapted and trained from scratch on MP-HAM-110K. What remains unclear is the exact code version and modification set for the non-SOC adaptation, together with its optimizer, schedule, training duration, checkpoint selection, and seed protocol. The detailed 100-epoch, three-seed protocol is stated for DP-Learner and its ablations. Without comparable information for the retrained baseline, the 33.4% improvement on MP-HAM-110K is harder to attribute specifically to the architecture.

The scope of the SOC results should also be stated more prominently. Appendix K explicitly restricts the learned correction to the averaged real spin-conserving sector, and Section 6 explains that the reported metric includes spin-flip blocks with negligible errors. This does not invalidate the comparison, particularly if both methods use the same convention, but the aggregate MAE does not isolate performance on the learned components. A sector-wise breakdown would make the empirical claim easier to interpret.

Finally, the efficiency results measure complete deployed predictors under different configurations. NextHAM uses four sub-models across four GPUs, while DP-Learner uses one model on one GPU. The reported time includes input transfer, forward computation, output conversion, and saving; the memory figure sums peak reserved memory across NextHAM's four devices. These are useful measurements, and the paper discloses their definitions. However, they do not isolate forward-computation speed or per-device memory requirements. The headline efficiency claims should consistently reflect this measurement scope.

Questions for authors

Clarification requests (answers could change my assessment).

  1. Please specify the complete output mapping in Eq. (109), including extraction from the common orbital layout and the SOC post-processing described in Appendix K. Can you show that these steps preserve the rotation and inversion laws claimed for the final prediction?

  2. How are reverse-pair Hermitian reciprocity and on-site Hermiticity enforced? Please report residuals for these conditions and, if available, numerical rotation and inversion equivariance errors measured on the final post-processed Hamiltonians.

  3. What is the exact test-time Gauge MAE definition? In particular, how is the energy offset chosen, how is the evaluation mask constructed and normalized, and how are reverse-pair duplicates and blocks outside the graph cutoff handled? Is the offset criterion the same as the joint training criterion in Appendix I?

  4. For NextHAM on MP-HAM-110K, please provide the code version, non-SOC modifications, training hyperparameters, checkpoint rule, and number of seeds. How comparable is its training budget to that of DP-Learner?

  5. Can the SOC errors be reported separately for the learned real spin-conserving components and the components inherited from H(0)? Please also clarify whether NextHAM uses exactly the same output restriction in the reported comparison.

Suggestions for revision (not conditions for acceptance).

  1. Frame the novelty primarily in terms of the persistent feature layout and the permitted update operations. Eqs. (43)–(45) acknowledge that the complete direct-product and irreducible direct-sum representations have the same information content. The phrase “fundamental representation mismatch” should therefore be distinguished from a loss of representational information.

  2. Keep the complexity claims tied to the operator classes, coupling paths, multiplicities, and dense-computation assumptions specified in Appendix H. A compact statement of these assumptions alongside the main claims would help prevent an overly broad interpretation.

  3. Describe the reported timing as inference-pipeline latency and the memory result as aggregate peak reserved memory. A supplementary breakdown of transfer, forward computation, output processing, and saving, together with per-device memory and timing synchronization details, would clarify the sources of the measured gains.

  4. Extend the dataset documentation with the exact construction and extraction procedure for H(0), the reciprocal-space training-sample selection rule, and versioned structure identifiers and split files. Appendix J already provides substantial DFT and basis information; these additions would further support reconstruction of the benchmark.

  5. Figure 2 already shows the repeated modules and the H(0) bypass. Its clarity could be improved by explicitly labeling the features passed from layer k to layer k+1 and distinguishing initialization from operations repeated at every layer.

5 comments

We sincerely thank the reviewer for the careful reading of our manuscript and the thoughtful, constructive comments. The feedback has helped us strengthen the theoretical analysis and clarify how the formulation is implemented. We have addressed each point below and made revisions to our paper accordingly.

