acceptodds
Under review as a conference paper at ICLR 2027

Machine Learning Force Fields Without Precomputed Labels: Variational Kohn-Sham Training

Abstract

Molecular dynamics at the accuracy of density functional theory (DFT) is limited by the cost of a self-consistent field (SCF) calculation at every step. Machine learning force fields (MLFFs) remove that cost from the simulation by regressing energies and forces onto reference DFT calculations, but move it to the dataset: every training label is itself a converged SCF calculation, and a new system, level of theory or range of conditions requires a new set of them. We merge labelling and training into a single optimisation. Because the Kohn-Sham energy is a variational functional of the electronic density, a network predicting the electronic state of a molecule can be trained by minimising the Kohn-Sham energy of its own prediction, averaged over geometries taken from an inexpensive empirical force field trajectory; the physics supplies the training signal and no labelled dataset is precomputed. The network emits a matrix in the atomic-orbital basis whose eigendecomposition yields an admissible density, energies follow from the functional and forces from the standard analytic nuclear gradient at that density, so inference requires no SCF iteration. Across ten drug-like and small organic molecules the model reaches a mean force error of meV/Å against converged DFT, and on a temperature-extrapolation benchmark it reaches meV/Å force RMSE far outside the training distribution, against for the strongest supervised baseline. Against supervised training of the same network on the same geometries it is as accurate or better while using –% less GPU time, and it drives stable molecular dynamics at a third of the cost per step of DFT.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning force fields, variational training, Kohn-Sham density functional theory

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_PwhdOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a variational Kohn–Sham training framework for machine-learning force fields without precomputed DFT energy and force labels. Experiments on molecular systems show competitive force accuracy and promising generalization.

Critical strengths
  • Eliminating the need for precomputed DFT labels could reduce the data-generation burden for machine-learning force fields.
  • The density-matrix parameterization, eigendecomposition gradient treatment, and analytical force construction address implementation challenges, supported by a relatively detailed mathematical analysis.
Critical weaknesses
  • The paper lacks comparisons with several relevant lines of prior work, including variational neural-network DFT-based training [1, 2], label-free electronic-structure learning [3], and methods incorporating SCF procedures [4]. The authors should clarify what is fundamentally novel beyond adapting these existing ideas to force-field prediction.
  • The method still requires electronic-structure calculations during training and inference. A comprehensive end-to-end comparison with supervised training is lacking.
  • Performance depends strongly on the sampling temperature and configuration coverage. More controlled experiments are needed to disentangle the benefits of variational training from those attributable to the training-data distribution. Furthermore, for larger molecules, the number of configurations required for convergence and the resulting model performance remain unclear.

[1] Li Y, Tang Z, Chen Z, et al. Neural-network density functional theory based on variational energy minimization[J]. Physical Review Letters, 2024, 133(7): 076401.

[2] Mathiasen A, Helal H, Balanca P, et al. Reducing the cost of quantum chemical data by backpropagating through density functional theory[J]. arXiv preprint arXiv:2402.04030, 2024.

[3] Zhang H, Liu C, Wang Z, et al. Self-consistency training for density-functional-theory hamiltonian prediction[J]. arXiv preprint arXiv:2403.09560, 2024.

[4] Wang Z, Liu C, Zou N, et al. Infusing self-consistency into density functional theory hamiltonian prediction via deep equilibrium models[J]. Advances in Neural Information Processing Systems, 2024, 37: 89652-89681.

Questions for authors

Related concerns are discussed in the weakness section.

3 comments

Response to Reviewer Pwhd

W1 — Prior work [1]–[4]; what is fundamentally novel

The revised paper cites [1]–[4], D4FT (Li et al., 2023), Hassan et al. (2025) and Dong et al. (2025), and credits them in Sections 1 and 2 (in blue) for the objective, the eigendecompose-and-occupy density construction, the minimal-basis anchor and the SCF-initial-guess use of Appendix I.

What is novel is what turns the objective into a force field, which none of [1]–[4] attempts. (i) Forces: none of the four computes one ([2] states they "can be derived from the Hamiltonian" and does not test it; [3] leaves the MD22 force labels unused). The loss reaches the predicted matrix only through its occupied subspace, so its eigenvalues are never trained (Section 4.2); forces read from them are 518× worse than ours and the uncanonicalised Rayleigh-quotient readout 132× worse (Table A7). The canonicalised readout of Eq. (5), exact to 10⁻⁹ meV/Å at a converged density, is ours. (ii) Single-pass inference: [2] follows the network with three DFT iterations and [4] iterates it to a fixed point; ours is one forward pass, one Fock build, one gradient, and Appendix H shows why even one Roothaan step overshoots. (iii) A backward pass with only occupied–virtual denominators (Eq. 6), where [3] clips eigenvalue differences and skips steps and [1], [2] describe no treatment; a primitive, not a new idea, and the revised Section 4.2 and Appendix G cite PySCFAD (Zhang & Chan, 2022) and Kasim (2020). To our knowledge no prior model trained this way has been evaluated as a force field.

W2 — Electronic-structure calculations remain at training and inference; end-to-end comparison

Table 2 is an end-to-end comparison: its supervised side pays labelling plus training and its variational side training only, with the same network, geometries, optimiser, schedule, seed and GPU type, scored on identical frames against our own converged SCF. On all four molecules the variational model has the lower force and energy error at the lower total cost; the aspirin and azobenzene gaps lie far outside the seed-to-seed spread of the new Appendix N, the ethanol and malonaldehyde force gaps inside it. Appendix F repeats the comparison on 3BPA at ωB97X/6-31G(d), where the two routes cost the same and the variational model's force error is about half the supervised one on every test set.

