acceptodds
Under review as a conference paper at ICLR 2027

Improving Kohn–Sham Hamiltonian Learning with Grassmannian Supervision

Abstract

Predicting the Kohn–Sham Hamiltonian in density functional theory has become an important task in AI for science, as it can accelerate electronic structure calculations and enable the prediction of multiple molecular properties from a single model. However, element-wise Hamiltonian regression does not explicitly enforce the accuracy of the occupied orbital subspace, which determines the density matrix from which ground-state observables are evaluated. This subspace therefore provides a natural auxiliary target for improving downstream property predictions. Yet, directly supervising orbital coefficients in an atomic orbital basis presents two challenges: (1) Gauge redundancy. Rotations among occupied orbitals preserve their subspace and density matrix, making the coefficient representation non-unique. (2) Gradient instability. Orbital-based losses require differentiation through an eigensolver, whose eigenvector derivatives can become unstable near degeneracies. To address these challenges, we propose GrassDist, a principled framework that aligns predicted and reference occupied subspaces using a Grassmannian geodesic loss on the Grassmann manifold. GrassDist is invariant to gauge transformations among occupied orbitals by construction, hence resolving the ambiguity in coefficient supervision. We additionally incorporate two gradient-stabilization treatments to ensure stable optimization. Across multiple benchmarks, GrassDist achieves substantial improvements over various downstream properties, while enabling stable training and faster convergence. On QH9, it reduces density matrix and total energy errors by one to two orders of magnitude compared with the Hamiltonian-only baseline and WALoss. GrassDist also exhibits strong data efficiency: on MD17 uracil, using only 10% to 20% of the training data already matches or surpasses the full-data baseline. The resulting model further generalizes to molecules larger than those seen during training. These results establish GrassDist as a significant methodological advance toward reliable and scalable molecular modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Chemistry, Density Functional Theory, Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_aLNROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adds an auxiliary loss to Kohn–Sham Hamiltonian prediction. The predicted Hamiltonian is diagonalised during training, the occupied eigenvectors are Löwdin-orthonormalised, and the squared Grassmannian geodesic distance to the reference occupied subspace is penalised. A Taylor expansion of arccos$^2$ near σ=1\sigma = 1, a small-gap truncation of the eigensolver backward pass and an auxiliary orbital-energy loss LϵL_\epsilon complete the method ("GrassDist"). With SPHNet and QHNet backbones, the paper reports large reductions in density-matrix and non-self-consistent total-energy errors on QH9, smaller gains on MD17, and much better size extrapolation to QMugs molecules with 10–17 heavy atoms, compared with element-wise Hamiltonian regression and WALoss.

Critical strengths
  1. Well-motivated target. Element-wise Hamiltonian error is a poor proxy for the quantities users care about, and the occupied subspace is the right gauge-invariant object. The outlier case study (App. C.7) makes this concrete: two molecules have a HOMO/LUMO occupation swap (d_Gr ≈ pi/2) and energy errors above 1000 kcal/mol.
  2. Large empirical effect on density-derived quantities. On all four QH9 splits and both backbones, d_Gr, P, μ\mu and E_tot improve substantially, median E_tot error drops by 4–12×. The QMugs transfer result (Table 8) is significant: both baselines collapse beyond 13 heavy atoms while GrassDist's quality degrades smoothly.
Critical weaknesses

W1. The paper does not isolate what its title claims: the benefit of the Grassmannian geodesic formulation, as opposed to any occupied-subspace (density-matrix) supervision.

  • The geodesic loss coincides with simpler losses. For small principal angles:

    • the geodesic loss is Σθ2\Sigma \theta^2;
    • the chordal loss is Σsin⁡2θ=12‖X^X^T−X∗X∗T‖F2\Sigma \sin^2 \theta = \frac{1}{2} ‖\hat{X}\hat{X}^T − X^*X^{*T}‖_F^2. This is a plain Frobenius loss on the density matrix in the Löwdin basis, and needs no SVD, arccos or Taylor patch;
    • the Procrustes-aligned coefficient loss is min⁡Q‖X^−X∗Q‖F2=2Σ(1−σi)\min_Q ‖\hat{X} − X^*Q‖_F^2 = 2\Sigma(1 − \sigma_i). This is exactly the leading term of Eq. 9.