1. Symmetry and Hermiticity of the final Hamiltonian (Weakness 1 and Questions 1–2)

We thank the reviewer for identifying the need to connect the internal equivariant features to the final physical Hamiltonian. In our nonmagnetic non-SOC model, the predicted real-space Hamiltonian blocks are real and thus invariant under time reversal, which acts by complex conjugation in the orbital representation. Rotation and spatial inversion act by real matrices and preserve this property. In the SOC model, the correction is a complex orbital–spin matrix, and the readout averages the real parts of the two spin-conserving blocks. We must therefore explicitly establish that this operation enforces time-reversal symmetry while preserving spatial equivariance.

We have extended the theory to show that the SOC readout produces a time-reversal-invariant correction while preserving rotation and spatial-inversion equivariance. The physical initial Hamiltonian H(0)H^{(0)} used here is itself time-reversal symmetric and obeys the stated spatial transformation laws. Adding the correction to H(0)H^{(0)} preserves these properties, and reciprocal-block assembly then enforces exact Hermitian reciprocity. The assembled SOC Hamiltonian therefore satisfies Hermiticity and time-reversal symmetry jointly with rotation and spatial-inversion equivariance. Numerical tests further verify the spatial equivariance of the assembled output.

Theorem 3.6, “Symmetry preservation of the SOC Hamiltonian readout,” establishes the readout's symmetry guarantees, with a complete proof in Appendix F.1. Appendix K specifies Hermitian assembly in Eq. (160), and the derivation immediately after Eq. (161) shows why it preserves those guarantees. Appendix G.6 now specifies the output mapping from the original paper's Eq. (109) in Eqs. (128)–(129). The additions are marked in blue. Below, we explain how physical orbital extraction, the SOC readout, and Hermitian assembly connect the internal equivariant features to the Hamiltonian used for physical-property calculations.

a. Extracting the physical orbital blocks preserves equivariance.

The element-dependent mask selects each atom’s actual orbital basis, retaining complete angular-momentum multiplets and using the same selection for both spins. Let EξE_\xi be the orbital selection matrix and Cξ=Eξ⊗I2C_\xi=E_\xi\otimes I_2 its SOC lift. From the accumulated common-layout correction ΔFab(K)(R)\Delta F_{ab}^{(K)}(\mathbf R), the physical correction feature is

Fabout(R)=qab(R)CaΔFab(K)(R)Cb†.F_{ab}^{\mathrm{out}}(\mathbf R)=q_{ab}(\mathbf R) C_a \Delta F_{ab}^{(K)}(\mathbf R) C_b^\dagger.

Here qabq_{ab} is the optional real distance-dependent factor, with qab=1q_{ab}=1 when disabled. It is invariant under the symmetry actions considered here. Complete-multiplet selection gives

\qquad E_\xi\mathsf P_{\mathrm{common}}^{\mathrm{orb}} =\mathsf P_\xi^{\mathrm{orb}}E_\xi.$$ Thus extraction preserves the rotation and inversion transformation laws. Because $E_\xi$ is real and acts identically on both spins, its SOC lift also intertwines time reversal. The conversion between orbital-major and spin-major layouts is a fixed basis permutation, with the symmetry representations permuted consistently. **b. The SOC readout adds a real spin-scalar correction to $H^{(0)}$.** In orbital-major order, the implemented projection is $$\mathcal P_{\mathrm{SOC}}(F)=A(F)\otimes I_2, \qquad A(F)=\frac12\operatorname{Re} \left(F_{\uparrow\uparrow}+F_{\downarrow\downarrow}\right) =\frac12\operatorname{Re}\operatorname{Tr}_sF.$$ The prediction before Hermitian assembly is therefore $$\widetilde{H}_{ab}(\mathbf{R}) = H_{ab}^{(0)}(\mathbf{R}) + \mathcal{P}_{\mathrm{SOC}}\left(F_{ab}^{\mathrm{out}}(\mathbf{R})\right).$$ The same real orbital correction is added to both spin-conserving blocks. The spin-flip blocks and the imaginary parts of the spin-conserving blocks are retained from $H^{(0)}$. In the external spin-major layout, the correction is $I_2\otimes A(F)$. Spin tracing and taking real parts commute with the orbital selection in part a, as shown in Appendix F.1. For non-SOC systems, the readout directly adds the selected real orbital correction to $H^{(0)}$.

c. The projection is compatible with all three symmetry actions.

