acceptodds
Withdrawn from ICLR 2027

Accelerated Neural Riemannian Flows for Kohn-Sham Density Functional Theory

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Density functional theory (DFT) underlies many electronic structure calculations. Typically, a key step in this framework is the iterative self-consistent field (SCF) procedure, which remains computationally expensive. Recent learning-based approaches accelerate parts of this pipeline by improving Hamiltonian prediction or initialization, but still retain the bottleneck at its root. We propose a gauge-symmetric neural Riemannian flows framework for SCF-free electronic structure computation. Our approach directly propagates electronic states on the Grassmann manifold via gauge-equivariant dynamics, rather than relying on a conventional SCF solver. This construction removes the structural bottleneck imposed by the SCF solver and makes the entire propagation trajectory differentiable, suitable for end-to-end learning of the electronic solver. Furthermore, we recast the propagation process at the operator level into a prefix-scan compatible form for efficient parallel execution and acceleration on modern GPUs. Empirically, our method reaches the electronic ground state without SCF, consistently improves energy and orbital related metrics, and achieves speedups of up to 30 times over prior approaches on MD17 molecules. These results point to a new SCF-free direction for electronic structure computation and highlight the potential of learned electronic solvers that are both accelerator-friendly and aligned with end-to-end differentiable learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
riemannian optimization, grassmann manifold, electronic struture prediction, gauge symmetry

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LWsnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces riemannian flow matching for the evolution of molecular orbital CC for the Grassmannian manifold Gr=St/U(k)\text{Gr} = \text{St}/\text{U}(k). The authors provide the infinitesimal derivative of the quotient manifold (i.e., horizontal derivative) and the corresponding flow matching training loss. The results show the effectiveness of the method.

Critical strengths
  1. Modeling the orbital iteration as a quotient-space flow matching is clearly novel.
  2. The results for MD17 and MD22 show the effectiveness of this method.
Critical weaknesses
  1. Stiefel manifold is usually denoted as VkV_k instead of St\text{St}.
  2. Equation (2) P=CC†SP=CC^\dagger S is not density matrix. The density matrix is generally defined by CC†CC^\dagger. I also did not find the exact equation in [1] that you referred to. This is also used in your loss definition
  3. Flow matching itself is defined in the Euclidean space. Please cite Riemannian flow matching [2], quotient space diffusion [3] and related works to make equation (8) and your Appendix more meaningful.
  4. For your inference geodesic walk, I cannot understand from your writing line 270-305 that you are walking from the initial point t=0 directly to all other time points like what QHFlow did, or you are walking under successive time frames. Your line 296 shows you walk successively, while your line 300 and Figure 2 suggest the other. I also do not recommend introducing jargon like "quantum prefix-scan" as the word "quantum" does not make sense here.
  5. In Figure 1, what are the black curves and red curves crossing the horizontal space? Similarly, the black curve in Figure 2.
  6. Please cite QHNetv2 [4] and its results.

[1] Eric Cancès and Gero Friesecke. Density Functional Theory. Springer, 2023.

[2] Chen, Ricky TQ, and Yaron Lipman. "Flow matching on general geometries." International Conference on Learning Representations. Vol. 2024. 2024.

[3] Xu, Yixian, et al. "Quotient-space diffusion models." arXiv preprint arXiv:2604.21809 (2026).

[4] Yu, Haiyang, et al. "Efficient prediction of so (3)-equivariant hamiltonian matrices via so (2) local frames." arXiv preprint arXiv:2506.09398 (2025).

Questions for authors

See weaknesses.

1 comment

We appreciate your careful reading of our work and the constructive suggestions you have offered. We address the points raised in the following response and invite you to share any additional questions or comments as the discussion proceeds.


W1. Stiefel manifold is usually denoted as V_k instead of St.

The notation St\mathrm{St} is widely adopted in the Kohn–Sham DFT literature[1], and we therefore consider its use appropriate in our formulation.