    They differ only at relative order θ2\theta^2. At the reported d_Gr ≈ 5×10⁻³ (2–7×10⁻² for the baseline), every angle lies deep inside the u<10−3u < 10^{-3} Taylor branch, and the three losses agree to better than 0.1%. They diverge materially only near θ≈π2\theta \approx \frac{\pi}{2}, i.e. for rare occupation swaps.

  • The gauge-redundancy ablation (Sec. 3.3) is weak.

    • Coefficient losses already cut E_tot from 4.87 to 0.18–0.19 kcal/mol. Most of the gain therefore comes from any orbital-level supervision.
  • Closely related work is not discussed. This includes:

    • direct one-electron density-matrix learning (e.g. Shao et al., 2023);
    • self-consistency training (Zhang et al., ICML 2024), cited only for the truncation trick;
    • DEQH (NeurIPS 2024);
    • prior Grassmann treatments of occupied subspaces in quantum chemistry (Van Voorhis & Head-Gordon, 2002; Polack et al., JCTC 2021).

As it stands, the evidence supports "supervising the density matrix through the eigensolver helps a lot". It does not support "the Grassmannian geodesic helps". A density-matrix (chordal) loss and a Procrustes-aligned coefficient loss, trained to convergence under the main protocol, are the decisive missing baselines.

W2. The headline gains are inflated by the evaluation protocol and by an additional loss term.

  • Two molecules account for the "over 100-fold" improvement.

    • E_tot is evaluated on the first 800 test configurations.
    • The two Table 11 outliers (2213.94 and 1321.25 kcal/mol) are in that set, as Fig. 1a shows. Together they contribute ≈4.42 of the baseline's 4.87 kcal/mol mean (≈91%). This suggests the 800-configuration subset is not representative.
  • An extra loss term confounds Table 1. All main results use (wH,wGr,wϵ)=(1,10,0.25)(w_H, w_{Gr}, w_\epsilon) = (1, 10, 0.25), but LϵL_\epsilon appears only in App. B.2. Table 7 shows that LGrL_{Gr} does not work on its own without LϵL_\epsilon, so its role should be clearly stated.

  • Comparisons are not equally rigorous.

    • The default weights appear to be chosen from test-split metrics (Tables 5, 6, 10).
    • WALoss uses a weight tuned in its original paper for a different dataset and backbone.
    • No training cost is reported for the per-step float64 eigendecomposition and SVD, although the paper claims "scalable".
Questions for authors
  1. Under the main protocol (residual parameterisation, 200k steps, test split, ≥3 seeds), how do (a) the chordal/density-matrix Frobenius loss in the Löwdin basis, (b) a Procrustes-aligned coefficient loss and (c) the geodesic loss without the Taylor branch compare with GrassDist on all Table 1 metrics and on QMugs? What fraction of principal angles ever takes the arccos branch during training?
  2. What are the stable-iid baseline E_tot, P and d_Gr means with the two Table 11 molecules excluded? Please report medians, 5/95 percentiles and the fraction of occupation-swap failures (any θi>π4\theta_i > \frac{\pi}{4}) for every metric and split, on a common, randomly drawn evaluation set. How are the "first 800" configurations distributed over molecules in the dynamic splits?
  3. Please add the H + LϵL_\epsilon case (1:0:0.25) to Table 1, QMugs, MD17 and the QHNet ablation. Which weights were used on MD17?
  4. Why does the purely CoccC_{occ}-MAE-trained model have higher CoccC_{occ} MAE than the Hamiltonian-only baseline?
  5. Were wGr,wϵw_{Gr}, w_\epsilon and the coefficient weights selected on validation or test metrics? What exactly is the validation loss used for checkpoint selection for each method? Was the WALoss weight re-tuned on QH9?
  6. What is the wall-clock time and memory per training step relative to the baseline, and how does it scale with the number of basis functions?
  7. How is E_tot obtained — a single non-self-consistent energy evaluation at $\hat{P}?
  8. Why are the MD17 and QHNet baselines so much better than the published numbers, and why is ethanol 3–4× worse than the larger molecules?
  9. The paper discusses QHFlow2, which reports accurate non-self-consistent energies on QH9 from Hamiltonian-only training. How does GrassDist compare, or combine, with such architecture-side improvements?
17 comments
AuthorsOpenReview ↗

Response to W1 & Q1: Grassmannian geodesic formulation and occupied-subspace supervision.