Theorem 3.6 establishes this compatibility from the implemented mapping. In the real spherical-harmonic basis, Dξ(R)D_\xi(R) is real and D~ξ(R)=Dξ(R)⊗U1/2(R)\widetilde D_\xi(R)=D_\xi(R)\otimes U^{1/2}(R). The common spin rotation cancels under the spin trace.

Tr⁡s(D~aFD~b†)=Da(Tr⁡sF)Db⊤.\operatorname{Tr}_s\left(\widetilde D_aF\widetilde D_b^\dagger\right)=D_a(\operatorname{Tr}_sF)D_b^\top.

Taking the real part commutes with the real orbital representations, so the projected correction obeys the same rotational law. The same argument applies to spatial inversion because the orbital parity matrices are real and spin is unchanged.

For time reversal,

TΘ(F)=JaF‾Jb†,Jξ=ILξ⊗iσy.\mathcal T_\Theta(F)=J_a\overline{F}J_b^\dagger,\qquad J_\xi=I_{L_\xi}\otimes i\sigma_y.

Time reversal complex-conjugates the spin trace.

Tr⁡sTΘ(F)=Tr⁡sF‾.\operatorname{Tr}_s\mathcal T_\Theta(F)=\overline{\operatorname{Tr}_sF}.

Taking its real part gives A(TΘF)=A(F)A(\mathcal T_\Theta F)=A(F). Since A(F)⊗I2A(F)\otimes I_2 is real and spin scalar, it is itself time-reversal invariant. Together, these results give

PSOC(TgF)=TgPSOC(F)for g=R,I,Θ.\mathcal P_{\mathrm{SOC}}(\mathcal T_gF)=\mathcal T_g\mathcal P_{\mathrm{SOC}}(F)\qquad\text{for }g=R,\mathcal{I},\Theta.

TΘPSOC(F)=PSOC(F).\mathcal T_\Theta\mathcal P_{\mathrm{SOC}}(F)=\mathcal P_{\mathrm{SOC}}(F).

Adding this correction to the physical H(0)H^{(0)} therefore preserves spatial equivariance and time-reversal symmetry. The hidden correction need not itself be time-reversal invariant. The spatial compatibility follows from the spin-trace and real-representation identities above, independently of the time-reversal argument.

d. Hermitian reciprocity is enforced before downstream calculations.

After loading the predictions for band structures and other physical-property calculations, we assemble each reverse pair using

Habphys(R)=12[H~ab(R)+H~ba(−R)†].H_{ab}^{\mathrm{phys}}(\mathbf R)=\frac12\left[\widetilde H_{ab}(\mathbf R)+\widetilde H_{ba}(-\mathbf R)^\dagger\right].

Applying this jointly to (a,b,R)(a,b,\mathbf R) and (b,a,−R)(b,a,-\mathbf R), with both members present in the block collection, enforces

Habphys(R)=Hbaphys(−R)†.H_{ab}^{\mathrm{phys}}(\mathbf R)=H_{ba}^{\mathrm{phys}}(-\mathbf R)^\dagger.

For a=ba=b and R=0\mathbf R=0, this is ordinary on-site Hermitian symmetrization. For a=ba=b and R≠0\mathbf R\ne0, the blocks at opposite lattice translations are paired. In exact arithmetic, the assembled reverse-pair residual Habphys(R)−Hbaphys(−R)†H_{ab}^{\mathrm{phys}}(\mathbf R)-H_{ba}^{\mathrm{phys}}(-\mathbf R)^\dagger and the on-site residual Haaphys(0)−Haaphys(0)†H_{aa}^{\mathrm{phys}}(\mathbf 0)-H_{aa}^{\mathrm{phys}}(\mathbf 0)^\dagger are both zero by construction. Consequently, the Fourier-assembled Hamiltonian satisfies Hphys(k)†=Hphys(k)H^{\mathrm{phys}}(\mathbf k)^\dagger=H^{\mathrm{phys}}(\mathbf k) for every k\mathbf k.

On the 3,000 Materials-HAM-SOC test structures, the maximum and mean absolute residuals for both reverse-pair reciprocity and on-site Hermiticity were zero after symmetrization, in FP32 and FP64, for both original and transformed inputs.