W3 — Dependence on sampling temperature and coverage; configurations needed for larger molecules

To separate sampling from objective we gave MACE our training distribution outright: maintainers' rMD17 settings, our 94,950 ethanol geometries sampled at 1600 K with a converged label for each, 20 epochs (1.9 million frame visits, twice our 950,000), scored on the same complete 1,000-frame test set against our converged SCF. MACE is one training run; our entries are the mean ± SD over three training seeds:

ethanol modelforce MAE (meV/Å)energy MAE (meV)
Supervised MACE (94,950 labels), same geometries1.420.57
Variationally trained QHFlow2 (ours, 3 seeds), same geometries, no labels0.64 ± 0.140.010 ± 0.002

Given our geometries and a label for each, MACE does not approach our accuracy (Table A15, new Appendix N): the training distribution alone does not produce the accuracy, and the objective alone (Table A6) does not produce the robustness under extrapolation.

Larger molecules: Appendix E trains one model across the eight galectin-8 ligands (52 to 64 atoms) with the budget of a single rMD17 molecule, 95,000 geometries and 237,500 steps in total, i.e. 11,875 geometries per ligand. It reaches 0.90 to 0.96 meV/Å force MAE on every ligand (Table A5), against a mean of 0.98 meV/Å for the per-molecule rMD17 models: for drug-like molecules of this size the configurations consumed and the accuracy reached are comparable to the small-molecule case, with an eighth of the geometries per molecule.

I appreciate the authors' substantial efforts in revising the manuscript and conducting additional experiments. Most of my original concerns have been adequately addressed, and the revised manuscript is significantly improved. I am therefore willing to increase my rating.

We sincerely thank you for the thoughtful feedback and positive assessment of our revised manuscript.

Reviewer_7hVFOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces variational Kohn–Sham training, where a neural network predicts the electronic density through an intermediate matrix without precomputed DFT labels. Since Kohn–Sham DFT minimizes an energy functional of the electronic density, the network can be trained by minimizing this functional without requiring precomputed labels. The results show that the proposed method achieves higher accuracy than the evaluated MLIPs and is faster than conventional DFT.

Critical strengths

The idea is technically solid and novel, and the method is clearly structured.

The paper targets the gap between DFT and MLIPs. It is faster than DFT and slower than MLIPs, but achieves higher accuracy than the evaluated MLIPs.

Critical weaknesses
  1. Detailed performance analysis is missing.

While the total training cost is compared with supervised training of the same backbone, including DFT labeling, inference speed is only compared with DFT. For a more complete comparison, inference speed should also be compared with MLIPs.

  1. Limited comparison with recent MLIP architectures.

The main comparison with MLIPs is limited to NequIP, Allegro, BOTNet, and MACE, while the proposed model uses an SO(2)-based backbone.

There are recent SO(2) convolution models with molecular dataset benchmarks. For example, DPA4 (https://arxiv.org/abs/2606.02419) and eSEN (https://arxiv.org/abs/2502.12147) report accuracy results on the SPICE-MACE-OFF dataset (https://arxiv.org/abs/2312.15211). Also, UMA (https://arxiv.org/abs/2506.23971) reports accuracy results on the OMol25 dataset.

Although the authors provide an accuracy comparison with supervised training of the same backbone, the proposed method could also be compared with recent models achieving state-of-the-art accuracy.

  1. Limited evaluation of electronic properties.

Since the proposed model predicts an electronic density through molecular orbitals, it could potentially be used more broadly than just predicting energies and forces. However, quantitative evaluation of other electronic properties is missing.

Questions for authors
  1. Could the authors provide an inference speed comparison with MLIPs?

  2. Could the authors provide an accuracy comparison with eSEN and DPA4 on the SPICE-MACE-OFF dataset under matched DFT settings? This would not only compare the proposed method with recent state-of-the-art architectures but also evaluate its generalization across different molecules and chemical elements, which has already been extensively studied for universal MLIPs.

  3. Could the authors provide errors in the predicted density matrices or occupied orbital subspaces compared with converged DFT? Alternatively, could the authors provide accuracy results for molecular properties other than energies and forces that can be obtained from the predicted electronic density?

4 comments

Response to Reviewer 7hVF (part 1 of 2)

W1 / Q1 — Inference speed against MLIPs

All methods on one NVIDIA L40S (the GPU of Tables 2 and A4), one protocol (batch size 1, 10 warm-up and 50 timed passes, forces by automatic differentiation): MACE-OMOL-0 in its extra-large configuration (52.4M parameters, 6 Å cutoff, as shipped), our model, and DFT warm-started from the previous frame's converged density, as a DFT-MD code runs it. Median time per frame:

ethanol (9 atoms)malonaldehyde (9)aspirin (21)azobenzene (24)Ac-Ala3-NHMe (42)
MACE-OMOL-0 extra-large (ms)29.529.540.140.479.8
Variationally trained QHFlow2 (ours) (ms)981083073211128
warm-started DFT (ms)444518140015264013

Per geometry we are slower than MACE by 3.3× (ethanol) to 14× (Ac-Ala3-NHMe) and faster than warm-started DFT by 3.6× (Ac-Ala3-NHMe) to 4.8× (malonaldehyde, azobenzene). The method competes with supervised MLFFs on accuracy and on independence from precomputed labels, not on throughput per geometry. The gap to MACE is structural (Section 6): inference is a network forward pass followed by a Fock build and an analytic nuclear gradient, both full-basis operations, so removing the SCF loop leaves a small constant factor over DFT rather than the orders of magnitude a fitted potential delivers.