We thank the reviewer for the comprehensive and constructive feedback. We address your concerns point by point below.

  1. Local equivalence of occupied-subspace losses. We would like to clarify that our central contribution lies in demonstrating the benefit of occupied-subspace supervision for downstream physical properties, with the Grassmannian geodesic loss providing a natural geometric formulation. We agree with your mathematical analysis that geodesic, chordal, and Procrustes losses are locally equivalent for small principal angles. However, local equivalence near convergence does not imply equivalent optimization dynamics or out-of-distribution generalization. Our preliminary comparison in Appendix C.14 showed that geodesic loss converges faster than chordal loss during early training without the MINAO-residual parameterization. To further examine their performance after convergence, we now conduct a controlled comparison using the main MINAO-residual parameterization and matched supervision. Results on QH9-stable-iid : | Method | H (10⁻⁶ Eh) | d_Gr (10⁻³) | P (10⁻³) | E_tot (kcal/mol) | |---|---:|---:|---:|---:| | Chordal | 180.29 ± 1.33 | 4.84 ± 0.21 | 0.27 ± 0.02 | 0.042 ± 0.006 (0.040) | | Geodesic-Taylor | 182.10 ± 0.16 | 4.72 ± 0.03 | 0.26 ± 0.01 | 0.041 ± 0.001 (0.030) | | Δ | −1.0% | +2.5% | +3.8% | +2.4% (+25.0%) |

Geodesic-Taylor (GrassDist in main text) loss achieves better performance across d_Gr, P, and E_tot (mean and median). With orbital-energy supervision included, the two losses achieve more comparable in-distribution performance, while geodesic loss retains a clear advantage in QMugs size extrapolation. Specifically, chordal loss yields 3.8%, 13.2%, and 43.8% higher energy MAE for molecules with 13, 15, and 17 heavy atoms, respectively. These findings support the value of the geodesic formulation, particularly for generalization beyond the training distribution.

  1. Gauge-redundancy ablation. We partially agree with the reviewer that occupied-subspace supervision is a major source of improvement. Indeed, demonstrating its importance is a central motivation and conceptual contribution of our work, as occupied-subspace supervision has not yet become a widely adopted principle in Hamiltonian prediction. Nevertheless, the results in Sec. 3.3 further demonstrate the benefits of our geometric formulation. Compared with coefficient-level supervision, GrassDist reduces total-energy MAE from 0.18 to 0.04 kcal/mol and dipole-moment error from 13.91 to 6.04, corresponding to reductions of 77.8% and 56.6%, respectively. Moreover, even with a small geodesic-loss weight and without orbital-energy supervision, GrassDist-lite improves total-energy MAE from 0.18 to 0.12 kcal/mol and dipole-moment error from 13.91 to 9.16. These results provide additional evidence for the benefit of our manifold-based supervision beyond naive coefficient-level objectives.

  2. Discussion of related work. We thank the reviewer for highlighting these relevant studies. We have expanded Sec. 4 and added the earlier quantum-chemistry Grassmann studies to Sec. 2.3. Their distinctions from GrassDist are summarized below: • Shao et al. (2023) directly learn the one-electron density matrix as a surrogate for electronic-structure calculations, whereas GrassDist predicts the Hamiltonian and explicitly supervises its induced occupied subspace. • Zhang et al. (ICML 2024) exploit Hamiltonian self-consistency for training without reference Hamiltonian labels, whereas GrassDist directly aligns predicted and reference occupied subspaces through a Grassmann manifold loss. • DEQH (NeurIPS 2024) incorporates self-consistency through an implicit fixed-point architecture, which is complementary to our loss-level approach. • Van Voorhis & Head-Gordon (2002) and Polack et al. (2021) employ Grassmann geometry for orbital optimization and accelerated MD via density-matrix extrapolation in conventional electronic-structure calculations, respectively. In contrast, GrassDist introduces Grassmannian supervision into equivariant Hamiltonian learning.