W2. Equation (2) P=CC†SP = CC^{\dagger} S is not density matrix...

We thank the reviewer for the comment. The conventional AO density matrix is expressed as CC†CC^\dagger, while CC†SCC^\dagger S is also a representation of the density projector in a nonorthogonal basis. Our formulation adopts the latter representation to describe electronic states on the Grassmann manifold, and we will clarify this distanction.

W3. Flow matching itself is defined in the Euclidean space...

Our formulation of Eq. (8) follows directly from the standard Riemannian geometric construction of a pullback metric, as explicitly derived in Appendix B.1. The suggested works are likewise built upon these established geometric foundations, but pursue fundamentally different objectives. Their connection to our approach is therefore limited to sharing certain mathematical concepts, rather than directly enriching the formulation of Eq. (8) or its derivation in the Appendix.

W4. For your inference geodesic walk...

Our method performs successive propagation mathematically, but not sequentially in computation. Specifically, the model predicts the sequence of local propagation operators independently of the intermediate states reconstructed by the scan. Their cumulative compositions are then evaluated in parallel via an associative prefix-scan, yielding all intermediate electronic states from the initial state. Thus, Equation in line 296 describes the mathematical composition of successive updates, whereas Figure 2 illustrates their parallel evaluation.

We understand that the term "quantum prefix-scan" may be ambiguous in this context. To avoid potential confusion and better reflect the underlying geometric structure, We will revise the terminology accordingly.

W5. In Figure 1, what are the black curves...

In Figure 1, the black curves represent gauge fibers corresponding to redundant orbital representations that are identified under the quotient projection onto the Grassmann manifold. The red curves illustrate horizontal directions representing physically meaningful subspace evolution. In Figure 2, the black curve depicts the propagation trajectory of the orbital coefficient matrices predicted by our model. We will clarify these visual elements in the figure captions.

W6. Please cite QHNetv2 and its results...

We intended to include QHNetv2 as an additional baseline. However, upon examining the repository provided in the QHNetv2 paper, we found that its implementation code is still not publicly available, making it difficult to reliably reproduce and evaluate the method. Nevertheless, we will cite QHNetv2 in the related work.


Reference

[1] Altmann, R., Peterseim, D., & Stykel, T. (2024). Riemannian Newton methods for energy minimization problems of Kohn–Sham type. Journal of Scientific Computing, 101(1), 6.

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

The paper proposes an ML surrogate using Riemannian Flows on the Grassmann manifold to circumvent the computational cost of iterative KS-DFT solvers. It highlights a Theoretical Framework for KS-DFT (1.) and an (associative) scan-compatible operator framework (2.) as its core contributions.

  1. rests on a local-convexity theorem
  2. on a "gauge-symmetric quantum bundle action" which exploits the associativity of orbital rotation matrices.

Based on these they train a sequence-to-sequence network in the form of a 1-D CNN which effectively only takes the intial guess molecular orbital coefficients (C_0) but no These contributions are then tested on two MD simulation trajectories from MD17 and MD22.

Critical strengths

None

Critical weaknesses

W1: Missrepresentation of Prior Work

The related work states that the cited Hamiltonian models "are typically designed and used as initializations or auxiliary components for SCF calculations" (p. 6, ll. 318–320), and the abstract presents SCF-free computation as the novelty of this submission (p. 1).

  1. QHFlow aims to "bypass" SCF "by directly predicting the Hamiltonian" (Kim et al., 2025, abstract).
  2. WANet aims to determine "the Hamiltonian without relying on SCF iterations" (Li et al., 2025).
  3. PhiSNet derives energies and forces directly from its predictions (Unke et al., 2021, abstract).
  4. DEQHNet describes itself as "a Hamiltonian solver" (Wang et al., 2024.