The "1.9–3.4×" of Sections 1 and 6 refers to the Ac-Ala3-NHMe MD step of Table A4: 3.72 s/step for warm-started DFT against 1.09 s/step for our model with the efficient inference path of Section 4.4 (3.4×) and 1.98 s/step with the standard path (1.9×); the revised Sections 1 and 6 say so. The table above is the new Appendix M (Table A14).

W2 / Q2 — eSEN / DPA4 on SPICE-MACE-OFF; comparison with pretrained models

A comparison on a pretrained model's own dataset would measure the method where the labels are already paid for. Variational training re-performs the labelling work (one Fock build per frame visit, as per SCF iteration), so a model of SPICE or OMol25 trained without their labels would spend compute of the same order; it has no advantage there and we claim none. SPICE's level of theory is in addition outside our training layer (no meta-GGA kinetic-energy-density term, no D3 term). The paper addresses the other case: a pretrained model whose labels are at a different level of theory from the one the study is defined at, or whose training distribution does not cover the target molecule or region of configuration space; there the user has to label the target molecules at the target level and train or fine-tune a model, and that labelling is what our method removes.

On SO(2)-based models: our backbone is itself an SO(2)-equivariant eSCN-style network (Kim et al., 2026). Tables 2 and A6 compare that same network trained by the supervised route against variational training on the same geometries at matched DFT settings: variational training is at least as accurate on each of the four rMD17 molecules of Table 2 and about twice as accurate on every 3BPA test set.

Response to Reviewer 7hVF (part 2 of 2)

We ran the pipeline a user would run with a pretrained model for the eight galectin-8 ligands of Appendix E (52–64 atoms, ωB97X/6-31G(d)): the 11,875 geometries per ligand our model consumed were labelled at the paper's reference settings, MACE-OMOL-0 (extra-large, pretrained on OMol25) was fine-tuned on all 95,000 labels with its maintainers' protocol, and both were scored on the held-out frames and references of Table A5 (revised Appendix E). Force MAE (meV/Å), energy MAE and energy MAE after removing the per-ligand offset (meV), 100 frames per ligand:

ligFligOMeligNHMeligNMe2ligOEtligPyrligOHligNH2
MACE-OMOL-0 fine-tuned, force2.142.032.052.102.032.092.042.03
Ours, force0.940.910.950.940.930.960.900.93
MACE-OMOL-0 fine-tuned, energy2.119.514.6020.163.689.939.908.84
Ours, energy0.310.320.340.370.360.410.310.32
MACE-OMOL-0 fine-tuned, centred energy1.871.531.651.741.842.071.741.84
Ours, centred energy0.060.060.060.080.070.070.060.07

Our model is the more accurate on every ligand, on forces and on energies, with none of the per-ligand offsets that dominate the fine-tuned model's energy error. Its accuracy is set by training time, which can be extended at any time, not by a label count fixed before any model is seen.

W3 / Q3 — Electronic properties

100 test frames per molecule for the four molecules of Table 2, where a variationally trained ("var.") and a supervised ("sup.") model of the same network exist. ‖ΔP‖/‖P‖ is the relative Frobenius error of the density matrix; sin θ_max the sine of the largest principal angle between the model's and the SCF's occupied subspaces (the gauge-invariant form of the orbital cosine similarity that Zhang et al., 2024, report); Δμ the dipole error and Δq the Mulliken charge MAE, both from P_θ directly:

molecule‖ΔP‖/‖P‖ (var.)‖ΔP‖/‖P‖ (sup.)sin θ_max (var.)sin θ_max (sup.)Δμ, D (var.)Δμ, D (sup.)Δq, e (var.)Δq, e (sup.)
ethanol7.93e-41.79e-32.00e-43.25e-40.00200.00210.000200.00053
malonaldehyde1.44e-32.89e-34.45e-45.00e-40.00640.00610.000400.00123
aspirin4.50e-33.17e-27.03e-42.00e-30.01440.01390.000760.00629
azobenzene3.18e-31.92e-27.91e-41.29e-30.01440.01500.000550.00555

Both densities are close to the SCF density in every measure. On all four molecules the variationally trained model's density is the closer one in ‖ΔP‖/‖P‖ and Δq, and Δμ is equal within noise, consistent with what the two objectives train: the variational loss is a functional of the density itself, the supervised loss fits Fock-matrix elements. Already in the paper: the model density halves the SCF iteration count as an initial guess (Tables A8–A9), a use Zhang et al. (2024; 33–46% fewer iterations) and DEQHNet (26–35%) reported before us; the revised Appendix I credits both. The table above is the new Appendix L (Table A13).

Thank you for the thorough and thoughtful responses to my questions and concerns. I appreciate the additional experiments and clarifications, which have been very helpful in understanding the scope and contributions of the work.

The responses have addressed several of my concerns, and I will maintain my original score.

Thank you for your thoughtful feedback and careful consideration of our responses. We appreciate your time and are glad that the additional experiments and clarifications helped address your concerns.

Reviewer_V6BQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper trains an electronic-structure model without precomputed DFT labels by minimizing the Kohn–Sham energy of its own predicted density over classical-MD geometries. An eigendecomposition-based parameterization guarantees an admissible density, and forces are computed at inference from one Fock build and the analytic gradient, without SCF iteration. The authors claim this removes the labelling cost that conventional MLFFs shift to dataset construction. Validation is on rMD17 and 3BPA, with a same-network supervised comparison and appendix experiments.