AuthorsOpenReview ↗

Continuing the response to W1 & Q1

  1. Comparison with additional loss variants. Following your suggestion, we have further extended the ablation to include chordal loss, Procrustes-aligned coefficient loss, and geodesic loss without the Taylor branch, under the main training protocol (Appendix C.9).
Subspace lossH (10^-6 Eh)d_Gr (10^-3)P (10^-3)E_tot (kcal/mol)eps_occ (10^-6 Eh)mu (10^-3 a.u.)S_acc (%)
Chordal180.29 ± 1.334.84 ± 0.210.27 ± 0.020.04 ± 0.00198.10 ± 9.366.29 ± 0.3961.69 ± 0.11
Procrustes180.84 ± 0.334.94 ± 0.210.28 ± 0.030.05 ± 0.00208.68 ± 20.766.61 ± 0.5361.60 ± 0.08
Geodesic180.85 ± 1.325.08 ± 0.370.28 ± 0.020.05 ± 0.02205.62 ± 12.146.58 ± 0.5961.73 ± 0.02
Geodesic-Taylor182.10 ± 0.164.72 ± 0.030.26 ± 0.000.04 ± 0.00191.93 ± 3.566.02 ± 0.0661.64 ± 0.12

The results show that these occupied-subspace objectives achieve comparable in-distribution performance when training converges, further supporting the effectiveness of subspace supervision. However, removing the Taylor branch introduces greater numerical variance.

  1. Frequency of the arccos branch. We additionally tracked approximately 8.44 million principal angles across two complete 200k-step training runs using different seeds. Only 0.0066% of angles enter the arccos branch overall, with the fraction decreasing from 1.21% during the first 1k steps to 0.00014% after 100k steps. This confirms that as training progresses, nearly all principal angles fall into the Taylor branch.
AuthorsOpenReview ↗

Response to W2-1 & Q2: Evaluation-set representativeness and outlier analysis.

We thank the reviewer for raising this concern. We provide additional evaluations to clarify the representativeness of our results and the influence of extreme outliers.

  1. Evaluation on the full test sets. Evaluating energy and SCF acceleration is computationally expensive, and prior studies typically report SCF acceleration on subsets of 50 configurations [2,3,7]. Our original evaluation of 800 configurations per dataset already provides a relatively large-scale benchmark. We also explicitly reported both mean and median errors to distinguish extreme failures from typical performance, and analyzed the two outliers in Appendix C.12 to illustrate the limitations of conventional Hamiltonian supervision. To directly address your concern, we have now extended the total-energy evaluation to the entire test set across all four QH9 splits, with the results summarized below.

Full-test EtotE_{\mathrm{tot}} MAE (kcal/mol):

DatasetBaseline E_tot mean (median)GrassDist E_tot mean (median)Fold (mean / median)
QH9-stable-iid2.8391 (0.2627)0.0385 (0.0331)73.7× (7.9×)
QH9-stable-ood0.6812 (0.2264)0.0392 (0.0367)17.4× (6.2×)
QH9-dynamic-geo0.8131 (0.2787)0.0256 (0.0248)31.8× (11.2×)
QH9-dynamic-mol68.2119 (1.2039)0.3805 (0.2296)179.3× (5.2×)

GrassDist maintains substantial improvements across all four full test sets, reducing mean and median energy errors by 73.7-fold and 7.9-fold on stable-iid, and by 179.3-fold and 5.2-fold on dynamic-mol, respectively. Although the two extreme outliers affect the original 800-configuration mean, these full-test results confirm that our main conclusions extend well beyond the original evaluation subsets.

  1. Results excluding the two outliers. We further recompute the mean errors on the original 800-frame stable-iid evaluation subset (800−2), after excluding the two molecules identified in Table 11: | Metric | Baseline | GrassDist | Fold (×) | |---|---:|---:|---:| | EtotE_{\mathrm{tot}} (kcal/mol) | 0.454 | 0.040 | 11.4× | | P (10−310^{-3}) | 1.747 | 0.263 | 6.7× | | dGrd_{\mathrm{Gr}} (10−310^{-3}) | 18.74 | 5.12 | 3.7× |

Even after removing these outliers, GrassDist reduces the errors by 11.4-fold, 6.7-fold, and 3.7-fold for EtotE_{\mathrm{tot}}, PP, and dGrd_{\mathrm{Gr}}, respectively. This demonstrates that the improvements are not driven solely by the two extreme cases.

  1. Error distributions across test splits. We have additionally included five tables in Appendix C.11 reporting the medians and 5th/95th percentiles across all metrics and QH9 splits. For brevity, we present the stable-iid results here.