Several of these works also report SCF initialization as an additional application not their sole purpose. Moreover, even in these SCF-Free frameworks, the "structural bottleneck" is not only "imposed by the SCF-solver" as the abstract claims, since a significant chunk of applications requires energies and hence a Fock build, which dominates the cost of each individual solver step.

W2: Insufficient & Methodologically Flawed Evaluations

Page 13 in the appendix states that while their method runs on GPU they compare the end-to-end speedups to baselines run on CPU

For fair comparison, all [baseline] methods are evaluated on the same set of fixed molecular geometries under a common PySCF-based electronic structure pipeline with the same settings used for data generation, which includes the basis set, exchange-correlation functional, and grid level.

W3: Misscallaneuous:

  • Missing attributions (see Ethics Concerns)
  • Writing and representation is lacking:
    • "time" is used conflated across actual MD simulation time and "pseudo / solver" time along the density optimization.
    • How are U_0 and U_ctx,U_ctx,k defined / obtained? (Appendix B4), how are you ensuring that U0U_0 is on the Grassmannian (if it is based on MINAO it is not), or are you starting from the first pure coefficients from pyscf, in that case you already spend one SCF step to obtain U0U_0
    • what do the dots and dashed lines mean in figure 4? are they network inference and SCF steps if yes why are the lines between the dots not straight?
  • No supplementary material or code provided.
    • This is especially problematic at it is very unclear, how some of the baselines are performing unexpectedly bad here.
    • Based on the numbers reported you are claiming an error of at least 65 Hatree (= 988 (yours) - 923 (QHFlow)) AFTER 13 SCF steps?! At zero steps the reported error of QHFlow is over 90 Hatree. This is orders of magnitude larger than errors reported in literature.
Questions for authors

Q1: Unclear In-/Outputs and Datasplit

What are the model inputs and outputs at evaluation time vs at training time? What is the train / test data split for each of the reported experiments (MD17, MD22, Twuseted Ethylene)?

Q2: Speedup calculations in Tab: 2

How have the speedups been calculated in this table? Since it reports energy, what is the end to end speedup to obtain energy? (It can't be 50x right? It should be asymptotically bounded at 1/131/13 since you also require at least one Fock build)

Q3: Benchmark Integrity

How do you explain the several orders-of-magnitude higher energy errors of the baselines compared to prior work?

3 comments

Thank you for your detailed review and for setting out your concerns. We explain our position in the reply that follows, addressing each point in turn. Please let us know if any concerns remain after considering our responses.


W1/Ethics 4. Misrepresentation of Prior Work...

The cited works share a common motivation of avoiding or reducing the computational burden of conventional SCF calculations. However, their ability and motivation does not necessarily mean that they have successfully replaced SCF in recovering accurate electronic ground states. As demonstrated in our experiments, their predictions still exhibit substantial accuracy gaps that are reduced through subsequent SCF refinement. Under such accuracy requirements, these predictions effectively serve as initializations or auxiliary inputs to conventional SCF calculations, rather than fully replacing the solver. Our discussion reflects this practical limitation, not a denial of their original objectives. We seek to overcome this limitation from a different perspective by actively leveraging the intrinsic geometry of electronic states as a fundamental inductive bias.

Regarding the computational bottleneck, we recognize that Fock construction can be expensive and remains relevant to electronic energy evaluation. However, the review appears to overlook an important distinction between the cost of individual electronic structure operations and the iterative computational bottleneck of SCF. Certain computational costs are unavoidable in electronic structure calculations, even when SCF iterations are eliminated. These costs are also present in existing approaches. For instance, QHFlow performs Fock construction during preprocessing to obtain its initial Hamiltonian. Indeed, the substantial cost of these underlying operations is precisely why reducing or eliminating repeated SCF iterations has been a central motivation of this line of research. The structural bottleneck arises from repeatedly constructing density-dependent operators, solving eigenvalue problems, and updating electronic states until convergence. The presence of these unavoidable computational costs therefore does not diminish the significance of eliminating the iterative SCF procedure itself.

