acceptodds
Under review as a conference paper at ICLR 2027

Equivariant Flow Matching for Electron Density Prediction

Abstract

Machine learning surrogates for density functional theory (DFT) have been increasingly used to reduce the cost of first-principles calculations. In this arena, predicting real-space electron densities offers a scalable and transferable initialization for self-consistent field (SCF) procedures. However, current methods face a clear dilemma. That is, grid-based architectures incur a high computational cost, while basis-set methods fail to capture the structural correlations inherent in the coefficient space. Here, we develop OrbFlow, an -equivariant generative model that predicts Gaussian-type orbital (GTO) coefficients via flow matching. OrbFlow retains the efficiency of a compact atom-centered basis while replacing pointwise regression with a learned probability path over the full coefficient space. It is trained through a two-phase trajectory curriculum that mitigates discretization drift during numerical integration. OrbFlow achieves state-of-the-art accuracy on QM9, reducing density error by 13.6% relative to the previous best model, and reduces error by 51% to 63% on every molecule of the MD benchmark relative to the strongest prior method sharing its basis. The predicted density also cuts SCF iterations by up to 68% with zero-shot transfer to unseen exchange-correlation functionals and recovers dipole and quadrupole moments to within a few percent of DFT references without any SCF calculation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
electron density prediction, equivariant flow matching, Gaussian-type orbital basis, SE(3) equivariance, density functional theory, quantum chemistry, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ae4aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets electron density prediction. They use the GTO auxiliary basis formulation of SCDP (Fu et al., 2024), which regressed the GTO coefficients with an eSCN architecture, by an EquiformerV3 (Liao et al., 2026) trained with flow matching. They use two flow steps during generation.

Critical strengths
  1. Slight improvement (~10% lower NMAE) on QM9 over SCDP + VO, L = 6 (Fu et al., 2024). But it is unclear if that is because of the different architecture or because of using two flow steps instead of direct prediction, especially since the proposed method is 1.6x slower.
Critical weaknesses
  1. No innovation. The paper swaps out the backbone for a newer architecture, and uses two flow steps instead of directly regressing
  2. The writing is verbose and claims known tricks it's innovations. E.g.
  • the "Two-Phase Trajectory Curriculum Learning" is half a page in methods, when they are really just training on the same flow times that they use during inference (Euler integration).
  • using equivariant networks ("Equivariant Flow and Isotropic Irrep Priors" in methods) is hardly their invention, but doesn't even cite any prior work.
  • verbose AI language throughout the paper
Questions for authors
  1. Does table 4 use a different validation set then table 1? If yes, can you add SCDP + VO, L = 6 (Fu et al., 2024) to table 4, to compare it to the direct regression (deterministic, single-step loss)?
  2. I don't understand the ablation of flow steps (Euler integration). It seems to sweep over training integration steps (phase-2) and inference steps separately. Why? Wouldn't it make more sense to keep them tied, and test more flow steps together? I don't see why more flow steps could reduce the accuracy
4 comments
AuthorsOpenReview ↗

Response to Reviewer ae4a

We thank the reviewer for the clear questions. The new experiments show that OrbFlow's gain over SCDP+VO comes from how it is trained, not from the newer backbone, and that on QHFlow2's network the same training beats QHFlow2's own recipe.

Slight improvement (~10% lower NMAE) on QM9 over SCDP + VO, L = 6 (Fu et al., 2024). But it is unclear if that is because of the different architecture or because of using two flow steps instead of direct prediction, especially since the proposed method is 1.6x slower.

Regarding the size of the improvement, OrbFlow lowers the error of SCDP+VO by 14.6% on QM9 (0.178% to 0.152%) and by 51% to 63% on each of the six MD molecules (Table 2). The gain is largest on MD and its size depends on the dataset, but it is consistent, since OrbFlow improves on SCDP+VO on QM9 and on every MD molecule.

Regarding the 1.6× slowdown, this is the MD latency (0.066 s against 0.041 s per molecule, Table 12), and on QM9 the ratio is 1.4× (0.33 s against 0.23 s, now in Section 4.5). The overhead comes from the Euler integration of the flow. OrbFlow evaluates the network once per Euler step, so its two-step rollout needs two forward passes where direct regression needs one. This adds 0.025 s per molecule on MD and 0.10 s on QM9, while building the density on the grid costs the same for both models.

The gain comes from the training, not from the architecture, and two flow steps alone do not explain it. We show this in two ways.

First, Table 4 fixes the backbone, basis, loss and training budget and changes only the starting point and the training loss. Like our other ablations, it uses a fixed 10% subset of QM9 (12,384 training and 1,000 test molecules), because a full QM9 run costs about 1,600 GPU-h and each ablation needs three seeds. All values are NMAE in %.