Median [P5, P95] on QH9 stable-iid (units as in Table 1):

metricBaselineWALossGrassDist
H38.341 [23.855, 65.520]53.803 [34.461, 88.426]177.306 [113.908, 265.670]
dGr14.120 [8.352, 26.647]21.623 [13.020, 40.497]4.279 [2.739, 6.593]
P1.198 [0.514, 3.417]1.863 [0.774, 5.516]0.237 [0.129, 0.433]
Etot0.263 [0.092, 0.958]0.629 [0.223, 2.306]0.033 [0.015, 0.073]
ϵΔ\epsilon_\Delta830.746 [85.386, 3035.945]1208.708 [124.133, 4039.046]528.079 [44.871, 2815.422]
ϵocc\epsilon_{\rm occ}133.681 [68.779, 308.603]183.600 [93.129, 418.205]163.901 [83.675, 364.641]
μ\mu10.511 [3.290, 31.041]15.979 [5.047, 45.706]4.751 [1.498, 13.505]
Sacc (%)66.667 [58.333, 75.000]69.231 [63.531, 75.000]61.538 [54.545, 66.667]

GrassDist consistently achieves better median and percentile results for dGrd_{\mathrm{Gr}}, PP, EtotE_{\mathrm{tot}}, μ\mu, and SaccS_{\mathrm{acc}}, supporting improvements in typical-case performance beyond the extreme outliers. Similar trends are observed across the other QH9 splits (Appendix C.11).

4.Occupation-swap failure rates. Following your suggested criterion (any principal angle θi>π/4\theta_i>\pi/4), we further quantify occupation-swap failures across all QH9 splits:

modelstable-iidstable-ooddynamic-geodynamic-mol
Baseline0.16% (21/13084)0.03% (3/9335)0.03% (3/9990)4.12% (416/10100)
WALoss0.11% (14/13084)0.12% (11/9335)0.00% (0/9990)2.34% (236/10100)
GrassDist0.00% (0/13084)0.00% (0/9335)0.00% (0/9990)0.00% (0/10100)

Notably, GrassDist exhibits zero occupation-swap failures across all four splits, whereas Baseline and WALoss exhibit nonzero failure rates on several splits, particularly dynamic-mol. This further highlights the benefit of occupied-subspace supervision in preventing catastrophic prediction failures.

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

This work studies the Kohn–Sham Hamiltonian learning problem. GrassDist introduced in this paper provides Grassmannian geodesic loss on the Grassmann manifold that is invariant to orthogonal gauge transformations of the occupied orbitals. The gradient stabilization technique is also introduced to address the issue when the predicted and reference subspaces are close to each other. The experiments on several benchmark problems support the benefit of introducing the gauge invariant loss; The downstream tasks for predicting the properties improves and also generalizes to molecules larger than those seen during training.

Critical strengths

All of those supports the key contributions of the paper:

  • Strong empirical results that support the benefit of introducing gauge invariant loss
  • Practical technique to avoid the gradient instability when the predicted and reference subspaces are close to each other
  • The loss is simple to add to existing Hamiltonian prediction models
Critical weaknesses

The Grassmannian geodesic distance itself is well established. However, introducing it and showing the empirical evidence has the value.

Questions for authors

Since the gradient stabilization technique is one of the main methodological contributions, could you add an empirical comparison with / without the gradient stabilization technique?

3 comments
AuthorsOpenReview ↗

Reponse to Q1: Ablation of gradient stabilization techniques.

We thank the reviewer for this constructive suggestion. We have conducted a 2×2 ablation study of our two gradient stabilization techniques, StableEigh (SE) and the Taylor branch (TB), to evaluate their contributions to training stability. Appendix C.10 presents representative training curves under identical random seeds and hyperparameter settings across all four variants. We further quantify the percentage of training steps exhibiting loss >10> 10 and gradient spikes during the first five epochs (~16,000 steps):

VariantLoss >10 steps (%)Gradient norm > 1e6 (%)Gradient norm > 1e10 (%)
full3.60.40.0
Taylor-only8.97.77.1
StableEigh-only26.921.920.1
all-off17.514.14.0

The complete GrassDist model exhibits the lowest frequency of loss anomalies and extreme gradients. Removing either SE or TB substantially increases numerical instability, demonstrating their complementary roles in stabilizing the optimization process. We have included these ablation results and corresponding discussions in Appendix C.10 of the revised manuscript.

Thank you for the additional experiment. I keep the original rating (=3), but acknowledges that this strengthens the message of the paper.

