- 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
-
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.
-
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
-
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?
-
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?
-
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?
-
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
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 measurement | Corresponding computation | Mean time (ms/structure) | Share of CUDA-event time | Trainable parameters |
|---|---|---|---|---|
| Fixed spin–orbital layout permutations | Reorder spin and orbital indices at input and at the correction output (Appendix K). | 0.179 | 0.21% | 0 |
| Invariant descriptors and feature assembly | Compute radial features and subblock Frobenius norms, then assemble the invariant inputs (steps 2–3, 7, and 14). | 10.007 | 11.52% | 0 |
| Invariant MLPs, excluding attention | Initialize 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). | 22.423 | 25.81% | 21,035,481 |
| Directional operator construction | Combine the bond direction with the rotation generators to construct and (step 9, Eq. (110)). | 10.514 | 12.10% | 0 |
| Dense matrix products | Apply the two-sided directional product and the left, right, and two-sided context products (steps 10 and 16, Eqs. (113) and (124)). | 14.240 | 16.39% | 0 |
| Attention and aggregation | Predict attention logits, project values, and aggregate incoming edge features (steps 8 and 11–12, Eqs. (114)–(116) and (118)–(119)). | 14.698 | 16.92% | 5,409,864 |
| Gating and residual updates | Apply the subblock and edge gates, scale the residuals with LayerScale, and accumulate the updates (steps 6, 9–10, 12–13, and 16–18). | 11.231 | 12.93% | 36 |
| Other forward work and timing overhead | Remaining time within the profiled forward interval. | 3.580 | 4.12% | 0 |
| Full forward interval on the CUDA stream | Model forward through the orbital–spin correction output. | 86.875 | 100% | 26,445,381 |
| Forward-call latency on the host | Wall-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 , the input and output irreps both use the full irreducible decomposition of the 729-dimensional space , where . 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 to over the first five epochs, then follows cosine decay to 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 , 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 for both predictions and references at distances of 6 Å or more. The baseline uses the self-consistent 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 mesh of 1,000 points. At each 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 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 alignment | Hamiltonian | Mean MAE (eV) ↓ | eV | eV | eV | eV |
|---|---|---|---|---|---|---|
| Before additional rigid alignment | 0.5236 | 26.30% | 28.96% | 31.56% | 13.16% | |
| Before additional rigid alignment | NextHAM | 0.3879 | 18.03% | 50.66% | 28.26% | 3.03% |
| Before additional rigid alignment | DP-Learner | 0.1077 | 74.96% | 21.73% | 2.23% | 1.06% |
| After additional rigid alignment | 0.4832 | 27.40% | 29.26% | 30.30% | 13.03% | |
| After additional rigid alignment | NextHAM | 0.1241 | 63.46% | 33.10% | 2.96% | 0.46% |
| After additional rigid alignment | DP-Learner | 0.0885 | 79.06% | 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 % relative to NextHAM. After an additional per-structure, median-based shift of the predicted band energies, the reduction is %. The fraction of structures with an MAE below eV rises from % to % with the common Fermi reference alone, and from % to % 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 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.
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) and (b,a,−R) using
Habphys(R)=21[Hab(R)+Hba(−R)†].
For an on-site block (a=b and 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. 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) 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) and apply the transformation consistently to the atomic positions, periodic lattice vectors, atom-pair displacements, and 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.
For both outputs, the maximum absolute equivariance error under combined rotation and spatial inversion is below 6×10−6 eV in FP32 and falls below 5×10−15 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.