StartSingle-step lossRollout loss
Deterministic (from zero)0.280 ± 0.003 (direct regression)0.268 ± 0.003
Noise0.295 ± 0.0160.240 ± 0.005 (OrbFlow)

A deterministic second pass helps little (0.268%), and starting from noise alone does not help (0.295%). Only both together, a noise start and training on the model's own two-step rollout, lower the error by 14% (0.280% to 0.240%).

Second, we separated the backbone from the training on full QM9 by training EquiformerV3, OrbFlow's backbone, by direct regression (500k steps, against 450k for OrbFlow). We also trained SCDP+VO on the 10% subset, as the reviewer asks below. All values are NMAE in %.

BackboneTrainingFull QM910% subset
eSCN (SCDP+VO)Direct regression0.1780.327
EquiformerV3Direct regression0.1740.280 ± 0.003
EquiformerV3OrbFlow training0.1520.240 ± 0.005

On full QM9, changing only the backbone lowers the error from 0.178% to 0.174%, while changing only the training lowers it from 0.174% to 0.152%. About 85% of the improvement therefore comes from the training. On the 10% subset, at the same 90k-step budget as all other ablations, eSCN is further behind (0.327% against 0.280%). Its validation error was still falling at the end of training, so part of this gap is slower convergence. The training gives a similar reduction on top (0.280% to 0.240%). With the full training set, both backbones converge, the backbone difference almost vanishes, and the gain comes from the training.

No innovation. The paper swaps out the backbone for a newer architecture, and uses two flow steps instead of directly regressing

As shown above, neither the backbone nor the extra step explains the gain. The contribution is the training. The model starts from noise and is trained on the same few-step rollout it runs at inference. The change is simple, as are consistency models (Song, Dhariwal, Chen and Sutskever, 2023) and shortcut models (Frans et al., 2025), which also change only how a flow model is trained and keep the network. Like them, it matters because of what it gains, namely 14% at a fixed backbone and 85% of OrbFlow's improvement over SCDP+VO on full QM9.

The training also works on other networks. On QHFlow2's network (details in our reply to Reviewer cVZC), it reaches 0.240%, against 0.255% for QHFlow2's own flow-matching recipe. To our knowledge, training the few-step rollout of a flow model on the density loss is new for density and Hamiltonian prediction. QHFlow and QHFlow2 train only at single interpolated times.

AuthorsOpenReview ↗

Continued Response to Reviewer ae4a

the "Two-Phase Trajectory Curriculum Learning" is half a page in methods, when they are really just training on the same flow times that they use during inference (Euler integration).

We agree with this description. Phase 2 trains the model on the same Euler steps it uses at inference, and the revised method section describes it more briefly.

We see this simplicity as a practical advantage rather than a weakness. Phase 2 keeps the same network and the same parameters, and adds no teacher, no extra conditioning input and no inference cost. The model is run with the same two steps before and after it. This change alone lowers the error on the 10% subset from 0.295% (Phase 1 only) to 0.240%, and from 0.280% for direct regression with the same backbone. The two phases also matter in practice, since starting Phase 2 from Phase 1 reaches the accuracy of the alternatives in Table 5 with less compute (74 GPU-h against 81 to 182).

using equivariant networks ("Equivariant Flow and Isotropic Irrep Priors" in methods) is hardly their invention, but doesn't even cite any prior work.

We agree and now cite this work (Köhler et al. 2020, Xu et al. 2022, Hoogeboom et al. 2022, Klein et al. 2023, Song et al. 2023). The revision presents the isotropic prior over the irreps of the GTO coefficients as a design choice rather than a contribution, and Proposition 1 is now a short remark.

verbose AI language throughout the paper

As stated in our LLM usage disclosure, language models were used to organize and polish the text, and all content and results were checked by the authors. In the revision we removed claims the experiments do not test, such as "eliminate discretization drift" and "straightens the flow field", replaced general statements with numbers, and shortened the method section and the introduction. We would welcome pointers to passages that still read as verbose.

Does table 4 use a different validation set then table 1? If yes, can you add SCDP + VO, L = 6 to table 4, to compare it to the direct regression (deterministic, single-step loss)?

No, both tables use the same 50 validation molecules. Table 4 trains and tests on the fixed 10% subset of the standard training and test sets, which is why its errors are higher. As requested, SCDP+VO (L=6) trained on the same subset with the same budget is in the second table of our first answer.

It seems to sweep over training integration steps (phase-2) and inference steps separately. Why? Wouldn't it make more sense to keep them tied, and test more flow steps together? I don't see why more flow steps could reduce the accuracy.