AuthorsOpenReview ↗

Thanks for your careful review.

Thank you very much for taking the time to review our rebuttal and for your positive remarks on the additional experiments.

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

GrassDist supervises occupied orbital subspaces using a gauge-invariant Grassmannian geodesic loss, combined with Hamiltonian and orbital-energy losses. Numerical stabilization supports eigendecomposition backpropagation. Experiments on QH9, MD17, and QMugs, using two network backbones, report substantial improvements in density-matrix and total-energy prediction.

Critical strengths
  1. Physically meaningful supervision. The occupied subspace directly determines density-derived observables, addressing the mismatch between elementwise Hamiltonian error and physical-property accuracy.
  2. Substantial empirical gains. On QH9 stable-iid, mean energy error falls from 4.87 to 0.04 kcal/mol, and median error from 0.28 to 0.03. Results across datasets and backbones support broader usefulness.
Critical weaknesses
  1. The benefit of geodesic loss over simpler subspace losses is not yet clear. Chordal loss also provides gauge-invariant supervision of the occupied subspace. Appendix C.8 compares the two, but evaluates only 2,449 training configurations over the first ten epochs, using full-Hamiltonian prediction instead of the main residual parameterization. These results suggest an early optimization advantage, but do not establish better test performance after convergence.
  2. The contribution relative to closely related work needs clarification. Zhang et al. [1] use Grassmann geometry for density-matrix learning, while Lou et al. [2] incorporate density-matrix supervision into Hamiltonian learning. The paper should explain more precisely what its geodesic objective and training procedure add beyond these related approaches.
  3. Evidence for robustness and computational efficiency is limited. Only one training seed is reported, and energy/SCF metrics are evaluated on limited test subsets. There is also no isolated ablation of gradient truncation or measurement of runtime overhead. Consequently, consistency across runs, dependence on numerical stabilization, and practical computational savings remain unclear.
  4. Comparisons across baselines and neural-network architectures are limited. The main baseline comparisons use SPHNet, while the supplementary QHNet experiment compares GrassDist only against Hamiltonian-only training. It therefore remains unclear whether the advantages over alternative objectives consistently hold across architectures. Evaluating multiple strong baselines, including WALoss and chordal/projector supervision, on different backbones would strengthen the evidence for the method’s robustness and general applicability.

[1] Zhang, L., et al. A symmetry-preserving and transferable representation for learning the Kohn-Sham density matrix. arXiv:2503.08400, 2025. https://arxiv.org/abs/2503.08400 [2] Lou, Y., et al. A Fixed-Point Neural Operator for Size- and Functional-Transferable Hamiltonian Prediction. arXiv:2606.14498, 2026. https://arxiv.org/abs/2606.14498

Questions for authors
  1. Please compare geodesic loss with chordal loss, or an equivalent projector loss, on the held-out test set after convergence. Use the main MINAO-residual parameterization, matched orbital-energy supervision, comparable tuning budgets, and a common checkpoint-selection criterion.
  2. Please report variation across random seeds, sensitivity to the gradient-truncation threshold, and training failure rates. What is the additional training time relative to the baselines, and do fewer SCF iterations translate into lower total runtime?
  3. Please list the complete loss weights for each row of Table 2 and reconcile them with Appendix B.2. Which validation metric was used to select checkpoints?
6 comments
AuthorsOpenReview ↗

Response to W1 & Q1: Geodesic versus chordal loss.

We thank the reviewer for this constructive suggestion. We should have made this point clearer. Our central contribution lies in demonstrating the benefit of occupied-subspace supervision for downstream physical properties, with the Grassmannian geodesic loss providing a natural geometric formulation. Other gauge-invariant objectives, such as chordal loss, can also offer similar benefits, further validating our central claim. To address this concern, we have extended Appendix C.9 with a three-seed comparison of geodesic and chordal losses after convergence on held-out test sets including QH9 and QMugs. Both use the same MINAO-residual parameterization, matched orbital-energy supervision, comparable hyperparameter tuning budgets, and a common checkpoint-selection criterion.