W2. Insufficient & Methodologically Flawed Evaluations...

All baseline models were evaluated on GPU under the same hardware environment as our method. The subsequent PySCF calculations, including SCF refinement and electronic structure evaluation, were performed on CPU using identical computational settings across all methods. Thus, the reported inference times reflect a consistent evaluation pipeline, with GPU execution for neural network inference and CPU execution for PySCF-based computations.

W3/Q3. Writing and representation is lacking...How do you explain the several orders-of-magnitude...

In our experimental setting, all methods were trained and evaluated using randomly generated matrix and their corresponding reference electronic states, rather than relying on conventional MINAO initialization. Specifically, U0U_0 is constructed from the matrix, with UctxU_{\mathrm{ctx}} also taken from U0U_0, satisfying the required orthonormality constraints without any preceding SCF iterations. This introduces a more challenging and general learning and evaluation setting than those considered in the original baseline implementations. The larger energy errors should therefore be understood in the context of recovering accurate electronic states from such challenging initializations.

Regarding the notation, the time variable in our formulation refers exclusively to the pseudo-time of electronic-state propagation, rather than physical molecular dynamics time. In Figure 4, the markers indicate the initial, final, and predicted states, along with intermediate states obtained during SCF iterations. The dashed lines connect the initial state to the model-predicted state. We will make these point explicitly in the manuscript.

Q1. What are the model inputs and outputs...

During both training and evaluation, the model takes randomly generated matrix from the Stiefel manifold and predicts coefficient matrices that parameterize the horizontal dynamics. For MD17 and MD22, the full molecular configurations are divided into training, validation, and test sets with a split of 8/1/1. For twisted ethylene, we similarly follow the same splitting protocol using configurations generated from multiple molecular dynamics ensembles.

Q2. How have the speedups been calculated...

The reported speedups are calculated based on the measured inference time required to obtain the final occupied-orbital coefficient matrix. Specifically, the reported 49.88× speedup in Table 2 corresponds to Ac-Ala3-NHMe, calculated from the baseline inference time of 79.8 s and our inference time of 1.60 s. When including the subsequent energy evaluation, an additional 19.33 s is required, resulting in a total computation time of 20.93 s. Similarly, for salicylic acid, energy evaluation requires an additional 1.65 s, resulting in a total computation time of 2.14 s, including our model inference time of 0.49 s. Even without adding the corresponding energy evaluation costs to the baseline runtimes, these yield conservative speedups of 3.81× and 6.82×, respectively. These results still demonstrate substantial computational acceleration.

A more comprehensive accounting of computational costs would also need to consider the additional electronic structure calculations involved in prior methods, such as QHFlow. Our reported inference times exclude such preprocessing costs and focus on model inference and SCF refinement where applicable. Including additional energy evaluation costs for our method without equally considering the computational requirements of prior approaches would introduce an imbalance in the comparison and unfairly disadvantage our method.

Ethics. Missing attributions...

The references raised in the review relate to our work at different levels. Grassmann gradients, horizontal lifts, and orthogonality-preserving transformations are established mathematical foundations, already acknowledged in our manuscript through Edelman et al., Absil et al., and VandeVondele and Hutter. The analyses of Dai et al. and Cancès et al. concern related theoretical properties of Kohn–Sham gradient flows and Hessians. Likewise, the literature on neural manifold flows and Riemannian flow matching shares geometric concepts with our approach, while pursuing different learning objectives and computational settings.

These connections naturally arise from the shared mathematical foundations underlying these lines of research. Riemannian geometry provides a common framework for related constructions and analyses across different problems and applications. In this context, our manuscript acknowledges and discusses the relevant foundational literature underlying these shared mathematical concepts. We did not intend to overlook or diminish the contributions of the studies you mentioned, and we recognize that a more explicit discussion of these connections could further improve the presentation of our manuscript.