The sweep over training integration steps is already tied, since a model trained with K steps is also tested with K steps. On the 10% subset it gives 0.296%, 0.240% and 0.240% for K = 1, 2 and 3 (K=2 is OrbFlow, K=1 and K=3 are the last two rows of Table 14 in the revision), so accuracy saturates at K=2. We will add K=4 to the final version as suggested, and the caption now says that the steps are tied.

More steps hurt only in the separate inference sweep (Appendix E.4), which runs the K=2 model with more steps than it was trained with. The network predicts the endpoint, so the last Euler step returns its prediction at t = 1 − 1/K. Phase 2 trains t = 0.5 but no later time, so the error rises with more steps (0.539% at K=4). A model trained only with the single-step loss does not degrade with more steps (0.290% at K=2, 0.284% at K=16), but it never beats regression (0.280%). OrbFlow should therefore be run with the number of steps it was trained with.

We hope these controlled experiments address the reviewer's main concern, and we would be grateful if the reviewer would reconsider the score in light of them.

AuthorsOpenReview ↗

# Follow-up to Reviewer ae4a

We thank the reviewer again for the helpful review. As the discussion period ends in two days, we briefly recall our main answers.

  1. Architecture or training. With EquiformerV3 fixed, direct regression reaches 0.174% on full QM9 and OrbFlow 0.152%, against 0.178% for SCDP+VO, so about 85% of the improvement comes from the training. Two flow steps alone do not explain it, since a deterministic two-step model reaches 0.268% against 0.240% for OrbFlow (Table 4).
  2. Novelty. The contribution is a training method that trains the model on its own few-step rollout from noise. It keeps the network unchanged and lowers the error at a fixed backbone by 13% on full QM9.
  3. Speed. The 1.6× latency comes from the two Euler steps, the smallest number at which accuracy saturates.
  4. Citations and writing. The revision cites prior equivariant flow models and removes claims the experiments do not test.

We would be very grateful to know whether these answers address the reviewer's concerns, and we are happy to clarify anything further.

Dear Reviewer ae4a,

We understand this is a busy time and thank you again for your review. As the discussion period ends tomorrow, we would be very grateful to hear your thoughts on our responses, in particular whether the evidence that most of the improvement comes from the training, at a fixed backbone, addresses your concerns about novelty and contribution. If any questions remain, we would be glad to clarify them before the deadline.

Thank you very much for your time.

Reviewer_cVZCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes OrbFlow, which represents electron density using coefficients in an auxiliary GTO basis and predicts them through SE(3)-equivariant flow matching. An EquiformerV3 backbone takes the molecular structure, current coefficient state, and time as inputs and predicts the target coefficients. The final density is obtained through a two-step rollout starting from an isotropic noise prior. Training consists of Phase 1, which predicts the endpoint from states interpolated between noise and the target coefficients, and Phase 2, which differentiates through the same rollout used at inference to minimize the final density error. The authors report a density NMAE of 0.152% on QM9 and evaluate the method on six MD molecules, SCF initialization under different functionals, and dipole and quadrupole prediction.

Critical strengths

The experimental design helps disentangle the effects of noise initialization and rollout training. Table 4 crosses deterministic versus noisy inputs with single-step versus rollout losses while keeping the backbone, basis, parameter count, and optimization-step budget fixed. In particular, the deterministic baseline includes two network evaluations and step conditioning, helping distinguish the gains from simply using more inference computation. On the 10% QM9 subset, deterministic rollout achieves an NMAE of 0.268%, while combining noise and rollout training reaches 0.240%; adding noise alone does not improve performance. These results support using the two components together. It is also useful that the curriculum continues training the same model without a separate teacher and that GPU-hours are reported for several training configurations.

Critical weaknesses

Missing comparisons with major density prediction models make it difficult to assess the relative gains in accuracy and SCF acceleration.

BOA is neither cited nor compared, despite reporting lower errors on the same QM9 VASP density benchmark. Using the same split as Fu et al., BOA reports NMAEs of 0.1381% and 0.1339% for its small and large configurations, respectively, both below OrbFlow’s 0.152%. These results do not match model size or training and inference costs, so the numbers alone do not establish relative computational efficiency. However, a broad claim of state-of-the-art accuracy is difficult to support without including BOA. BOA should be included as a major baseline, with accuracy and computational cost reported together. Since BOA uses a quadratic basis-product representation, improvements over SCDP, which shares the same linear GTO basis, should be distinguished from superiority over density prediction methods more broadly. (BOA, Klockow et al., ICLR 2026)

Similarly, comparing only against SAD in Table 3 does not establish an advantage over existing learned initializers. Reporting total time including model inference is useful, but results using densities predicted by SCDP, ELECTRA, and BOA under the same solver settings are needed. In particular, comparing whether other models trained on PBE also improve initialization for B3LYP5 and PBE0 would help determine whether the observed functional transfer and time savings provide an additional benefit specific to OrbFlow’s training method. The comparison should include convergence rates and total time for model inference, density reconstruction, and solver execution, alongside SCF iteration counts.