Results on QH9-stable-iid

MethodH (10⁻⁶ Eh)d_Gr (10⁻³)P (10⁻³)E_tot (kcal/mol)ε_Δ (10⁻⁶ Eh)ε_occ (10⁻⁶ Eh)μ (10⁻³ a.u.)S_acc (%)
Chordal180.29 ± 1.334.84 ± 0.210.27 ± 0.020.042 ± 0.006 (0.040)902.71 ± 24.59198.10 ± 9.366.29 ± 0.3861.69 ± 0.11
Geodesic-Taylor182.10 ± 0.164.72 ± 0.030.26 ± 0.010.041 ± 0.001 (0.030)894.64 ± 19.19191.93 ± 3.566.02 ± 0.0661.64 ± 0.12
Δ−1.0%+2.5%+3.8%+2.4% (+25.0%)+0.9%+3.2%+4.5%+0.08%

The results show that the chordal loss achieves comparable in-distribution performance, with slightly lower Hamiltonian error but higher errors on the downstream ground-state observables. This indicates that gauge-invariant occupied-subspace supervision, rather than the particular metric on the Grassmannian, is the key source of improvement.

Energy MAE on QMugs

Method11 heavy atoms13 heavy atoms15 heavy atoms17 heavy atoms
Chordal0.290.832.497.25
Geodesic-Taylor0.290.802.205.04
Δ\Delta0+3.8%+13.2%+43.8%

Notably, geodesic loss retains an advantage in out-of-distribution size extrapolation on QMugs, where chordal loss yields 3.8%, 13.2%, and 43.8% higher energy MAEs for molecules with 13, 15, and 17 heavy atoms, respectively. These findings support the value of the geodesic formulation, particularly for size generalization, while acknowledging the effectiveness of other gauge-invariant alternatives. We have updated Appendix C.9 accordingly and moderated our discussion of chordal loss in the revised manuscript.

AuthorsOpenReview ↗

Response to W2

We thank the reviewer for pointing out these two relevant studies. Zhang et al. [1] directly predict the density matrix using symmetry-adapted ACE representations, trained with an unconstrained element-wise loss. The Grassmannian constraint is enforced through a post-hoc spectral retraction during inference, which fails to inform the model of the underlying Grassmannian geometry during training, as the model does not receive manifold-aware feedback during training. In contrast, GrassDist directly supervises the occupied subspace induced by the predicted Hamiltonian through a Grassmannian geodesic loss.

Lou et al. [2] incorporate an element-wise density-matrix MAE/RMSE loss without Löwdin orthonormalization, which differs from the chordal distance on the Grassmann manifold. To examine this distinction, we conduct a controlled comparison with the density-matrix loss (P-loss) using the same SPHNet backbone and training budget on QH9-stable-iid, with each method using its best validation setting (P-loss: w=0.05w=0.05; GrassDist: w=10w=10). GrassDist achieves better occupied-subspace and downstream predictions, including density-matrix error (0.24 vs. 0.41 ×10⁻³), geodesic distance (4.32 vs. 10.38 ×10⁻³), dipole moment (6.25 vs. 13.95 ×10⁻³ a.u.), total energy (0.03 vs. 0.12 kcal/mol), and SCF iteration ratio (61.7% vs. 66.5%; lower is better). We have expanded the discussion of these related works in the revised manuscript Sec. 4 and Sec. 2.3.

[1] https://arxiv.org/abs/2503.08400 Zhang, L., et al. A symmetry-preserving and transferable representation for learning the Kohn-Sham density matrix. arXiv:2503.08400, 2025.

[2] https://arxiv.org/abs/2606.14498 Lou, Y., et al. A Fixed-Point Neural Operator for Size- and Functional-Transferable Hamiltonian Prediction. arXiv:2606.14498, 2026.

AuthorsOpenReview ↗

response to W3 & Q2: Robustness and computational efficiency.

We thank the reviewer for these constructive suggestions. We address each concern below and have included the additional experiments and discussions in Appendix C.9, C.10 and C.15 of the revised manuscript.