Critical strengths
  • The label-free variational objective with admissibility by construction is a conceptually clean idea and largely beneficial once its robustness and generality are demonstrated.
Critical weaknesses
  • The benchmark choice is largely misaligned with the core claim. The method is motivated by the cost of DFT labeling, but it is mainly evaluated on rMD17, where DFT is cheap and the benchmark is already saturated with ML-based models. The cost comparison is made against a supervised pipeline built for this paper, not against how MLFFs are actually trained. The claim is general, but the evidence is narrow and comes from easy settings.
  • The evidence is not reliable. The out-of-distribution and matched-sampling claims are confounded by differences in sampling and data quantity relative to the baselines. The results are single-seed and lack uncertainty, and the precision reported in Table 1 is limited, so the table cannot support the model comparisons. Several quantitative statements lack supporting data, and key experimental settings are missing.
  • The core method lacks ablation. Components central to the method, such as the minimal-basis anchor and the parameterization choices, are not ablated. It is therefore unclear what drives the reported performance.
  • The core method lacks theoretical grounding. There is no complexity analysis of the variational objective. There is no evidence of training stability or convergence to the ground state. Whether the forces are consistent with the model’s own energy is not addressed. For a paper proposing a new framework, these omissions are critical.
Questions for authors
  1. Can you demonstrate the claimed advantage in a regime where DFT labelling is genuinely expensive, against standard MLFF pipelines at their actual cost? For example, the OMol25 [1] dataset provides an electronic-structure subset of closed-shell molecules, including DFT compute cost for each instance.
  2. Do the main accuracy claims hold under comparisons that control sampling and data quantity, across multiple seeds?
  3. Which components contribute to the performance, and to what extent? In particular, what happens without the minimal-basis anchor?
  4. What is the computational complexity of the objective, what ensures stable convergence to the ground state, and are the reported forces the gradient of the model’s energy?

[1] Levine, D. S., Shuaibi, M., Spotte-Smith, E. W. C., Taylor, M. G., et al. The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models. arXiv:2505.08762, 2025. v1

3 comments

Response to Reviewer V6BQ (part 1 of 2)

W1 / Q1 — Benchmarks misaligned with the claim; OMol25; standard pipelines at their actual cost

Where a pretrained model matches the level of theory a study requires, or generalises well to the target system, using it is the efficient choice. The paper addresses the other case, labels at a different level of theory or a target the pretrained model was not trained for, where the user has to label the target molecules at the target level and train or fine-tune on them; that labelling is what our method removes. On OMol25's own data the labels are already paid for, and variational training re-performs the labelling work (one Fock build per frame visit), so we claim no advantage there.

Every comparison in the paper (Tables 1, 2, 3 and A6) is against models trained on labels of the same molecules at the same level of theory, the pipeline of that case; in Table 2 those labels take a third (ethanol) to four fifths (aspirin, azobenzene) of the supervised pipeline's GPU time. Our training costs the same order of GPU time as labelling the same geometries (revised Section 6; Section 5.3's sentence suggesting otherwise is removed).

We ran that pipeline for the eight galectin-8 ligands of Appendix E (52–64 atoms, ωB97X/6-31G(d)): the 11,875 geometries per ligand our model consumed were labelled at the paper's reference settings, MACE-OMOL-0 (extra-large, pretrained on OMol25) was fine-tuned on all 95,000 labels with its maintainers' protocol, and both were scored on the held-out frames and references of Table A5 (revised Appendix E). Force MAE (meV/Å), energy MAE and energy MAE after removing the per-ligand offset (meV), 100 frames per ligand:

ligFligOMeligNHMeligNMe2ligOEtligPyrligOHligNH2mean
MACE-OMOL-0 fine-tuned, force2.142.032.052.102.032.092.042.03
Ours, force0.940.910.950.940.930.960.900.93
MACE-OMOL-0 fine-tuned, energy2.119.514.6020.163.689.939.908.84
Ours, energy0.310.320.340.370.360.410.310.32
MACE-OMOL-0 fine-tuned, centred energy1.871.531.651.741.842.071.741.84
Ours, centred energy0.060.060.060.080.070.070.060.07

Our model is the more accurate on every ligand, on forces and on energies, with none of the per-ligand offsets that dominate the fine-tuned model's energy error. Its accuracy is set by training time, not by a label count fixed in advance.

W2 / Q2 — Confounded comparisons; single seed; precision; missing settings

We gave MACE our training data outright: maintainers' rMD17 settings, our 94,950 ethanol geometries (1600 K) with a converged label for each, 20 epochs (1.9 million frame visits, twice ours), scored on the same 1,000 test frames against our converged SCF. MACE is one training run; our row is the mean ± SD over three seeds (Table A15; every other table reports seed 11, 0.511 meV/Å, the best of the three):

ethanol modelforce MAE (meV/Å)energy MAE (meV)
Supervised MACE (94,950 labels), same geometries1.420.57
Variationally trained QHFlow2 (ours, 3 seeds), same geometries, no labels0.64 ± 0.140.010 ± 0.002

With geometries, their number and sampling temperature fixed, and a label for each, MACE stays above every seed, so Table 1's differences are not produced by data quantity or sampling alone; Tables 2 and A6 show the same within one network.