The derivation after Eq. (161) in Appendix K shows why Hermitian assembly preserves the symmetries of the prediction. For local unitary matrices UaU_a and UbU_b representing a rotation or spatial inversion,

(UbH~baUa†)†=UaH~ba†Ub†,\left(U_b\widetilde H_{ba}U_a^\dagger\right)^\dagger=U_a\widetilde H_{ba}^\dagger U_b^\dagger,

so the adjoint of the reverse block transforms in the same way as the original block, and their average preserves spatial equivariance. Time reversal commutes with taking the adjoint and with averaging using real weights, so the assembly also retains time-reversal symmetry whenever the prediction has it before assembly.

In our original workflow, we applied this symmetrization before computing downstream physical properties. We have now added it to the inference code so users can obtain Hermitian Hamiltonians at inference. Applying symmetrization before scoring reduces DP-Learner's test Gauge MAE from 0.841 to 0.806 meV on Materials-HAM-SOC and from 1.595 to 1.541 meV on MP-HAM-110K. Nevertheless, NextHAM computes Gauge MAE before Hermitian symmetrization. To compare the two models on equal terms, we report DP-Learner’s original Gauge MAEs of 0.841 and 1.595 meV in Table 1, with both models evaluated at the same stage of inference.

e. Numerical verification of spatial equivariance of the final SOC Hamiltonian.

We tested the complete prediction pipeline on 3,000 structures from the Materials-HAM-SOC test set. For each structure, we first predicted the final Hamiltonian from the original geometry and H(0)H^{(0)}. We then jointly transformed the geometry and H(0)H^{(0)} by spatial inversion combined with a random rotation in SO(3)\mathrm{SO}(3) and repeated the prediction. The resulting Hamiltonian was compared with the correspondingly transformed prediction from the original input.

In this test, both paths used the time-reversal symmetry restriction and Hermitian assembly described above, so this test evaluates the final Hamiltonian used for downstream calculations. The same FP32-trained model was evaluated in both FP32 and FP64 to assess the effect of numerical precision. For each precision, we pooled the absolute elementwise equivariance errors across all Hamiltonian matrix entries in all 3,000 test structures and computed their global maximum and mean.

Evaluation precisionMaximum equivariance error (eV)Mean equivariance error (eV)
FP325.896×10−65.896\times10^{-6}2.520×10−62.520\times10^{-6}
FP644.758×10−154.758\times10^{-15}6.031×10−176.031\times10^{-17}

The error magnitudes and their decrease at higher precision are consistent with floating-point roundoff. These results support the rotation and spatial-inversion equivariance of the complete prediction pipeline.

2. Training and evaluation details (Weakness 2 and Questions 3–5)

a. Energy-reference alignment in training and evaluation.

We thank the reviewer for asking us to clarify the evaluation procedure. Hamiltonians must be compared with a common energy reference. In the nonorthogonal atomic-orbital basis used here, changing the energy zero corresponds to H↦H+μSH\mapsto H+\mu S, where SS is the overlap matrix. This shifts all generalized eigenvalues by μ\mu while preserving eigenvectors, band gaps, and other energy differences. Gauge alignment accounts for this freedom before measuring the Hamiltonian error.

Training and evaluation determine the offset using different criteria. During training, the offset μ\mu minimizes the joint real-space and reciprocal-space objective in Eq. (158), with its analytical solution given in Eq. (159) of Appendix I. For evaluation, we follow the gauge alignment of Wang et al. [1] as implemented in the released NextHAM evaluator [2]. We fit one offset in real space by least squares and then compute the MAE of the aligned Hamiltonians.

For each test structure, the evaluation offset is

μeval=∑R,αβRe⁡[conj⁡([Hpred(R)−Hgt(R)]αβ)S(R)αβ]∑R,αβ∣S(R)αβ∣2+ϵμ.\mu_{\mathrm{eval}} = \frac{\displaystyle\sum_{\mathbf{R},\alpha\beta} \operatorname{Re}\left[\operatorname{conj}\left(\left[H^{\mathrm{pred}}(\mathbf{R})-H^{\mathrm{gt}}(\mathbf{R})\right]_{\alpha\beta}\right) S(\mathbf{R})_{\alpha\beta}\right]}{\displaystyle\sum_{\mathbf{R},\alpha\beta}\left|S(\mathbf{R})_{\alpha\beta}\right|^2+\epsilon_{\mu}}.