The benefits of the GTO coefficient representation and supervision are not sufficiently separated from those of the proposed training procedure.

OrbFlow constructs its training paths using coefficients obtained by fitting DFT densities to an auxiliary GTO basis, and trains by comparing the reconstructed predictions with the original DFT densities. If an important motivation is to extend equivariant learning methods used in Hamiltonian prediction, such as QHFlow (Kim et al., 2025), to this coefficient space, a natural comparison would apply existing models to the same GTO representation and supervision. Flow-matching methods such as QHFlow and QHFlow2 (Kim et al., 2026), as well as regression models using backbones such as QHNetV2 (Yu et al., 2025), could provide useful baselines. Such comparisons would help distinguish the benefits of the GTO representation and supervision from those of the proposed flow and rollout training.

Questions for authors

Accuracy and cost comparison with BOA. Could you include BOA in Table 1 by training and evaluating it on the same QM9 VASP data and test split? Please report parameter counts, training costs, and total density prediction times measured under the same hardware and evaluation conditions, and clarify the scope of the state-of-the-art claim in light of BOA’s published results.

Existing models using the same GTO representation. How do OrbFlow’s accuracy and computational cost compare with QHFlow (Kim et al., 2025) or QHFlow2 (Kim et al., 2026)-style flow matching and deterministic regression models using the QHNetV2 (Yu et al., 2025) backbone when trained on the same GTO coefficient prediction task?

Training curves and the additional benefit of Phase 2. Could you show the training losses for Phase 1 and Phase 2, along with validation density NMAE measured under identical inference settings, as functions of optimization steps and cumulative GPU-hours?

Learned initialization baselines in Table 3. When using densities predicted by SCDP, ELECTRA, and BOA trained on the same PBE data, how do convergence rates, SCF iteration counts, and total calculation times compare with OrbFlow for PBE, B3LYP5, and PBE0? Please match electron-count normalization, grids, solver settings, and convergence criteria within each functional, and report model inference, density reconstruction, and solver times separately. This would help assess whether the functional transfer and time savings provide additional benefits over other density predictors.

6 comments
AuthorsOpenReview ↗

# Response to Reviewer cVZC

We thank the reviewer for the careful review and for noting the design of Table 4. We address all four requests below.

Could you include BOA in Table 1 by training and evaluating it on the same QM9 VASP data and test split? Please report parameter counts, training costs, and total density prediction times measured under the same hardware and evaluation conditions, and clarify the scope of the state-of-the-art claim.

BOA is now in Tables 1 and 2, and the state-of-the-art claim is narrowed. OrbFlow is now claimed to be the most accurate model among linear expansions in an atom-centered GTO basis (14.6% below SCDP+VO on full QM9). BOA's released QM9 checkpoints were trained on the same split, so we evaluated them and our models on the same 10,000 test molecules and grid, on one H200 with batch size 1. We also include direct regression with OrbFlow's own backbone, which separates the cost of the backbone from the cost of our training.

ModelNMAE [%]Split 2 [%]ParamsPrediction [s/mol]Training [GPU-h]
SCDP+VO, L=60.1780.19161M0.23≈990
EquiformerV3, direct regression0.1740.18899Mn/a≈620
OrbFlow0.1520.16599M0.33≈1,600
BOA small0.1370.144102M0.64≈49 (est.)*
BOA large0.1320.138102M0.70≈91 (est.)*

Prediction time is network inference plus density evaluation on the full grid, averaged over 200 test molecules after warm-up. BOA also needs 0.09 s of CPU preprocessing per molecule. Training cost is logged wall-clock time on H200 GPUs, validation included.

*Estimated. BOA does not report its training time, and we did not retrain BOA but evaluated its released checkpoints. We ran its released training code on one H200 with the settings stored in each checkpoint, timed 400 steps after warm-up, and multiplied the measured time per step by the checkpoints' step counts (validation not included).

Accuracy and prediction time. OrbFlow predicts twice as fast as BOA, and no model in the table is both faster and more accurate than OrbFlow. The difference is in building the density on the grid (0.61 s against 0.17 s), because BOA's quadratic form is evaluated from products of pairs of basis functions. The same representation explains much of BOA's accuracy, since on four MD molecules BOA is below the fitting floor of our basis (Appendix A.4). Our contribution is the training, which needs only a network that predicts coefficients, so it can also be applied to BOA.

