acceptodds
Withdrawn from ICLR 2027

Learn the Field, Construct the Matrix: Rethinking Targets for Kohn–Sham DFT

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Machine learning is increasingly coupled with Kohn–Sham density functional theory (DFT), yet it remains unclear which electronic representation provides the most effective interface to the solver. We compare four candidates—charge density , effective potential , Hamiltonian , and density matrix —on a unified 43k-crystal benchmark, evaluating native prediction accuracy, closed-loop self-consistent-field (SCF) acceleration, and downstream property recovery. Although and lie closest to the eigensolution, their learned counterparts realize little of the large acceleration available to exact matrix states. In contrast, learned and accelerate SCF on 90.3% and 84.3% of test crystals, with median loop speedups of and . More strikingly, matrices reconstructed from learned real-space fields outperform dedicated matrix-input routes on 97.6–99.9% of test crystals after one solver step, while residual further improves downstream recovery under matched electronic work. These results show that proximity to the eigensolution is a poor criterion for choosing a learning target. Instead, effective ML–DFT interfaces should expose quantities from which the solver can analytically reconstruct constrained electronic states, rather than requiring neural networks to learn those constraints directly.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
density functional theory, effective potential, electronic structure, equivariant neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_5jgDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the choice of learning targets for ML period DFT initialization. It empirically demonstrates that VeffV_\textrm{eff} is advantageous over other learning targets.

Critical strengths

This paper shows that VeffV_\textrm{eff} can be a better learning target for the dataset and ML models the author tested.

Critical weaknesses
  1. This paper doesn't demonstrate enough contribution to the field or meet the conference's standards. It cites "Chen et al., 2026" (https://arxiv.org/abs/2609.15151) multiple times and acknowledges that this paper is a follow-up work to "Chen et al., 2026". However, the overlap between "Chen et al., 2026" and this work is very large: their dataset is the same, and the SCF benchmark methods are the same. This paper also reuses HH, DD, and ρ\rho evaluation from "Chen et al., 2026", without making new progress in these three targets. For the ML model architecture part, this paper uses an existing framework, ChargE3Net. This work added VeffV_\textrm{eff} to the existing dataset, but calculating VeffV_\textrm{eff} is easy given that HH and DD already existed in "Chen et al., 2026". Another new thing (compared to "Chen et al., 2026") in this work is the downstream properties evaluation. However, downstream properties evaluation is not a new task and is widely benchmarked.

  2. The experiments in this paper are not able to prove that VeffV_\textrm{eff} is broadly more advantageous than other prediction targets. Thus, the conclusion in the paper has a narrow scope. Specifically:

  • The most important conclusion in this paper, that VeffV_\textrm{eff} can be advantageous for SCF initialization, is known in the field. And the related work includes (https://arxiv.org/abs/2509.25724, https://arxiv.org/abs/2604.21657)
  • Even in the period system this paper focused on, it fails to establish that VeffV_\textrm{eff} is generally/broadly more advantageous. Only one dataset has been evaluated, and changing the dataset can potentially reverse the conclusion, given that the current gap between VeffV_\textrm{eff} and the competing method is very close. To show a general advantage, the author needs justifications from physical intuition, evidence across datasets and calculation settings. One thing the author can do is to show a metrics v.s. dataset size curve/table. This can help us understand if their ranking flips with different data volume.
  • Taking one step back, this paper also doesn't provide strong enough evidence for the provided dataset. The biggest problem is that the DD baseline is too weak. As shown in Table 3, the SAD default takes 16 cycles to converge. However, as a residual learning approach based on SAD, HamGNN-Matrix-∆D still needs 16 cycles to converge. Without the residual, ML-based initialization even needs more SCF cycles to converge. So either the dataset chosen is very hard for this task, or the baseline method is not well optimized/suitable for this dataset. Showing VeffV_\textrm{eff} beating a poor baseline is not unexpected. Please clearify if the advantage of VeffV_\textrm{eff} comes from a better learning target, or better/suitable model architecture.
Questions for authors

See weakness

Reviewer_4ENwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper estimates several targets from kohn-sham dft, hypothesizes that the targets of charge density and hamiltonian are perhaps historical artifacts rather than optimal targets. They investigate targeting hamiltonian, charge density, VeffV_{eff}, and density matrix. They introduces a relatively large benchmark, finds speedups in many cases that use charge density or VeffV_{eff} albeit with modest median speed increases, and further show that real space fields are faster in almost all cases than matrix based versions.