Here R\mathbf{R} denotes the lattice translation, and α,β\alpha,\beta index the matrix entries of the stored real-space Hamiltonian blocks, following Appendix I. This single real scalar aligns the energy reference of the entire structure. We then compute the aligned MAE over the entries selected by the evaluation mask.

MAEG=∑R,αβMαβ(R)∣[Hpred(R)−Hgt(R)−μevalS(R)]αβ∣∑R,αβMαβ(R)+ϵMAE.\mathrm{MAE}_{\mathrm{G}} = \frac{\displaystyle\sum_{\mathbf{R},\alpha\beta} M_{\alpha\beta}(\mathbf{R})\left|\left[H^{\mathrm{pred}}(\mathbf{R})-H^{\mathrm{gt}}(\mathbf{R})-\mu_{\mathrm{eval}}S(\mathbf{R})\right]_{\alpha\beta}\right|}{\displaystyle\sum_{\mathbf{R},\alpha\beta} M_{\alpha\beta}(\mathbf{R})+\epsilon_{\mathrm{MAE}}}.

The sums use the stored blocks of the 8 Å graph, including on-site blocks. Each stored directed block contributes once. When both (a,b,R)(a,b,\mathbf{R}) and (b,a,−R)(b,a,-\mathbf{R}) are present, both are counted. Blocks outside the graph cutoff are excluded from the fit and the MAE. Within the evaluation graph, a block outside the learned correction range is scored using its H(0)H^{(0)} value. Following the setting of NextHAM, we compute the reported Gauge MAE on these predictions before the downstream Hermitian assembly in part 1(d).

The mask Mαβ(R)M_{\alpha\beta}(\mathbf{R}) selects the physical orbital entries used in the MAE. The least-squares fit uses the stored Hamiltonian and overlap arrays before the MAE mask is applied. The numerical stabilizers are ϵμ=10−6\epsilon_{\mu}=10^{-6} and ϵMAE=10−7\epsilon_{\mathrm{MAE}}=10^{-7}.

For SOC systems, α,β\alpha,\beta include both orbital and spin indices, so the metric includes all four spin blocks. Each entrywise error is measured by its complex modulus. The overlap is spin diagonal, so the common energy shift acts identically on the two spin-conserving blocks and leaves the spin-flip blocks unchanged.

Let MAEG(s)\mathrm{MAE}_{\mathrm{G}}^{(s)} denote the Gauge MAE of test structure ss. We first compute this metric within each structure and then average over test structures.

MAE‾G=1Ntest∑s=1NtestMAEG(s).\overline{\mathrm{MAE}}_{\mathrm{G}} = \frac{1}{N_{\mathrm{test}}}\sum_{s=1}^{N_{\mathrm{test}}}\mathrm{MAE}_{\mathrm{G}}^{(s)}.

Each structure therefore receives equal weight in the reported result. Values computed in eV are converted to meV. DP-Learner, NextHAM, the architectural ablations, and the H(0)H^{(0)} reference are evaluated using this same protocol, ensuring a consistent comparison after alignment of the energy reference.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Letter to AC

Dear AC,

We thank the reviewers for their detailed feedback. During the discussion, we added the requested experimental comparisons, extended the theoretical analysis, clarified technical details, and revised the manuscript. Our responses to the individual reviewers provide the full details. For the Area Chair’s convenience, we summarize the most important points below.

  1. Hamiltonian MAE with the same four distance ranges. At the reviewer’s request, we report the training and evaluation details for four DP-Learner submodels, one for each of the four distance ranges used by NextHAM. Under this matched partition, DP-Learner reduces the mean Gauge MAE on the testing set from 1.417 to 0.893 meV on Materials-HAM-SOC and from 2.394 to 1.638 meV on MP-HAM-110K, reductions of 37.0% and 31.6% relative to NextHAM. The accuracy advantage therefore persists when both architectures use the same distance partition. Within DP-Learner, however, splitting the model by distance lowers the MAE on the training set but raises it on the testing set on both benchmarks, consistent with greater overfitting. We therefore use a single DP-Learner across all distance ranges as the default in the paper.