Seeds: the same three seeds give 1.24 ± 0.20 meV/Å on malonaldehyde and 1.67 ± 0.11 on aspirin (energies 0.038 ± 0.008 and 0.142 ± 0.007 meV), an order of magnitude above the frame standard error, so single-seed differences below about two seed SDs are not established. The comparison above and the 3BPA comparisons (Table A6) lie far outside that margin; the ethanol and malonaldehyde gaps of Table 2 do not (new Appendix N). The revised paper reports the ethanol mean ± SD (Table A15), frame standard errors (Table A2) and the missing settings (Table A1: seed, test-set sizes, grid, precision, GPU types).

W3 / Q3 — Ablations: the minimal-basis anchor and the parameterisation

The force readout is already ablated in Table A7 (Appendix H).

Anchor: two ethanol variants in the Table A1 configuration, no anchor (the network output is the predicted matrix) and the anchor as an input feature only, 1,000 test frames against our converged SCF:

outputforce MAE (meV/Å)energy MAE (meV)one Roothaan step (meV/Å)
MINAO anchor residual (paper)0.511 ± 0.0050.00770.508
no anchor0.952 ± 0.0090.02610.577
anchor as an input feature only1.169 ± 0.0080.02900.632

Response to Reviewer V6BQ (part 2 of 2)

The anchored model is better than the no-anchor variant on 96.7% of frames and than the input-feature variant on 99.8%, yet the no-anchor model trains to within 2× of it on forces and 3.5× on energies: the anchor is a better starting point, not the source of the physics. It is a prior-work design (QPT; Zhang et al., 2024); the ablation is the new Appendix K.

Parameterisation: Eq. (4) is the construction SCF itself uses, every output is an admissible density, and Eq. (6) has only occupied–virtual denominators; we did not ablate it.

Loss: Li et al.'s self-consistency residual ‖H_pred − F[P(H_pred)]‖² with weight λ (rebuilt target detached) makes the raw predicted matrix a Fock operator but does not improve the reported force of Eq. (5), which never uses its eigenvalues. Aspirin, 1,000 test frames, force MAE against our converged SCF, paired per-frame difference (within a pair only λ differs):

configurationλ = 0 (meV/Å)λ = 300 (meV/Å)paired difference (meV/Å)
13.7M-parameter model, full step budget2.21 ± 0.102.07 ± 0.09−0.145 ± 0.026
42.8M-parameter model, half step budget2.50 ± 0.102.81 ± 0.11+0.308 ± 0.020

The term moves the force by 7% in one direction and 12% in the other; it is not what makes the method work (Appendix K).

W4 / Q4 — Complexity, stability, convergence to the ground state, forces vs. energy

Ground state. Every stationary point of the objective is a self-consistent Kohn–Sham solution, and the variational bound ranks them, so descent selects the ground state. The bound: E(P) ≥ E₀(R) for every admissible P (symmetric, PSP = P, correct trace), with equality only at the ground state, and Eq. (4) makes P_θ admissible for every network output. Stationarity: by the chain rule through Eq. (4), dE/dM_θ is a sum over occupied–virtual pairs of F_ai/(ε_i − ε_a) times eigenvector products, which vanishes if and only if the occupied–virtual block of F[P_θ] vanishes in the eigenbasis of M_θ, the self-consistency condition. A self-consistency residual alone vanishes at every such solution, ground state or not, and ranks none of them; this is why the energy is the objective.

Stability. The only denominators of Eq. (6) are occupied–virtual gaps of M_θ (Appendix G). The objective does not constrain that gap, but the MINAO anchor starts from a matrix with a physical gap and the trained models keep one: the smallest over 100 test frames of each Table 2 molecule is 1.64 eV (azobenzene) (revised Appendix G).

Complexity. One step on B geometries costs B Fock builds (O(N²N_aux) density fitting plus the exchange–correlation quadrature), B O(N³) eigendecompositions and one network forward and backward pass; the backward pass through the physics is free, since ∂E/∂P = F. Training performs one Fock build per frame visit, at our budget the Fock builds of labelling every training geometry at about ten SCF iterations (revised Section 6). Inference is one MINAO guess, one forward pass, one Fock build and one gradient.

Forces vs. energy. The reported force (Eq. 2 at P_θ, canonicalised by Eq. 5) omits the response term Tr(F[P_θ] dP_θ/dR), so it is not the gradient of E_θ(R) = E_KS(R, P_θ(R)). Central finite differences (±0.01 and ±0.02 Å, Richardson-extrapolated, network re-run), 100 test frames per molecule, every component, mean ± SE:

aspirinethanol
Reported force vs. converged DFT force, meV/Å1.82 ± 0.030.53 ± 0.01
Reported force vs. exact gradient of the model energy, meV/Å2.02 ± 0.030.72 ± 0.01
Exact gradient of the model energy vs. converged DFT force, meV/Å0.087 ± 0.0020.013 ± 0.0003

The reported force is not conservative at the level of its own error; the exact gradient, conservative by construction, is an order of magnitude closer to DFT, as the objective implies (second order against first order in the density error). The revised Section 6 and the new Appendix J report both; the analytic exact gradient is the next step. Two sentences changed: Section 5.4's finite-difference sentence described a check of the reference forces and is removed, and the grid-response term claimed in Section 4.3 and Table A1 was omitted for model and reference alike; Table A1 and Appendix J now say so.

Presentation. The revised Section 2 states which parts follow QPT (objective, density construction, anchor, geometry pool) and which are specific to a force field (force readout, backward pass). Section 4 then proceeds objective (4.1, Eq. 3) → admissible density from the network's matrix (4.2, Eq. 4) → energies and forces read off that density without SCF iteration (4.3, Eq. 5) → architecture, geometry source and efficient inference (4.4); Appendix G is the backward pass, Appendix H the force-readout ablation.