Training cost. Training is OrbFlow's main cost, and BOA is both more accurate and, by our estimate, cheaper to train. BOA's training and prediction costs differ because they use very different numbers of grid points. Its loss uses 5,000 to 6,000 sampled points per molecule (ours uses 60,000), while prediction evaluates the full grid of about 600,000 points. Its quadratic form makes each point costly, which dominates prediction time but adds little to training. Three points put the number in context. (i) Against the previous best linear-expansion model the increase is moderate. SCDP+VO, logged the same way on the same GPUs, takes about 990 GPU-h, so OrbFlow costs 1.6× as much for a 14.6% lower error. (ii) The increase comes from our training, not from the backbone. The same EquiformerV3 trained by direct regression costs about 620 GPU-h, less than SCDP+VO, and our training adds about 2.6× on top (74 against 28 GPU-h on the 10% subset), mostly from Phase 2, which runs the network twice and backpropagates through the rollout. This compute buys accuracy that regression does not reach by training longer. Regression's validation error is flat at the end of training, and on full QM9 regression trained for 500k steps, more than OrbFlow's 450k, stops at 0.174%. On the subset, OrbFlow already passes regression's final accuracy after 35 to 46 GPU-h. (iii) Training is paid once, and we release the trained checkpoints, while prediction time is paid for every molecule. We did not tune our training cost, for example by using fewer grid points in the loss. Section 4.5 and Appendix D of the revision report these numbers.

AuthorsOpenReview ↗

Continued Response to Reviewer cVZC

How do OrbFlow's accuracy and computational cost compare with QHFlow or QHFlow2-style flow matching and deterministic regression models using the QHNetV2 backbone when trained on the same GTO coefficient prediction task?

We ran both comparisons on our GTO coefficient task and on the 10% QM9 subset of our ablations. We used the network of QHFlow2 (Kim et al., 2026), the most recent of these methods, and replaced its Hamiltonian heads, which output matrix blocks for pairs of atoms, with our per-atom readout of GTO density coefficients (33.3M parameters).

With the QHFlow2 recipe, we follow its released settings. Each atom gets a random time, the network predicts the final coefficients from the noisy input (endpoint prediction, Gaussian prior with σ = 0.1), and inference uses 3 Euler steps. Since our basis has no initial-guess coefficients, we predict the full target instead of QHFlow2's residual. With our training, Phase 1 uses the same kind of endpoint prediction, and Phase 2 trains on the model's own 2-step rollout with the loss on the final density, and inference uses the same 2 steps. For deterministic regression we used QHNet (Yu et al., 2023), the public predecessor of QHNetV2, with the same readout, loss, subset and budget. Times are on one A100 (batch size 1, full grid).

Model, 10% subsetNetworkTest NMAE [%]ParamsTraining cost [GPU-h]Prediction time [s/mol]
QHFlow2 recipeQHFlow20.25533.3M28.60.32
OrbFlow training (ours)QHFlow20.24033.3M45.30.27

On the same network, our training lowers the error of the QHFlow2 recipe by 6% (0.255% to 0.240%) and predicts faster, since it needs two network passes instead of three. Its extra training cost comes from backpropagating through the rollout in Phase 2. It also matches our EquiformerV3 model on this subset (0.240%) with a third of the parameters.

The QHNet regression (21.9M parameters) is not finished yet. Our first run had a bug in our adaptation of its output layer, and the corrected run did not finish in time for this reply. We will post its result in this thread as soon as it is available.

Could you show the training losses for Phase 1 and Phase 2, along with validation density NMAE measured under identical inference settings, as functions of optimization steps and cumulative GPU-hours?

Yes. A new figure in Appendix E shows OrbFlow's training loss and validation density NMAE through Phase 1 and Phase 2 against optimization steps, always under the same inference setting (K=2) (10% subset, three seeds). We give the corresponding cumulative GPU-hours in the text below, since a Phase 2 step costs more than a Phase 1 step. To isolate the effect of Phase 2, the figure also shows the same Phase 1 checkpoint trained further with the Phase 1 loss instead of switching, and direct regression with the same backbone.

Validation NMAE drops as soon as Phase 2 starts, from 0.54% at the end of Phase 1 to 0.44% one epoch later (means over three seeds), and reaches 0.221% to 0.227% on the 50 validation molecules. Continuing Phase 1 instead ends at 0.267% to 0.294%, and direct regression levels off at 0.263% to 0.267% after its 28 GPU-hours. OrbFlow passes this level after 35 to 46 GPU-hours and ends 15% lower after 74 to 83. We also removed the "2.5× less compute" statement from the abstract. Section 4.6 now states it only against joint optimization (74 against 182 GPU-hours).