Critical strengths
  • The dataset seems to be a nice contribution. There are many papers that create fairly limited datasets, but this one is larger.
  • The hypothesis is clearly stated and explored, the thinking is well organized and their results are modest but clearly a gain. It also sets up a whole other direction of study for this growing field.
  • The results against methods they compared with are strong and support that targets outside of the hamiltonian and density matrix could improve results and also on derived properties.
  • The metric for both fewer than sad and the time percentage are nice to show that the method actually provides acceleration.
Critical weaknesses
  • After a clear introduction, the results section quickly descends into quite some jargon and LLM-style writing (in my opinion). The pipeline itself should also be presented better, perhaps with a figure. There is quite some detail that is fairly confusing to me even though I'm rather familiar with hamiltonian prediction.
  • Accelerations are fairly modest. It seems like a lot of complexity for ~5-20% speedups. Using an LLM to optimize the solver in pySCF seems to plausible to offer a better speedup.
Questions for authors
  • The choice of grid seems arbitrary, but likely important. After all these quantities can change suddenly on the grid. Did you try any tests about grid size generalization? What would it take?
  • Are you advocating for your neural network too? That part wasn't so clear to me. It seems to perform well, but you don't focus on the network itself making a big improvement. Maybe that can be addressed? It should also live within the results tables where all methods that predict rho and V are presented at "equal weight."
  • Although you are working in the solid state chemistry regime there is a method that works across the periodic table with extremely strong results. Perhaps it makes sense to cite localized basis set Hamiltonian prediction methods such as HELM from Manasa Kaniselvan, et. al. 2025 and related? After all in table 1 you are attempting to summarize contributions from the past years and I believe HELM is by far the most performant and trained on the largest datasets.
Reviewer_vKg3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes learning the effective potential VeffV_\mathrm{eff} as a better electronic representation for machine-learning-assisted Kohn–Sham DFT. The authors compare four learning targets—charge density ρ\rho, effective potential VeffV_\mathrm{eff}, Hamiltonian H, and density matrix D—using representative backbone architectures for each representation on a unified benchmark of approximately 43,000 crystals, evaluating SCF acceleration and direct downstream property prediction. The experimental results support the advantages of learning VeffV_\mathrm{eff} in both SCF acceleration and direct downstream property prediction.

Critical strengths

-The paper constructs a benchmark dataset for effective-potential learning, providing a useful resource for future research.

-The paper provides detailed implementation settings and evaluation protocols for its effective-potential learning pipeline, facilitating the reproduction of its results.

Critical weaknesses

-The experimental design does not adequately support the paper's central objective of comparing different electronic representations. The models use different architectures and training strategies, making it difficult to attribute the performance differences to the learning targets themselves. Notably, (ρ\rho) and (VeffV_{\mathrm{eff}}), both predicted using ChargE3Net, achieve nearly identical SCF acceleration, whereas the matrix models use different backbones and perform substantially worse. The authors also acknowledge that their results rank specific pipelines rather than representations. Therefore, the current experiments do not provide convincing evidence for the paper's broader conclusions about learning-target selection.

-The practical computational benefits of the proposed approach are questionable. Predicting (VeffV_{\mathrm{eff}}) requires an additional projection onto the atomic-orbital basis to construct the Hamiltonian, introducing approximately 15.3 seconds of overhead compared with direct Hamiltonian prediction in downstream property calculations. Moreover, the reported SCF-loop speedup is only 1.17×, which decreases to 1.11× when including neural-network inference. The aggregate timing ratio is merely 1.047×, even before accounting for all workflow overheads. Such limited acceleration raises concerns about the practical value of the proposed approach.

-The paper is poorly organized and difficult to follow. The narrative is frequently interrupted by excessive technical details, disrupting the logical flow and obscuring the central scientific message. The overall presentation lacks a clear and coherent storyline, making the paper unnecessarily difficult to follow.

Questions for authors

-Why do the reproduced Hamiltonian and density-matrix models perform so poorly? The reported results for DeepH-E3, NextHAM, and HamGNN-Matrix appear substantially weaker than expected from their original publications, with some models even performing worse than the SAD-derived initial guesses. Can the authors explain these performance gaps and clarify whether the models were faithfully reproduced and sufficiently optimized for the current benchmark?