I thank the authors for their response. Although the authors added several experiments, they are insufficient to address the concerns raised. Most of them still evaluate only a few molecules. The required improvements to the main results are not reflected, and small-scale experiments are instead added only in the response. I maintain my score.

Q1. OMol25 labels can serve as the reference for evaluation, so their availability is not a reason to omit the experiment. The added result in the revised Appendix E is based on eight ligands from one congeneric series with a small basis. It also reports no cost for either pipeline. This does not demonstrate the method in a regime where labeling is genuinely expensive.

Q2. Table 1, the main result, remains at one-decimal precision for a single best-seed run. Its precision and seed statistics must be improved in Table 1 itself rather than only in the new appendix tables. The added comparison in Table A15 is based on a single molecule and a single MACE run, which hardly supports a comprehensive evaluation of the proposed method. Appendix N generalizes the seed spread of ethanol to all of Table 1, which the data does not support.

Q3. The anchor ablation in Appendix K is based on a single small molecule and a single seed. The parameterization was not ablated.

Q4. The complexity analysis appears only in the rebuttal and not in the paper. The argument for convergence to the ground state does not hold as a guarantee for a non-convex objective with parameters shared across geometries. Stability is supported only by the gap of the final model on 100 test frames, with no training dynamics reported. Appendix J confirms that the reported forces are inconsistent with the model’s own energy, and this remains unresolved.

Reviewer_rH2xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a recipe for training ML force fields without precomputed DFT labels for energies/forces. Instead of label supervision, training directly minimises the Kohn–Sham energy of the electron density matrix predicted by the NN, over geometries sampled from classical MD. At inference, it allows to predict forces by rebuilding the Fock matrix from the density predicted by the network, and applying the analytic gradient (no SCF iterations required).

On the rMD17 benchmark, the proposed method shows lower force error than four MLFF baselines on 9/10 molecules, and on 3BPA it extrapolates to higher temperatures far better than supervised baselines. Against supervised training of the same network and on the same geometries, it reaches similar or better performance at 12-17% lower total cost. It also runs stable MD at a third of DFT’s cost per step.

Critical strengths
  • The reported experiments and evaluations are conducted in a rigorous way. I appreciated the authors reporting, in tables, the scores of independent converged SCF against the benchmark labels, which sets the noise floor clearly (and is often an overlooked detail). Ablation on the training signal in Table 2 is clean. Sampling breadth is well separated from the objective, with the 300 vs 1600K comparison on 3BPA cleanly showing how much robustness comes from broader geometries and how much from the loss itself. Limitations of the method are also framed fairly.
  • The experiments cover a wide range of settings, with results on rMD17, 3BPA, a 42-atom peptide MD run and a multi-ligand model at 50+ atoms. Using the prediction as an SCF initial guess (about 50% fewer iterations) is a nice practical side result.
  • I appreciated the authors providing a thorough diagnostics of the numerical aspects of the method (gating the backward pass against finite differences, verifying that forces don’t depend on the unconstrained blocks of the network output, ablating the force readout, checking that every frame across all evaluations gives a valid density). The explanation of why plug-in eigenvalues fail will be useful to anyone building on Hamiltonian predictors.
Critical weaknesses

Major

  • The core idea in the proposed method (using the KS energy of the model’s predicted electronic state as minimisation objective for training) is essentially the same objective as in Mathiasen et al. 2024 (https://arxiv.org/abs/2402.04030). Very similar ideas also appear in Li et al., PRL 2024 (https://arxiv.org/abs/2403.11287), and, with a self-consistency loss, in Zhang et al., ICML 2024 (https://proceedings.mlr.press/v235/zhang24ak.html). The QPT method from Mathiasen et al. also shares several implementation and engineering similarities (MINAO residual output, the prefetched pool of reused geometries), on top of the core framing (merging labelling and training). While I recognise that the force-field evaluation angle was not explicitly tackled in previous literature, the novelty of contribution 1 must be reframed, and all these previous papers (none of which currently appears in the bibliography) compared against. Swapping in the self-consistency loss on the same backbone would be an easy and informative ablation.
  • The analysis reported in Table 1 confounds several factors, and doesn’t establish a clear cost-benefit trade-off. The variational models train on 95k geometries against the 950 labelled datapoints for the baselines. By Table 2, labelling all 95k aspirin geometries has comparable costs to the full variational training, so the baselines’ 950 labels correspond to ~1% of the electronic-structure compute. The models also differ in geometry distribution and architecture. They evaluate the DFT functional at inference at about a third of the cost of DFT per step, whereas the baselines are orders of magnitude cheaper per step. The authors should report training and inference costs for the baselines, to compare against the numbers given for the variational method. Without them, the table doesn’t show what the accuracy gain costs, in practice. Showing where the Pareto frontier lies (accuracy vs total compute) would make the results much easier to interpret, and much more valuable. It is unclear to me what’s the practical niche that the proposed method is aiming to address. Since the proposed models are trained per molecule, the paper should state the break-even point, to clarify when a user should prefer this method to running DFT-MD directly (or even extended-Lagrangian BOMD without SCF optimisation, that should be included in the comparison as a training-free alternative) or fitting a cheap MLFF.