AuthorsOpenReview ↗

Continued Response to Reviewer cVZC

When using densities predicted by SCDP, ELECTRA, and BOA trained on the same PBE data, how do convergence rates, SCF iteration counts, and total calculation times compare with OrbFlow for PBE, B3LYP5, and PBE0?

We ran this comparison with SCDP+VO and BOA. All models are trained on the same PBE data and tested on the same molecules (200 for PBE, 100 for B3LYP5 and PBE0). Densities are scaled to the number of valence electrons, and every predicted density uses the same PAW occupancies from the converged calculation, as in Table 3, because VASP does not use a density without them (see our reply to qf6S). Within each functional, all runs use the same grid, solver settings and convergence threshold (10⁻⁵ eV). We report the SCF iteration count, which measures the quality of the initial guess independently of hardware (RIC is relative to SAD), and the prediction time per molecule on one H200 (batch size 1), split into network inference and density reconstruction on the full grid.

PBE, 200 moleculesConverged [%]SCF iterationsRIC [%]Inference [s]Reconstruction [s]
SAD (VASP default)10020.9 ± 1.1100n/an/a
SCDP+VO10011.6 ± 2.155.40.060.17
BOA (large)10014.5 ± 1.169.40.09 (+0.09 CPU)0.61
OrbFlow10011.5 ± 1.754.80.160.17

At PBE, OrbFlow and SCDP+VO need about the same number of iterations, and BOA needs more, although its density error is the lowest. A lower density error on the grid therefore does not by itself mean fewer SCF iterations. The SAD and OrbFlow values agree with Table 3, which used 500 molecules (21.3 and 11.6 iterations, RIC 54.5%). The final version will add the solver (VASP) times of all four initial guesses, measured in a single matched batch, and the B3LYP5 and PBE0 runs.

Reviewer_qf6SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed a new flow matching approach to predict the electron density (represented by basis function coefficients). Results are evaluated by density error, SCF initialization, and downstream properties.

Critical strengths

Combining with flow matching, this paper demonstrates near SOTA-level accuracy in the electron density prediction task. In addition, this paper also introduces two-phase curriculum learning, further reducing the training time. The predicted electron density reduced the SCF cycles needed for the DFT task and displayed zero-shot transferability.

Critical weaknesses

I have some minor concerns about this work:

  1. It seems that another approach offers better performance on the same dataset (ICLR 2026, https://openreview.net/forum?id=HDdkFjFEZd). They obtained 0.1339 on the QM9 dataset, compared with 0.152 NMAE in this work. Could the authors comment on the performance gap between this work and the existing work?
  2. As mentioned in the introduction section (Liu et al., 2026), there are two advantages to using electron density over the Hamiltonian matrix: electron density can generalize to larger molecules, and it has good zero-shot transferability. This work demonstrates the latter, but omits the first. Could the author comment on the performance on an OOD dataset?
  3. Could the author comment on the choice of VASP? What's the applicability of OrbFlow to be used with all-electron molecular packages, such as PySCF. For a complete VASP/PAW charge-density initialization, the real-space grid (predicted by OrbFlow ) alone is not the whole input; PAW one-center augmentation occupancies are also needed (See https://arxiv.org/abs/2609.21759). Could the author clarify how the augmentation occupancy was predicted? Is it based on the ground truth, or the VASP default? If the answer is the former, then how does OrbFlow perform for molecules without ground truth?
Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer qf6S

We thank the reviewer for the helpful questions. They led us to add BOA as a baseline, an OOD test on larger molecules, and a clarification of how our SCF tests use the PAW occupancies.

They obtained 0.1339 on the QM9 dataset, compared with 0.152 NMAE in this work. Could the authors comment on the performance gap between this work and the existing work?

The gap comes mainly from the representation. BOA writes the density as a quadratic form in the basis functions, while OrbFlow, like SCDP, predicts a linear expansion in an atom-centered basis. On four of the six MD molecules, BOA is below the fitting floor of our basis, i.e. the lowest error any linear model in this basis can reach (Appendix A.4).

Molecule (NMAE [%])Our basis floorOrbFlowBOA (small)
Ethanol0.8501.0780.710
Benzene0.3990.4210.361
Resorcinol0.5150.6630.371
Ethane0.8070.8710.772

We now cite BOA (Klockow et al., ICLR 2026), include it in Tables 1 and 2, and restrict our claims, including in the abstract, to linear expansions in an atom-centered basis. Among these, OrbFlow improves on the strongest previous model, SCDP+VO, by 14.6% on full QM9.