-How are the physical constraints of the predicted density matrices enforced? Appendix Table 9 lists electron-number conservation, idempotency, and occupation bounds as constraints requiring explicit enforcement or learning, while Section 4.4 reports substantial occupation-bound violations. Can the authors clarify how these constraints are incorporated into the evaluated models and whether they are effectively satisfied by the predictions?

Reviewer_UD48OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors addresses the problem of representation choice to solve KS DFT prediction through ML-based approach. They propose learning the effective potential and reconstructing electronic matrices through the DFT solver rather than predicting them directly. Using a benchmark of approximately 43k inorganic crystals, the results report prediction accuracy, SCF acceleration, and downstream property recovery. The evaluations show that learned real-space fields generally outperform the evaluated matrix-learning approaches, although all experiments are conducted at the PBE level of theory.

Critical strengths
  1. Interesting and potentially impactful research direction. Rather than focusing solely on improving prediction accuracy, the paper investigates which electronic representation provides a more effective interface between ML and DFT solvers. This is an important question for developing practical ML-accelerated electronic-structure calculations.

  2. Systematic evaluation across multiple representations. The unified comparison of four electronic representations, including closed-loop SCF acceleration and downstream property recovery, provides useful evidence beyond conventional prediction-error benchmarks.

Critical weaknesses
  1. Limited demonstration of the broader applicability and practical impact. The central conclusion—that learning real-space fields and reconstructing matrices is preferable to directly learning electronic matrices—is supported only at the PBE level of theory. However, a major motivation for ML-assisted DFT is accelerating computationally expensive electronic-structure calculations, including higher-level functionals such as r²SCAN, HSE06, and ωB97M-V. For these functionals, additional orbital-dependent or nonlocal contributions can change the Hamiltonian construction and the relative advantages of different learning targets. Therefore, the current results provide limited evidence that the proposed methodological insight extends to more accurate electronic-structure methods.

    Alternatively, the authors could establish practical significance within the PBE setting by demonstrating advantages for substantially larger systems, where computational savings become more consequential. The current benchmark is dominated by relatively small unit cells, leaving this potential application insufficiently explored.

  2. Limited demonstrated computational benefit. Although the proposed potential-based initialization accelerates SCF iterations, the reported speedup decreases to 1.111× when ML inference is included, and the aggregate runtime improvement is only 1.047×, excluding some I/O overhead. This makes it difficult to assess the practical impact of the method in its current evaluation setting.

Questions for authors
  1. Can the authors provide preliminary evidence supporting either of the following directions? Both would strengthen the paper's contribution, and demonstrating one would be sufficient to address my main concern.
    • Extension toward higher-level functionals: Demonstrate, even on a limited scale, whether the proposed field-learning and solver-reconstruction principle remains useful for a more accurate functional, potentially with modifications to accommodate its additional operator terms. Relevant examples include the ωB97M-V-based OMol25 dataset [1] and the r²SCAN-based MatPES dataset [2]. The existing DeepH-hybrid study [3] also provides relevant context for learning electronic representations beyond semilocal DFT.
    • Practical scalability within PBE: Alternatively, demonstrate that the proposed approach provides meaningful computational benefits for substantially larger systems or supercells, and strengthen the paper's motivation and experimental analysis around large-scale semilocal DFT applications. Evidence of favorable end-to-end computational savings as system size increases would help establish practical impact without requiring higher-level functionals.
  2. How do the authors position their central contribution in light of these two possibilities? Is the intended claim a general principle for selecting ML representations across electronic-structure methods, or a practical acceleration strategy for large-scale semilocal DFT? Clarifying this scope and providing supporting evidence in either direction would substantially strengthen the paper.

1] Levine, D. S., et al. "The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models." arXiv:2505.08762, 2025. https://arxiv.org/abs/2505.08762

[2] Kaplan, A. D., et al. "A Foundational Potential Energy Surface Dataset for Materials." arXiv:2503.04070, 2025. https://arxiv.org/abs/2503.04070

[3] Tang, Z., et al. "A Deep Equivariant Neural Network Approach for Efficient Hybrid Density Functional Calculations." Nature Communications, 15, 8815, 2024. https://doi.org/10.1038/s41467-024-53028-4

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

Related papers

Discussion (0)

Sign in to comment.