Regarding the ethics-related concern raised in the review, we would appreciate further clarification on the basis of this assessment. While the review provides a list of related references, it remains unclear what criteria or considerations led to this concern. Could you elaborate on the standards applied and explain how they relate to the citation in our manuscript? Such clarification would help us better understand the concern and address it more constructively.

Thank you for the responses. None of my concerns regarding the soundness, baseline evaluation (see the PySCF sanity-check table in my top-level comment), or attribution of prior work have been adequately addressed, and the practical utility of the proposed approach is not apparent. I maintain my score of 1.

  • Ill-posed inputs and missing geometry conditioning (Q1 / W3): Per the rebuttal and Algorithm 1, the network fθ(tk,Uk)f _ \theta(t _ k, \mathbf{U} _ k) receives only a randomly sampled U0∈St(M,N)\mathbf{U} _ 0 \in \mathrm{St}(M, N) without nuclear coordinates r\mathbf{r}, atomic numbers, or S(r)\mathbf{S}(\mathbf{r}). A model conditioned solely on random noise U0\mathbf{U} _ 0 cannot identify which test geometry r\mathbf{r} is being evaluated, nor does the Euclidean update Uk+1=GkU0\mathbf{U} _ {k+1} = \mathbf{G} _ k \mathbf{U} _ 0 satisfy C†S(r)C=IN\mathbf{C}^\dagger \mathbf{S}(\mathbf{r}) \mathbf{C} = \mathbf{I} _ N.

  • Inconsistent timings and speedup claims (Q2 / W2): Including the 1.65 s (salicylic acid) and 19.33 s (Ac-Ala3-NHMe) energy evaluations disclosed in the Q2 reply reduces the claimed 30–50x speedups to ~4–7x (2.14 s total on salicylic acid, vs. 4.65 s for 9 full SCF cycles from MINAO on a laptop CPU) while leaving un-converged DIIS residuals (0.36–1.77 Eh). Moreover, in Table 1, 9 PySCF SCF steps take 2.56 s for PhiSNet yet 12.6–14.6 s for SPHNet/WANet on the same molecule despite <0.05 s GPU inference, indicating severe timing inconsistencies.

  • Unresolved attribution and ethical concerns (W1 / Ethics): Presenting the known Grassmann gradient/Hessian of KS-DFT (Edelman et al., 1998; Dai et al., 2020; Cancès et al., 2021) and Grassmann horizontal lift/orbital transformation (Edelman et al., 1998; VandeVondele & Hutter, 2003) as novel theorems/propositions of this submission (Theorem 3.1, Proposition 3.2, Proposition B.7)—alongside unattributed Riemannian flow matching formulations—goes beyond general background discussion. For the record, I declare that I have no conflict of interest with any of the missing attributions flagged in my review. Coupled with the implausible baseline numbers in Tables 1–2 and the absence of code, I reiterate that I have serious ethical concerns regarding this manuscript.

Reviewer_hzG9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces an SCF-free framework for KS-DFT. It combines two of the existing approaches: Riemannian optimization for DFT [1], and the associative parallel prefix-scan, to predict the occupied orbital coefficients.

[1] https://doi.org/10.1007/s10915-024-02612-3

Critical strengths

None

Critical weaknesses

This paper introduces a promising theory and framework. However, there are many discrepancies between the theory and the implementation/experiments.


Specifically, the major weaknesses are:

  • This paper used a gauge-invariant loss. But this loss does not automatically make the learned parameterization (neural network) gauge-equivariant. More evidence are needed to establish the claimed gauge symmetry.

    • On the one hand, in Appendix B.4, if we rotate the occupied frame U0U_0 by a matrix WW that satisfies W†W=INW^\dagger W=I_N, U′=U0WU'=U_0W, and suppose the complementary frame also changes by VV: U0,⊥′=U0,⊥VU_{0,\perp}'=U_{0,\perp}V (U0,⊥U_{0,\perp} is U0U_0's orthogonal complement). Then, in order to make the update generator Ω[K]\Omega[K] gauge-invariant, KK must also changed (equivariantly) accordingly, K′=V⊤KWK'=V^\top KW.

