acceptodds
Under review as a conference paper at ICLR 2027

Stepwise Refinement and Label-Free Error Ranking for Kohn–Sham Hamiltonians

Abstract

Machine-learned Kohn–Sham Hamiltonians can bypass costly self-consistent density functional theory (DFT) calculations, but their practical use requires identifying predictions with large errors that should be recomputed with DFT. In this work, we introduce HIVE (Hamiltonian Iterative Velocity Estimator), an equivariant Hamiltonian prediction model that also provides an error ranking at no additional cost. During training, the model sees one randomly sampled point on a deterministic path from zero to the target Hamiltonian. At inference, we repeatedly apply the same field to its evolving Hamiltonian prediction, a process we call the rollout. With this strategy, averaging the stepwise estimates gives the final Hamiltonian, while their dispersion provides a label-free score for ranking likely failures without additional evaluations beyond the rollout. We test HIVE across four periodic benchmarks with PBE and HSE06 labels, and the results improve the best reported Hamiltonian mean absolute error on three and match it on the fourth. More steps improve accuracy without retraining, mainly because the finer rollout produces better estimates rather than because averaging cancels noise. On newly DFT-labelled structural shifts, the dispersion ranks per-structure error with Spearman correlations of 0.98–0.99 across all three PBE systems. Overall, HIVE provides both an accurate Hamiltonian prediction and a practical uncertainty signal without training a separate model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Hamiltonian prediction, equivariant neural networks, density functional theory, uncertainty quantification, error ranking, neural ODE, self-ensembling, materials science

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bgUeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work introduces HIVE, which repeatedly applies a single equivariant model to refine Hamiltonian predictions and rank per-structure prediction errors within the same rollout. It uses a flow-based ODE formulation similar to QHFlow, but constructs a deterministic path starting from a zero source so that every step estimates the same target Hamiltonian. The stepwise predictions are averaged to obtain the final Hamiltonian, while their dispersion provides an error-ranking score without requiring labels at inference or a separate uncertainty model. The authors report lower Hamiltonian MAE than previous results on three of four periodic benchmarks, while matching the prior result at the reported precision on the fourth. They also evaluate accuracy across different step counts using the same checkpoint, assess error ranking under structural shifts using newly computed DFT labels for three PBE systems, and separate the gains from averaging from improvements in the individual step estimates.

Critical strengths

The joint construction of prediction and error ranking within a single rollout is interesting. Since every step estimates the same Hamiltonian target, the roles of the mean and dispersion are clear. The ranking score requires neither an additional ensemble nor extra field evaluations beyond the rollout already used for prediction. I also appreciate that the authors clearly explain the inability to detect errors shared across steps and interpret the score as a ranking signal rather than a calibrated error estimate. In Hamiltonian learning, where additional data can be obtained through DFT calculations, deciding which predictions should be recomputed is an important problem. The paper explains this need intuitively and helps draw attention to it.

The experiments and analyses provide thorough support for the central claims. By varying only the step count at a fixed checkpoint, the authors show that accuracy improves over the first few steps on all four benchmarks and distinguish the benefits of averaging from improvements in the individual step estimates. For structural shifts, they compute new DFT labels and include comparisons against displacement RMSD, ranking within individual displacement levels, and risk–coverage evaluation on the original test sets. These controls help assess whether the dispersion captures information about structure-specific prediction errors.

Critical weaknesses

Evidence for the practical utility and robustness of error ranking remains limited. The high pooled correlations in Figure 3 span a wide error range across different displacement levels. Although within-level and original test-set evaluations are also provided, the in-distribution correlation for bilayer graphene is 0.442, lower than the pooled result, and the main analyses focus on one checkpoint per system. In addition, the active-learning pilot in Appendix C.4 shows no consistent advantage over random selection when acquisition is restricted to new candidates. The demonstrated acquisition benefit is therefore mainly avoiding redundant selection of structures already seen during training. I acknowledge that the authors explicitly state these limitations, but it remains difficult to judge how effectively error ranking translates into more efficient data acquisition.

