- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper introduces H2Net, an equivariant Hamiltonian model with a deep inner branch and a shallow outer branch motivated by the interaction ranges of the HSE functional. Equivariant gated matrix features (EGMF) inject physical matrices into initial node and edge features, including a non-self-consistent Hamiltonian and PBE predictions from Uni-HamGNN. The proposed VSLoss evaluates the predicted occupied density matrix under the reference Hamiltonian and is used alongside matrix and band losses during fine-tuning. Experiments cover three HSE datasets and four PBE datasets, with additional comparisons on the original DeepH-hybrid datasets. The reported benefits include lower training memory, improved matrix accuracy, and better downstream band and wavefunction metrics in several settings.
Critical strengths
-
The asymmetric architecture addresses a concrete computational problem. In Table 2, H2Net reduces training memory by 24.3%–40.5% while lowering Hamiltonian MAE on all three HSE datasets. For MoS2, the error falls from 0.082 to 0.064 meV. Table 1 also demonstrates that the model can accommodate a deeper network with f orbitals within an 80 GB budget, where the multilayer HamGNN baseline runs out of memory.
-
The ablations establish contributions from both the architecture and matrix inputs. Table 6 compares HamGNN, H2Net without EGMF, and the full model. On SiO2, their errors are 0.509, 0.256, and 0.228 meV, respectively. Table 3 tests EGMF with two backbones on four datasets, and Table 5 examines the input matrix and injection location. This is substantially more informative than comparing only the complete model against a coordinate-based baseline.
Critical weaknesses
-
The added machine-learning contribution is not sufficiently isolated from existing physical-input and residual-learning strategies. Hybrid-functional Hamiltonian prediction is already established by DeepH-hybrid [1], and NextHAM [2] uses initial Hamiltonian blocks as node and edge descriptors together with residual prediction. EGMF combines a CG decomposition, equivariant linear projections, and residual injection. Its global scalar gates can be absorbed into the learned projections; their role in optimization is not evaluated. Table 3 compares EGMF with residual-only learning, omitting the combination of matrix inputs and a nonlinear residual correction. The main HSE comparisons also lack a single-branch baseline receiving the same physical inputs. These controls would assess the proposed full model against stronger uses of the available prior information.
-
The closest loss baselines are missing. Appendix D explicitly relates VSLoss to occupied–unoccupied coupling penalties and discusses WALoss [3] and NextHAM's reciprocal-space objective [2]. Table 4 compares only matrix supervision, BandLoss, VSLoss, and their combination. It does not establish whether the predicted-projector trace objective improves on existing reference-eigensystem supervision. A comparison from the same pretrained checkpoint, with matched fine-tuning budgets, would be more informative than further comparisons against matrix loss alone.
-
The physical interpretation of VSLoss needs to be narrower. Equation (11) is a variational objective for occupied states under a fixed reference Hamiltonian. Its minimum identifies the reference occupied subspace when a gap is present. It leaves orbital energies and rotations within the occupied or virtual subspaces unconstrained. The replacement of by the reference also omits self-consistent density-response terms; proximity to the reference alone does not establish that the resulting gradient accurately follows the true DFT energy gradient. The Ritz bound supports occupied-subspace supervision. It does not establish general ground-state energy minimization or individual-orbital accuracy. These distinctions should be maintained in the main text and Appendix D.
-
The wavefunction evaluation does not directly test the quantity controlled by the loss, and statistical reliability is unclear. Appendix F compares energy-ordered individual eigenvectors, including five unoccupied bands, through a phase-invariant overlap. This metric remains sensitive to rotations within degenerate subspaces and to near-degenerate band reordering. No occupied-projector or density error is reported. The number of independent training runs and uncertainty across runs are also unspecified. This matters for small differences: on HSE MAPbI3, joint fine-tuning changes band MAE from 13.12 to 13.09 meV and wavefunction dissimilarity from 5.56% to 5.52% (Table 7). The objectives also produce trade-offs: on PBE MAPbI3, BandLoss alone gives a lower band MAE than the joint loss, 9.61 versus 12.32 meV.
-
The efficiency evidence covers target-model training memory, not the complete workflow. There are no runtime measurements for matrix preparation, Uni-HamGNN inference, H2Net prediction, or VSLoss fine-tuning. Batch size, numerical precision, and the memory measurement procedure are not specified. The reported memory savings are useful, but their scope needs to be clear. In particular, the extra PBE model and the eigendecomposition used during fine-tuning introduce costs that should be accounted for separately. Table 1 demonstrates greater depth within a fixed memory budget; its H2Net configurations use more memory than the one-layer baseline.
[1] Tang et al. A deep equivariant neural network approach for efficient hybrid density functional calculations. Nature Communications 15, 8815 (2024).
[2] Yin et al. Advancing Universal Deep Learning for Electronic-Structure Hamiltonian Prediction of Materials. ICLR 2026.
[3] Li et al. Enhancing the Scalability and Applicability of Kohn-Sham Hamiltonians for Molecular Systems. ICLR 2025.
Questions for authors
-
On one representative dataset, please compare VSLoss with a closely related existing objective, such as a reciprocal-space WALoss or NextHAM-style subspace loss, starting from the same pretrained checkpoint and using the same fine-tuning budget.
-
Please clarify the incremental value of EGMF over matrix inputs combined with residual prediction. A focused comparison using the same backbone and input matrices would help. For H2Net, specify the Uni-HamGNN checkpoint, whether it is frozen, its pretraining-data overlap with the evaluation sets, and the mapping between the input and target orbital bases.
-
Using the existing test predictions, please report an occupied-subspace error and separate occupied and unoccupied wavefunction results. Confirm that all models use the same evaluation k points. State the number of training runs, checkpoint-selection rule, and whether the loss variants share their starting checkpoint and fine-tuning budget.
-
Please report representative preprocessing, inference, and fine-tuning timings, and define the memory measurement protocol. Clarify code and dataset availability. The discussion of VSLoss should separate the fixed-reference Ritz guarantee from the approximation to true self-consistent energy optimization.