    • However, from Algorithm 1, it seems the neural network only takes PkP_k as an input. So if the input U0U_0 is rotated by WW, the neural network will still make the same prediction because the neural network's input, PkP_k, is gauge-invariant (doesn't change w.r.t. rotation). Please also define the local tensor xkx_k.

    • Therefore, the designed neural network doesn't seem to keep the gauge-invariant property. To settle this, the author needs to show that, for a fixed molecular geometry, they construct several equivalent representations, U0,U0W1,U0W2,…U_0, U_0W_1, U_0W_2, \ldots. Then show that the neural network offers the exact same downstream property prediction (i.e., energy), and explain how the prediction satisfies K′=V⊤KWK'=V^\top K W.

  • The reported baseline methods have a significantly higher error. For example, PhiSNet (NeurIPS 2021, https://arxiv.org/abs/2106.02347) reports a 1e-4 level orbital energy MAE on its own, yet in this paper, their error becomes 3.85 EhE_h (Table 5, aspirin, same molecules as in the PhiSNet paper). Similar issues happen in all baseline methods.

  • In the Dataset Statistics part, please clarify the size of the training, validation, and testing set. Do these three sets together cover the entire trajectory from the MD17 or MD22 datasets? More importantly, does the testing set covers the diverse enough structures? I.e., testing structures come from the beginning to the end of the MD trajectory.

  • From the paper's Algorithm 1, it seems that the model didn't take the molecular geometry rr as an input. Then, how does the model predict the molecular property without knowing the geometry? It is also unclear how UctxU_\textrm{ctx} is obtained. Does UctxU_\textrm{ctx} come from a previous MD snapshot? Does UctxU_\textrm{ctx} come from the current structure and an initial guess, such as MINAO in pyscf? If we have a standalone molecule, without any trajectory history, how do we construct UctxU_\textrm{ctx}?


For minor weaknesses:

  1. This method is not an equivariant neural network. How would it respond to molecular geometry rotations?
  2. The proposed method seems to have no transferability across molecules. But the baseline methods it compared are designed to model diverse molecules and structures. I.e., QHFlow is equivariant and is trained with different molecules and can make predictions on unseen molecules (different numbers of atoms). Table 4 seems to hard-code the matrix size. Together with the previous point, I think this paper presents a much narrower approach. No cross-molecule transferability, and no equivariance.
  3. Please clarify if the inference time in Table 1 includes the 9 SCF cycles for the baseline method. If they indeed include the time of 9 SCF cycles, PhiSNet takes 2.56s in total; does this mean the 9 SCF cycles take ~2.5 s, so the neural network inference takes ~10s for SPHNet? Why is that slow?
  4. In Table 7, 13 SCF cycles are near enough to converge from scratch. If the molecules were initialized from a pure numerical method, such as MINAO, the error after 13 cycles is around ~1e-9 relative to the ground truth. With MINAO initialization, it usually takes ~16 SCF cycles to converge. But the baseline initialization method in the paper still has a very large error after 13 cycles. And some of the 13 SCF results even have a higher energy than 0 SCF. These need an explanation.
  5. Applying Riemannian descent to Kohn–Sham DFT to bypass the SCF procedure was already established (1998 paper, https://arxiv.org/abs/physics/9806030; and a recent work, https://doi.org/10.1007/s10915-024-02612-3). It is inappropriate to treat "new theoretical formulation" as the main contribution of this paper.
  6. Please clarify how the downstream property is obtained. The neural network predicts the occupied orbital coefficients CC; do we need another diagonalization to obtain the downstream properties?
Questions for authors

See weakness

3 comments

Thank you for your careful reading of our work and your detailed technical comments. In the following sections, we explain our reasoning and respond to the points you raised. We look forward to your feedback on these explanations.