Minor

  • The headline 5.6 versus 62.0 meV/A on 3BPA, reported in the abstract, compares a model trained at 1600K against baselines trained at 300K. The matched comparison (20.8 vs 62.0) is the fairer number that should come first. The abstract also describes the model scoring at 5.6 as “far outside the training distribution”, although it was trained on 1600K geometries and tested at 1200K, if I understand correctly.
  • Equation (2) leaves out the density-response term when evaluated at non-stationary densities: the resulting forces are therefore neither exact Born-Oppenheimer forces nor the gradient of the model’s own energy. The authors should discuss this and quantify how non-conservative the forces are, beyond the limited analysis of the 10 ps NVE test on one molecule currently given.
  • No measure of run-to-run variance.
  • Transferability across chemistry is barely tested.
  • The degeneracy-safe backward pass is standard engineering: it should be credited to existing differentiable quantum-chemistry tools.
Questions for authors
  1. Can you please explain how the proposed method differs from the previous papers mentioned above (QPT, Li et al. 2024, self-consistency training)? How does your variational loss compare with the self-consistency loss, when deployed on the same backbone?
  2. Can you quantify how large is the omitted density-response term in eq. (2), measured as the difference between the reported force and the exact gradient of the model's predicted energy? Is your approximation enough for MD to conserve energy across a longer NVE run?
  3. Can you provide a comparison of your model against the reported baselines that shows the overall training time for all methods and corresponding predictive performance, in a comparable setting (e.g., MACE trained on DFT labels at a total labelling+training budget equivalent to that of the variational training)?
  4. For the Ac-Ala3-NHMe model, what was the training cost and what is the minimum trajectory length needed for the proposed method to pay off against DFT-MD?
6 comments

Response to Reviewer rH2x (part 1 of 2)

Major weakness 1 / Q1 — Novelty relative to QPT, Li et al., Zhang et al.; the self-consistency loss

The revised paper cites QPT (Mathiasen et al., 2024), Li et al. (PRL 2024), Zhang et al. (ICML 2024), D4FT (Li et al., 2023) and DEQHNet (Wang et al., 2024), and credits them in Sections 1 and 2 (in blue) for the objective, the eigendecompose-and-occupy density construction, the minimal-basis anchor and the geometry pool.

What is new is what turns the objective into a force field, which none of them does. (i) Forces: none of them computes one. The loss reaches the predicted matrix only through its occupied subspace, so its eigenvalues are never trained (Section 4.2); forces read from them are 518× worse than ours and the uncanonicalised Rayleigh-quotient readout 132× worse (Table A7). The canonicalised readout of Eq. (5) is ours. (ii) Single-pass inference: QPT follows the network with three DFT iterations and DEQHNet iterates it to a fixed point; ours is one forward pass, one Fock build, one gradient. (iii) A backward pass through the eigendecomposition with only occupied–virtual denominators (Eq. 6), where Zhang et al. clip eigenvalue differences and skip steps; this is a primitive, not a new idea, and the revised Section 4.2 and Appendix G cite PySCFAD (Zhang & Chan, 2022) and Kasim (2020).

Self-consistency loss. We added Li et al.'s residual ‖H_pred − F[P(H_pred)]‖² with weight λ to the variational energy (rebuilt target detached, one seed per run). It makes the raw predicted matrix a Fock operator but does not improve the reported force of Eq. (5), which never uses that matrix's eigenvalues. Aspirin, complete 1,000-frame test set, force MAE against our converged SCF, with the paired per-frame difference (within a pair only λ differs):

configurationλ = 0 (meV/Å)λ = 300 (meV/Å)paired difference (meV/Å)
13.7M-parameter model, full step budget2.21 ± 0.102.07 ± 0.09−0.145 ± 0.026
42.8M-parameter model, half step budget2.50 ± 0.102.81 ± 0.11+0.308 ± 0.020

The term moves the force by 7% in one direction and 12% in the other; it is not what makes the method work (new Appendix K).

Minor weakness 2 / Q2 — Eq. (2) at a non-stationary density; non-conservative forces; longer NVE

We added an experiment measuring the difference between the reported force and the exact gradient of the model energy E_θ(R) = E_KS(R, P_θ(R)): central finite differences (±0.01 and ±0.02 Å, Richardson-extrapolated, network re-run at every displaced geometry) on 100 aspirin and 100 ethanol test frames, every Cartesian component. Mean absolute difference, mean ± standard error:

aspirinethanol
Reported force vs. converged DFT force, meV/Å1.82 ± 0.030.53 ± 0.01
Reported force vs. exact gradient of the model energy, meV/Å2.02 ± 0.030.72 ± 0.01
Exact gradient of the model energy vs. converged DFT force, meV/Å0.087 ± 0.0020.013 ± 0.0003

The reported force is not the gradient of the model energy: its distance from that gradient exceeds its own error against DFT. The exact gradient, conservative by construction, is an order of magnitude closer to the DFT force, as the objective predicts (E_θ − E_SCF is second order in the density error, whereas Eq. (2) at an approximate density is first order). Computing the exact gradient analytically needs the derivative of the network output and of the anchor and overlap integrals with respect to the nuclei, which we have not implemented; by finite differences it costs 12N energy evaluations per geometry. The revised Section 6 and the new Appendix J report this.

Two sentences of the paper changed. Section 5.4's "finite differences reproduce our analytic forces to 1.8×10⁻⁴ meV/Å" described a check of the reference forces, not the model's; it is removed, and Section 5.4 now points to the Conv. DFT column of Table 3 for the distance between the 3BPA labels and our converged SCF. Section 4.3 and Table A1 said the grid-response term is included; model and reference forces were both computed without it (grid level 3). Table A1 and Appendix J now say so; the term is identical for model and reference, so it cancels in every model-vs-DFT number, and the finite differences above include it on both sides.