  1. Variation across random seeds. We have extended our experiments on QH9 stable-iid to three independent training seeds:
MethodH (10⁻⁶ E_h)d_Gr (10⁻³)P (10⁻³)E_tot (kcal/mol)ε_Δ (10⁻⁶ E_h)ε_occ (10⁻⁶ E_h)μ (10⁻³ a.u.)S_acc (%)
Baseline40.82 ± 0.1869.54 ± 2.969.87 ± 0.763.16 ± 0.281743.69 ± 121.70491.31 ± 3.2523.13 ± 1.5467.19 ± 0.04
WALoss56.92 ± 0.0159.78 ± 2.179.24 ± 3.042.59 ± 0.162046.07 ± 35.93299.36 ± 0.6526.78 ± 0.4970.06 ± 0.06
GrassDist182.10 ± 0.164.72 ± 0.030.26 ± 0.010.039 ± 0.001894.64 ± 19.19191.93 ± 3.566.02 ± 0.0661.64 ± 0.12

Protocol: all metrics, including EtotE_{\mathrm{tot}}, use the full test split (13,084 frames); only SaccS_{\mathrm{acc}} is evaluated on the first 800 test configurations. The results remain consistent with our original findings, supporting the robustness of GrassDist. We note that single-seed evaluations have also been commonly adopted in prior Hamiltonian-learning studies [1, 2].

  1. Evaluation on larger test sets. Evaluating energy and SCF acceleration is computationally expensive, and prior studies typically evaluate SCF acceleration on subsets of 50 configurations [3,4,5]. Our evaluation on 800 test configurations per dataset already provides a relatively large-scale comparison. Nevertheless, we have now extended the energy evaluation to the entire QH9 stable-iid test set of 13,084 configurations. GrassDist maintains its advantage over the baseline, achieving nearly 9-fold and over 70-fold reductions in median and mean absolute energy errors, respectively. | Method | Mean | median | |---|---|---| | baseline | 2.84 | 0.26 | | GrassDist | 0.04 | 0.03 |

Since evaluating SCF acceleration requires full SCF calculations until convergence, we additionally examine the sensitivity of SaccS_{\mathrm{acc}} to evaluation subset size, using 200, 400, 800, and 1,200 configurations. Across these subsets, SaccS_{\mathrm{acc}} varies by at most 0.8 percentage points for each method, ranging from 67.17%–67.52% for Baseline, 69.98%–70.78% for WALoss, and 61.66%–62.29% for GrassDist. These results consistently support our original conclusions.

  1. Sensitivity to gradient truncation. Varying the StableEigh truncation threshold τ\tau from 1×10−41\times10^{-4} to 2×10−22\times10^{-2} yields similar training dynamics (updated Appendix Fig. 13). In contrast, removing StableEigh results in a 40% run-level instability rate (4 in 10 runs), defined as at least one step with a gradient norm exceeding 10610^6 within a run, demonstrating the importance of numerical stabilization.
  2. Additional training overhead. We report the per-iteration training time relative to Hamiltonian-only training on MD17 ethanol and QH9 stable-iid:
systemavg. #atomsbaselineGrassDistΔt
MD17 ethanol90.296 s0.320 s+24 ms (+8.1%)
QH9-stable-iid17.20.642 s0.828 s+186 ms (+29.0%)

The additional overhead mainly arises from eigendecomposition and subspace-loss computation, remaining modest for the molecular systems considered.

  1. Practical SCF acceleration. We additionally measure the total wall-clock time, including neural-network inference and SCF calculations, under the QH9 stable-iid setting of Table 1: | Method | S_acc | Mean wall time (s) | Relative wall time | |---|---|---|---| | MINAO | 100.00% | 7.45 | 100.0% | | Baseline | 67.29% | 5.27 | 70.8% | | WALoss | 70.29% | 5.38 | 72.3% | | GrassDist | 62.06% | 4.88 | 65.6% |

The wall-clock results are consistent with the SaccS_{\mathrm{acc}} trends. GrassDist reduces total runtime by 7.4% relative to Baseline and 9.3% relative to WALoss, confirming that fewer SCF iterations translate into practical computational savings. The wall-time ratios are slightly higher than the corresponding SCF iteration ratios due to fixed preparation costs (e.g., density-grid construction and electron-integral evaluation) that do not decrease with the number of SCF iterations.

[3] Zhang et al. "Self-consistency training for density-functional-theory hamiltonian prediction." arXiv preprint arXiv:2403.09560 (2024). [4] Wang, Zun, et al. "Infusing self-consistency into density functional theory hamiltonian prediction via deep equilibrium models." Advances in Neural Information Processing Systems 37 (2024): 89652-89681. [5] Yin, Shi, et al. "TraceGrad: a framework learning expressive SO (3)-equivariant non-linear representations for electronic-structure Hamiltonian prediction." arXiv preprint arXiv:2405.05722 (2024).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.