W1/Minor-W1. This paper used a gauge-invariant loss...This method is not an equivariant neural network...

Gauge symmetry in our work is established at the level of the learning objective and geometric propagation, rather than through explicit input-level equivariance of the neural network. While architectural gauge equivariance can be beneficial, it is an additional design choice, not a necessary condition for the theoretical formulation established in this work.

Ensuring equivariance of a neural network under different input representations and learning the physical evolution of electronic states are fundamentally different problems. The former constrains how predictions transform under changes of representation, whereas the latter requires distinguishing physically meaningful horizontal directions from redundant vertical gauge directions throughout electronic-state propagation. This presents a deeper geometric challenge that cannot be resolved merely by enforcing equivariance at the input level. Our work is motivated precisely by this challenge and develops a mathematical framework for learning and propagating electronic states while respecting their underlying geometric structure. We are concerned that such a narrow focus may obscure the broader motivation and central contributions of our work. We hope the reviewer will revisit the manuscript with openness to the scientific questions it addresses.

W2/Minor-W4. The reported baseline methods have a significantly higher error...In Table 7, 13 SCF cycles...

In our experimental setting, all models were trained to predict the displacement between random matrix and the reference electronic states, rather than relying on conventional MINAO initialization or other prior protocol. This introduces a substantially more challenging learning problem, which accounts for the differences in baseline errors compared to their original implementations. Similarly, convergence within 13 SCF iterations is not guaranteed under this initialization setting, unlike the MINAO-based scenario described by the reviewer. We will clarify this point in the revised manuscript.

W3. In the Dataset Statistics part...

The MD17 and MD22 trajectories were randomly split into training, validation, and test sets with an 8:1:1 ratio. This random splitting ensures that the test set includes diverse molecular configurations sampled throughout the trajectories, rather than being restricted to a particular temporal segment.

W4. From the paper's Algorithm 1...

Our model learns the energy-minimizing gradient flow on the Grassmann manifold, starting from random matrix on the Stiefel manifold defined by the geometry-dependent overlap matrix S(r)S(r). In our setting, UctxU_{\mathrm{ctx}} is simply taken as U0U_0, the initial orbital coefficients for the current molecular structure. This requires neither a previous MD snapshot nor MINAO initialization.

Minor-W2. The proposed method seems to have no transferability...

Cross-molecule transferability and rotational equivariance are important research directions, and existing approaches have made substantial efforts toward achieving these capabilities across diverse molecular datasets. The breadth of applicability should not be confused with the significance of the underlying problem being solved. Before extending a learned electronic solver across diverse molecular systems, a fundamental challenge is to establish whether electronic states themselves can be accurately learned and propagated without relying on conventional SCF procedures. Our work addresses this foundational challenge by establishing a geometric learning framework and demonstrating its feasibility on challenging molecular systems. Moreover, our geometric formulation is not inherently tied to a particular molecule and can be further developed to incorporate cross-molecule transferability and rotational equivariance.

Beyond this, our work offers a distinct perspective from existing Hamiltonian-centered approaches by directly modeling the flow of electronic states on the Grassmann manifold. We hope that introducing this perspective will broaden the conceptual and methodological scope of electronic structure learning and encourage further exploration of diverse approaches to learning electronic states.

Minor-W3. Please clarify if the inference time in Table 1...

The reported inference times in Table 1 include both neural network inference and the SCF cycles. SPHNet has a relatively high computational cost for neural network inference, which contributes to its longer overall runtime. In addition, the computational cost of SCF cycles can vary depending on the initialization, as different initializations may affect the computational workload within each SCF iteration.

Minor-W5. Applying Riemannian descent to Kohn–Sham DFT...

The cited works indeed establish important foundations for Riemannian optimization in KS-DFT, which we acknowledge. However, our theoretical formulation goes beyond classical Riemannian descent. We introduce learnable gauge-symmetric horizontal flows on the Grassmann manifold and establish associative, geometry-preserving bundle actions that enable parallel prefix-scan propagation.