The quadratic representation also makes BOA slower. On the same test set, grid and GPU (one H200, batch size 1), the released BOA checkpoints reach 0.132% (large) and 0.137% (small) at 0.70 s and 0.64 s per molecule, and OrbFlow reaches 0.152% at 0.33 s, about twice as fast, with a similar number of parameters (99M against 102M). Building the density on the grid dominates the difference (0.61 s for BOA against 0.17 s), because a quadratic form is evaluated from products of pairs of basis functions, while a linear expansion needs each basis function only once.

The two approaches are complementary. Our contribution is a training method rather than a representation. The model starts from noise and is trained on its own few-step rollout. It needs only a network that predicts coefficients and a differentiable density loss, so it applies in principle to BOA's quadratic coefficients. With the backbone fixed, it lowers the error of direct regression from 0.174% to 0.152% on full QM9, and it improves QHFlow2's network beyond QHFlow2's own training recipe (0.240% against 0.255% on our 10% subset, see our reply to Reviewer cVZC). Applying it to BOA is a natural next step, which the revision discusses.

Could the author comment on the performance on an OOD dataset?

We added an OOD test on larger molecules. From QM40 (Madushanka et al., 2024) we selected 50 molecules containing only H, C, N, O and F with 10 to 26 heavy atoms (QM9 has at most 9), in three size bins (10 to 13, 16 to 19 and 23 to 26 heavy atoms, with the number limited by the cost of the reference calculations), and ran VASP PBE with the QM9 settings. All models are trained on QM9 only.

The test exposed a size dependence in the previous OrbFlow model. It inherits from SCDP a normalization that divides the predicted coefficients by √N_orb of the whole molecule, a global scale that a local network cannot infer for molecules larger than those it was trained on. Its QM40 densities contain 12%, 22% and 33% too few electrons in the three bins. Replacing this normalization with a constant and briefly fine-tuning on QM9 removes the dependence. The error in the electron count then falls below 0.5% in all three bins, while the QM9 test error changes only from 0.154% to 0.163% (same 1,000 test molecules). Appendix G of the revision describes this change.

With this model, OrbFlow's density still accelerates the SCF on the larger molecules.

Initial guess (QM40, 50 molecules, PBE)SCF iterationsRIC [%]Time [s]RTR [%]
VASP default25.810012.2100
OrbFlow22.085.411.090.5

RIC and RTR are the mean SCF iteration count and wall-clock time relative to the default initial guess (lower is better). The reduction is smaller than on QM9 (RIC 54.5% for PBE), so the benefit decreases for molecules well beyond the training sizes, but it remains.

AuthorsOpenReview ↗

Continued Response to Reviewer qf6S

Could the author comment on the choice of VASP? What's the applicability of OrbFlow to be used with all-electron molecular packages, such as PySCF.

We used VASP because the QM9 reference densities (Jørgensen & Bhowmik, 2022) are VASP PAW densities, i.e. valence pseudo-densities without core electrons, smoothed near the nuclei. Running the SCF tests in the same code keeps the training labels and the solver consistent. OrbFlow itself is not tied to VASP. In PySCF, the predicted density can initialize an SCF by building a Fock matrix from it (Hartree and XC potentials on the integration grid) and diagonalizing it. Because PySCF is all-electron by default, the core electrons must be added, either as frozen atomic core densities or by running PySCF with pseudopotentials. Alternatively, OrbFlow can be retrained on all-electron densities, such as the PySCF QM9 set of Li et al. (2025), with a GTO basis that includes steep core functions. The final version will add this discussion. A quantitative evaluation in all-electron codes is future work.

Could the author clarify how the augmentation occupancy was predicted? Is it based on the ground truth, or the VASP default? If the answer is the former, then how does OrbFlow perform for molecules without ground truth?

The occupancies were taken from the ground truth, and we thank the reviewer for raising this and for the reference. OrbFlow predicts only the valence density, not the PAW augmentation occupancies. The revised paper will state this.

For a molecule without ground truth, VASP cannot use a predicted density on its own. Given a charge-density file without the occupancies, VASP reports the file as incomplete and starts from superposed atomic densities instead, so the predicted density is discarded. We confirmed this in our runs. Densities from two different models (SCDP+VO and BOA), written without occupancies, gave exactly the same mean SCF iteration count. This holds for every model that predicts only the density, including SCDP, ELECTRA and BOA, and we believe it deserves explicit mention in work on learned SCF initialization. The revision will state it in Section 4.3 and Appendix F.

The occupancies therefore have to come from a separate model. As Ærtebjerg et al. (2026, arXiv 2609.21759) show, they matter for SCF initialization, and their AugNet predicts them. AugNet is complementary to OrbFlow, since its occupancies combined with OrbFlow's density would give an initial guess that needs no ground truth, which we list as future work. Until then, all our SCF comparisons, including the comparison with SCDP+VO and BOA requested by Reviewer cVZC and the QM40 runs above, give every initial guess the same converged occupancies, so that they compare the densities on equal terms.