The practical advantage over existing methods is difficult to assess after accounting for model size and computational cost. Increasing K at a fixed checkpoint provides evidence that additional inference computation improves accuracy, but the baseline results in Table 1 are taken from the literature and do not match model size or training and inference budgets. HIVE uses larger models than some existing methods, so its resource requirements should be considered alongside the accuracy gains. It would be helpful to include parameter counts and inference times measured under consistent conditions in Table 1, and to report training and inference costs separately. This would clarify whether HIVE offers a better combination of accuracy and error ranking than existing predictors or ensembles under the same compute budget.

Questions for authors

Performance relative to computational cost. Could you report parameter counts and inference times measured under consistent conditions for the models in Table 1? With training and inference costs reported separately, and accounting for model size and geometry feature caching, how does HIVE compare with existing predictors or ensembles in Hamiltonian MAE and risk–coverage performance under the same inference budget?

Contributions of the time and Hamiltonian inputs. Could you compare the prediction accuracy of geometry-only, geometry+time, and geometry+time+H models using the same backbone and training conditions?

Reproducibility of error ranking. Does the in-distribution and within-level ranking performance of DMD_M remain similar across the already trained ensemble members?

Reviewer_1xzhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Hamiltonian Iterative Velocity Estimator (HIVE), a novel model for Kohn-Sham Hamiltonian prediction. HIVE augments Hamiltonian prediction with an error ranking by an equivariant prediction model. Benchmark experiments show promising performance of HIVE.

Critical strengths
  1. This paper makes a good novelty contribution in proposing HIVE for Hamiltonian prediction. The idea of predict error ranking and applying rollout in inference is useful and insightful to researchers in AI for quantum chemistry field.
  2. The writing of this paper is generally good, clear, method details and formulation are well elaborated.
Critical weaknesses
  1. Lack discussion about practical usage gap to replace pure DFT methods, and performance/computational cost tradeoff, particularly the large materials.
  2. The major contribution of this paper is the idea of “providing an error ranking”, or “the new model that provides an error ranking”? If it is the prior one, authors are recommended to do experiments and show that the idea works for multiple different networks (showing using error ranking achieves better performance than without error ranking on the same network architecture).

--------------------Post Rebuttal---------------------------
My concerns have been addressed in rebuttal so I increase my rate.

Questions for authors

No other questions, see "Critical Weaknesses" section.

Reviewer_bPQMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HIVE trains an equivariant field along a deterministic path from zero to the target Hamiltonian. The paper studies accuracy, displacement shifts and a small active-learning experiment.

Critical strengths

I find the idea in this paper is interesting as it establish a pathway that the Hamiltonian of materials are learned in a new way. This way, if well established and studied, can be quite useful for future developments.

Critical weaknesses

Although this paper is very interesting in terms of methodology development, I find this paper did not fully demonstrate its capability in actual applications. See below

Questions for authors
  1. I find the method is very new in this paper, but I don't think this paper properly tested or benchmarked the method enough.
  2. The accuracy and cost comparison is incomplete. Table 1 compares an eight-evaluation rollout with results taken from other papers, while Table 4 lists models with 64.5–143.6M parameters. The K sweep shows that more computation helps this trained model, but it does not show whether the same budget is better spent on a direct predictor or a smaller ensemble. The ensemble comparison also has unequal training budgets. Please report an accuracy–latency comparison and at least one direct-prediction baseline trained with a comparable backbone and budget. The phrase “no additional cost” should remain specific to computing dispersion after the rollout.
  3. At comparable measured inference time, how does HIVE compare with a direct predictor and a small ensemble?
  4. is it possible to establish some cross system benchmark using this approach, instead of training and testing on the same system?

Comments on the paper

1 comment
Reviewer_1xzhOpenReview ↗

Follow-up Response

I appreciate authors' efforts in rebuttal. My concerns have been addressed so I increase my rate. Please include all new results in the revised paper .

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.