acceptodds
Under review as a conference paper at ICLR 2027

Physics-Guided Learning of Kohn-Sham Hamiltonians for Periodic Solids

Abstract

Deep learning prediction of Kohn-Sham Hamiltonians overcomes the computational bottleneck of density functional theory (DFT) in large-scale electronic structure calculations. Most existing methods, however, fit Hamiltonians numerically without fully using the rich quantum-mechanical priors in DFT. We propose two methods guided by physical priors: the HSE Hamiltonian model H2Net and the variational surrogate loss (VSLoss). Inspired by the separation of short- and long-range interactions in the HSE functional, H2Net uses a dual-branch architecture. Fewer layers in the long-range branch limit the receptive field and reduce training memory requirements. The model also incorporates the non-self-consistent part of the Hamiltonian and a PBE Hamiltonian from a pretrained model as input features, explicitly introducing physical priors to improve prediction accuracy. VSLoss guides the model toward minimizing the ground-state energy, bringing the Hohenberg–Kohn variational principle into training to improve the accuracy of downstream physical properties calculated from the Hamiltonian. By improving model representations and optimization objectives, respectively, the two methods make comprehensive use of DFT priors from basis sets, pseudopotentials, functionals, and the variational principle. Together, they offer a new approach to learning Kohn–Sham Hamiltonians for periodic solids.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Kohn-Sham Hamiltonian Prediction, Physics-Informed Machine Learning, Equivariant Graph Neural Networks, Electronic Structure Calculations, Density Functional Theory, Scientific Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nFWbOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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.

  3. 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 HKS(P^)H^{\mathrm{KS}}(\widehat P) by the reference HH 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.

  4. 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.

  5. 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
  1. 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.

  2. 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.

  3. 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.

  4. 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.

Reviewer_R5mROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes using inductive biases to improve hamiltonian prediction. They suggest a novel equivariant architecture, a "multi-graph" representation, and a variational loss. The proposals attempt to improve representation, decrease memory footprint, and improve performance respectively. The authors also provide a small dataset on which they evaluate their models and a few simple benchmarks.

Critical strengths
  • Numerical results on extremely small self-made benchmarks are an improvement against a two baseline models.
  • There are a number of inductive biases they suggest and a seemingly novel low-cost equivariant architecture that are novel to my knowledge.
Critical weaknesses
  • The material is strewn throughout confusing sections that blend into one another. I was left without a clear image of the independent pieces proposed by this work. How does EGMF relate to H2Net and why are they not applied at the same time? How does the VSLoss relate as well? It feels like a pulp of different ideas. Unless this can be clarified and unified into a proposal, I don't think it can be accepted.
  • The comparisons, datasets, and baselines are all fairly small scale and on extremely limited numbers of examples.
Questions for authors
  • In the preamble I was not expecting you to introduce a new architecture. You described priors and I expected some kind of pre-training or something else... Perhaps you can be more explicit and terse in the intro about what you introduce and why.
  • I don't understand why your architecture accommodates basis set mismatch. The notion of comparing invariants for losses and errors would indeed do this, but that is a property of the error computation... Not the architecture, right? Explain and untangle this please.
  • Furthermore on the same subject, you write "When their maximum angular momentum l is the same but their numbers of irreducible representations differ, the equivariant linear layer automatically aligns the channels." What do you mean by this? Are you trying to align your invariant features or align the basis dependent features but across different bases? Perhaps use of clear notation is called for here.
  • Equation (4) is preceded by "Apart from HF exchange and the nonlocal pseudopotential, these terms can be written as..." but aren't you introducing exactly those terms? I am further confused by this.
  • The edges, Ein,Eout\mathcal{E}_{in}, \mathcal{E}_{out} which seem to be important to your method, are confusingly described in the main text. The key feature of how they interact is omitted and left for the appendix. Instead you kept some quantitative information about the relative size of the graphs, but that seems very uninformative to me.
  • Manasa Kaniselvan, et. al. 2026's model and follow up work are applicable across the periodic table, despite being trained in a non-periodic setting. Although direct comparison can be hard due to different functionals, I would think it would make sense to mention this work.
  • Why do you introduce your ideas in a different order than you evaluate them in? I find this confusing.
Reviewer_QJuqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces matrix-derived input features and a dual-branch equivariant architecture for HSE Hamiltonian prediction. A deeper inner branch and a shallower extended-range branch reduce training memory. It also proposes VSLoss, which uses the occupied states of the predicted Hamiltonian in a trace with the reference Hamiltonian, together with matrix and band supervision.

Critical strengths

The paper is trying to approach a practical difficulty in learning HSE Hamiltonians. The interesting dual-branch design is physically motivated and gives improvements in both prediction accuracy and training memory, including when the baseline can also use a deeper network. I also find the the evaluation of band and wavefunction errors in addition to Hamiltonian MAE quite useful.

Critical weaknesses

That said, I find the benchmark is not critical enough in this paper and also the innovation is quite incremental.

Questions for authors
  1. It is true that hybrid functional has some intrinsic difficulties, but we don't see enough envidencve from this paper, the authors only studied a few chemical systems, I want to see more broader benchmarks.
  2. In principle, the hydride functional's difficulty comes from its non-locality in its definition. In the approach proposed here, the authros used some inner and outer modules in the model. So, my question is, if we only apply the outer module to the existing Hamiltonian models, do we get similar improvements. Also, what if we apply the method proposed here to a non-hybrid functional's results? Do you also improve the results computed with, for example, PBE?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.