Our main contribution is the training method, which lowers the density error at a fixed backbone (0.174% to 0.152% on full QM9). SCF initialization is a downstream use of the density.

Thanks for answering my questions. Most of my concerns have been addressed, and I have increased my score to 3, weak acceptance.

I still have two remaining comments/concerns:


It's completely reasonable that one representation is advantageous over another representation. So when comparing OrbFlow against BOA, the author could show that the linear expansion is harder than the quadratic expansion, and that the linear expansion is more useful in some areas.

For the former, BOA's Appendix C3, Table 5, indeed suggested that linear expansion could be harder. Perhaps the author can find any other supporting evidence of the difficulties. For the latter, the revised manuscript compares the inference time, which partially addresses this problem. I am also wondering: in addition to the inference time, could the linear expansion be more advantageous from another perspective (such as VRAM or something else)?


By changing the normalization method from 1/N1/\sqrt N to a constant, the electron count error is reduced a lot. But the degradation of RIC is still concerning (54.5% -> 85.4%). Is it because we have a much larger density NMAE (not shown in the revised manuscript)? Could the author further comment on this issue?

AuthorsOpenReview ↗

Second response to Reviewer qf6S

We sincerely thank the reviewer for the follow-up and for raising the score. We address the two remaining points below and will add the new results to the paper.

The author could show that the linear expansion is harder than the quadratic expansion, and that the linear expansion is more useful in some areas. [...] In addition to the inference time, could the linear expansion be more advantageous from another perspective (such as VRAM or something else)?

Why the linear expansion is harder. Besides BOA's Table 5, we see two further signs.

  1. Expressiveness. The density is a sum of products of orbitals on pairs of atoms. A one-center linear expansion has to re-expand these products, which is why density-fitting bases are three to five times larger than orbital bases (Eichkorn et al., 1995) and why SCDP needed bond-midpoint virtual orbitals. The quadratic form represents the products directly. This matches our first reply, where BOA is below the fitting floor of our linear basis on four MD molecules.
  2. Positivity. A linear expansion has no built-in positivity constraint, since its l > 0 functions take both signs, so a non-negative density has to be learned from data. A quadratic form is better suited by construction, because a density written as products of basis functions, like the density-matrix form, is non-negative whenever its coefficient matrix is positive semidefinite. In practice, on 200 QM9 molecules, OrbFlow is negative on 1.8% of the grid points (less than 0.0002% of the charge), while BOA is negative on below 0.01% of the points in only 13 of 200 molecules.

Where the linear expansion is more useful. Besides the twofold faster prediction, it needs much less GPU memory. We measured peak memory on the same 200 QM9 molecules (one H200 with 141 GB, batch size 1, full grid, weights included) at equal chunk sizes of grid points. Values are means, with the largest molecule in parentheses.

ModelNetwork [GB]Density, 20k chunks [GB]Density, 100k chunks [GB]Full grid in one pass [GB]
OrbFlow1.183.5 (4.5)13.7 (18.9)64.8 (100.9)
BOA (small or large)0.487.5 (13.8)33.3 (61.1)out of memory

Evaluating products of pairs of basis functions makes BOA need 2.2 to 2.4 times the memory of OrbFlow, and about 3 times on the largest molecule. In addition, the electron count and the multipoles are linear in the coefficients. Liu et al. (2025, arXiv 2509.25724) also find that linear density coefficients transfer to molecules of 100 to 200 atoms (RIC 73% to 82%), while a density matrix, which is quadratic like BOA's form, degrades from 70.5% to 91.7% RIC on molecules three times larger.

But the degradation of RIC is still concerning (54.5% -> 85.4%). Is it because we have a much larger density NMAE? Could the author further comment on this issue?

We agree, and see two reasons.

The main reason is the short fine-tune. Due to the limited time, we did not retrain with the new normalization but fine-tuned the submitted checkpoint for only 7.5k steps, against 450k steps for full training. Since the new normalization rescales every output coefficient, this is not enough to relearn the output scale, as the higher QM9 error also shows (0.154% to 0.163%). We will retrain with the constant normalization from the start and report the QM40 RIC and the QM40 density NMAE per size bin in the final version.

Size extrapolation is also harder in general. The QM40 molecules have up to 2.9 times as many heavy atoms as QM9 and are more drug-like, and their default SCF already needs more iterations (25.8 against 21.3). Liu et al. (2025) observe the same trend, with RIC rising from 68.9% to 81.3% for their best linear model on molecules three times larger than in training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.