  2. Near-Fermi band-energy MAE in reciprocal space. On the testing set of Materials-HAM-SOC, we compare bands at the same band index and kk point on a 10×10×1010\times10\times10 mesh. With the DFT Fermi energy as the common energy zero, DP-Learner reduces the mean near-Fermi band-energy MAE by 72.2% relative to NextHAM. After an additional per-structure, median-based rigid shift of the predicted bands, the reduction is 28.7%. The gain in real-space Hamiltonian accuracy thus carries through to the relative positions of the bands in reciprocal space.

  3. Results on QH9-Stable-ID. We report the results of a molecular version of DP-Learner that retains the direct-product message-passing architecture. With one Euler step, DP-Learner achieves a Hamiltonian MAE of 6.4 µHa, compared with 7.1 µHa for QHFlow2-M using three Euler steps. DP-Learner also achieves lower MAEs for occupied-orbital energies, HOMO and LUMO energies, and the HOMO–LUMO gap in this comparison, with 25.30M trainable parameters compared with 44.25M for QHFlow2-M.

  4. Rotation and spatial-inversion equivariance of the final SOC Hamiltonian. Theorem 3.6 shows that the SOC readout preserves rotation and spatial-inversion equivariance and produces a time-reversal-invariant correction. The initial Hamiltonian H(0)H^{(0)} used in our nonmagnetic SOC calculations is itself time-reversal symmetric, so adding the correction to H(0)H^{(0)} yields a time-reversal-symmetric Hamiltonian. We also tested the complete pipeline, including Hermitian assembly, on 3,000 Materials-HAM-SOC structures under a random rotation combined with spatial inversion. The maximum absolute equivariance error of the final Hamiltonian is 5.896×10−65.896\times10^{-6} eV in FP32 and 4.758×10−154.758\times10^{-15} eV in FP64, confirming spatial equivariance numerically.

  5. Hermiticity of the assembled output. Appendix K shows how symmetrizing each reciprocal block pair enforces Hermiticity while preserving the spatial and time-reversal symmetry guarantees. On the same 3,000 test structures, we numerically verified the Hermiticity of the assembled Hamiltonian for both original and transformed inputs. Together, Theorem 3.6 and Appendix K establish that the final SOC Hamiltonian can simultaneously satisfy Hermiticity, rotation and spatial-inversion equivariance, and time-reversal symmetry in the nonmagnetic setting studied here.

  6. Component-wise inference time and full inference-pipeline latency. Our responses report component-wise timings and trainable parameter counts for DP-Learner on Materials-HAM-SOC. We also profiled NextHAM’s edge_lin, the layer that maps its final edge features to the irreducible coefficients of the Hamiltonian correction. This module alone takes 386.755 ms per structure per submodel by CUDA-event timing, averaged across the four A800 GPUs. DP-Learner’s entire forward pass on one A800 takes 86.875 ms by CUDA-event timing and 170.264 ms by host wall-clock timing. By the same CUDA-event measure, NextHAM’s output projection alone takes longer than DP-Learner’s full forward pass. For the full inference pipeline, including input transfer, network execution, output conversion, and saving to disk, Table 1 reports a speedup of up to 11.2× over NextHAM in mean latency per structure.

  1. SOC Hamiltonian MAE by spin block. Both DP-Learner and NextHAM correct the real parts of the two spin-conserving blocks and retain the remaining components from H(0)H^{(0)}. For DP-Learner, the real-part MAE in each learned block is 1.682 meV. Every retained real or imaginary component has an MAE below 0.001 meV. This breakdown shows how the learned correction and the physical initialization each contribute to the aggregate Gauge MAE of 0.841 meV.

  2. What DP-Learner passes between layers and how each block is updated. The revised Section 4 and Figure 2 describe how bond direction, attention, and atomic context update an atom-pair block while preserving its two orbital or orbital–spin indices. Algorithm 1 in Appendix G.1 presents one complete forward pass from H(0)H^{(0)} to the predicted Hamiltonian. In responses to Reviewer 6nkG, we clarify DP-Learner’s theoretical and methodological contributions compared to BOA [1] by distinguishing the prediction tasks, the features each method propagates, and the symmetries each addresses.

[1] Klockow et al. A Function-Centric Graph Neural Network Approach for Predicting Electron Densities. ICLR, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.