Response to Reviewer rH2x (part 2 of 2)

Longer NVE: a 100 ps microcanonical Ac-Ala3-NHMe run from an independent start, scored with the Table A4 estimator:

run lengthtotal-energy drift (meV/atom/ps), 95% CI
Variationally trained QHFlow2 (ours)100 ps+0.040 [+0.025, +0.055]
DFT10 ps+0.025 [−0.071, +0.121]

The model's drift is resolved from zero and lies inside DFT's own interval. No bond broke, and the force error against DFT along the 100 ps thermostatted run shows no trend with time.

Major weakness 2 / Q3 — Table 1 confounds data quantity with cost; baselines' costs; practical niche

We trained MACE with its maintainers' rMD17 settings on our 94,950 ethanol geometries (1600 K), each with a converged energy and force label, and scored both models on the same complete 1,000-frame test set against our converged SCF. MACE is one training run; our entries are the mean ± SD over three training seeds. Costs on the GPU of Table 2, MACE trained alongside the same GPU-occupancy process as our runs:

ethanol, same 94,950 training geometriesSupervised MACE (94,950 labels)Variationally trained QHFlow2 (ours, 3 seeds)
Force MAE (meV/Å)1.420.64 ± 0.14
Energy MAE (meV)0.570.010 ± 0.002
Labelling (GPU-h)11.00
Training (GPU-h)8.817.6
Total (GPU-h)19.817.6

With the same geometries and a label for each, MACE stays above every seed of our model at a comparable end-to-end cost (Table A15, new Appendix N, pointed to from Section 5.3; Table A1 names the GPU of every cost).

Extended-Lagrangian BOMD still performs one Fock build and one nuclear gradient per step, the same operation count as our inference, so its per-step cost is in our model's class rather than warm-started DFT's, and it applies to sequential trajectories only; we did not implement it. Per-geometry inference costs of a pretrained MLFF, our model and warm-started DFT on one GPU are in the new Appendix M, and the relation to pretrained force fields is in the revised Section 6.

Major weakness 2 / Q4 — Break-even against DFT-MD for Ac-Ala3-NHMe

Training took 42.3 GPU-h on an H200 (31.3 h of steps and about 11 h of validation); Table A4's MD timings are on an L40S. We re-timed the MD on an H200 with the same checkpoint, start state, integrator and thermostat, DFT warm-started from the previous step's density:

training on H200, MD on L40S (Table A4)training and MD on H200
Training (GPU-h)42.342.3
Warm-started DFT-MD (s/step)3.721.90
Variationally trained model (s/step)1.090.71
Saving per step (s)2.631.19
Break-even trajectory (ps)58128

The 100 ps run of Appendix D is past break-even in the first accounting and short of it in the second; the revised Appendix D states both. "Pooling two further starting states" now reads one further starting state.

Minor weaknesses 1, 3, 4, 5 — abstract; run-to-run variance; transferability; backward pass. Abstract: it now states the matched 3BPA comparison (both models trained on 300 K geometries, 20.8 against 62.0 meV/Å at 1200 K); the 1600 K number and "far outside the training distribution" are removed, likewise in Section 1. Variance: frame standard errors are in Table A2, the three-seed spread in Table A15, and Table A1 states the seed. Transferability: every model of Sections 5.2–5.4 is per molecule; Appendix E trains one model across eight ligands and tests on each, which the revised Section 6 describes as one model serving several molecules rather than transfer to unseen ones; a transferable model is the same training over a pool of molecules, which needs only geometries. Backward pass: the mathematics is classical (PySCFAD Section II.C; Pople et al., 1979; Kasim, 2020; Kasim & Vinko, 2021; Tamayo-Mendoza et al., 2018), and the revised Section 4.2 and Appendix G cite them; the one difference is that differentiating the projector removes every occupied–occupied and virtual–virtual term with no threshold, where their eigenvector-level passes exclude degenerate pairs by an equality mask.

I thank the authors for the thorough revision, which addresses most of my concerns, and I appreciate the corrections they volunteered. A few issues remain open in my view.

  1. The new Appendix J shows that the reported force is further from the model's own energy gradient than from DFT, and that the exact gradient is about 20–40× closer to DFT. The proposed readout is therefore weaker than the exact gradient of the model energy, which can be obtained by backprop (a single backward pass given differentiable integrals, rather than 12N finite-difference evaluations), as Hassan et al do to drive their sampling. Section 6 mentions this result, but it should be stated clearly as a limitation of the proposed force readout. Ideally, forces from the exact gradient should be reported.
  2. I don’t see the the 100ps NVE drift (reported in the rebuttal above) in the updated paper. Sec 5 and Appendix D still state that the drift is not resolved, while the rebuttal reports a drift resolved from zero. Falling inside the interval of a 10 ps DFT run does not establish equivalence. Please add the 100 ps result, correct the text, and ideally include a DFT NVE run of matching length.
  3. Every table still reports the best seed of three on ethanol, where the seed standard deviation is about 22% of the mean. Was this seed fixed before training? At that spread, some of the results (e.g. the malonaldehyde gap in tab 2 and the lambda effects in tab A12) likely fall within seed variance -- can they really be presented as conclusive?
  4. The conclusion's "order of magnitude more accurate" on 3BPA still uses the unmatched 1600K comparison.

I would be inclined to raise my score if these points are addressed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.