From the perspective of learning, the distinction is even more important. Existing learning-based approaches have predominantly focused on predicting Hamiltonians or related electronic quantities. Our approach instead makes the evolution of electronic states itself the object of learning. This shift in perspective calls for a mathematical framework that unifies learnable geometric dynamics with efficient parallel propagation. We therefore establishes a fundamentally new theoretical formulation for electronic structure learning, rather than merely adapting classical Riemannian descent to an existing learning pipeline.

Minor-W6. Please clarify how the downstream property is obtained...

The computation of downstream properties depends on the specific quantity of interest. Our model directly predicts the occupied orbital coefficients, from which the density matrix can be constructed. Properties involving the Fock matrix require its additional construction from the predicted density, while eigenvalue-based quantities, such as orbital energies and HOMO–LUMO gaps, require further diagonalization. Importantly, such operations are standard throughout the electronic structure prediction literature, appearing in both preprocessing and post-processing stages depending on the method. Hamiltonian-based approaches likewise rely on related electronic structure operations, such as diagonalization to recover orbital coefficients. These computational requirements are therefore shared across existing approaches rather than being unique to our framework.

The author's reply doesn't address my concerns:

W1/Minor-W1. This paper used a gauge-invariant loss...This method is not an equivariant neural network...

I don't think the author can claim a major theoretical breakthrough as the main novelty. Edelman et al. (https://arxiv.org/abs/physics/9806030) and Altmann et al. (https://link.springer.com/article/10.1007/s10915-024-02612-3), both cited in this paper, have already established the direct minimization of energy method using Riemannian optimization. The remaining challenge is the implementation, which the author just acknowledged they didn't address at all.

In simple terms, initial orbital coefficients can be rotated by unitary matrices. The ML model used in this paper will break completely under any such rotations because the proposed method didn't address the core equivariance challenge.

W2/Minor-W4. The reported baseline methods have a significantly higher error...In Table 7, 13 SCF cycles...

"displacement between random matrix and the reference electronic states": I don't think this can fully explain why the baseline is weaker than the original publication. While QHFlow indeed starts from a MINAO guess, methods such as QHNet do not need MINAO as a starting point. More specifically, QHNet is an end-to-end network, taking molecular geometry (atom type and its coordinates) as input to predict the Hamiltonian matrix. I am not clear on what "random matrix" means in the author's comment, and how it relates to QHNet. As the baseline method shows several orders of magnitude of error compared to the published checkpoint, this paper must disclose how they weaken these methods and why it is a legitimate choice.

Minor-W3. Please clarify if the inference time in Table 1...

"In addition, the computational cost of SCF cycles can vary depending on the initialization": this is generally not true. Each SCF cycle takes the exact same number of numerical operations. No matter what the initialization is, time spent in each cycle should be very similar. Since we now know Table 1 includes both neural network inference and the SCF cycles, then a remaining question is that, why SPHNet takes ~10s in inference? From the SPHNet paper, its inference takes less than 0.05 s. Why it become that slow in this work?


Together with my previous point, this paper claims accuracy and speed as one core contribution, but there are many uncertainties about the baseline methods' performance and inference time.

Comments on the paper

1 comment
Reviewer_DXoXOpenReview ↗

PySCF Sanity Check (Table 1, Salicylic Acid)

To save readers and the AC time scrolling through the thread: the rebuttal reveals that all baselines in Tables 1–2 were evaluated starting from random Stiefel matrices U0∈St(M,N)\mathbf{U} _ 0 \in \mathrm{St}(M, N). Running a direct sanity check in the paper's exact PySCF setup (def2-SVP, PBE, grid level 3; Apple M4 Pro 14-core CPU, 24 GB RAM, PySCF 2.15.0) shows two critical issues with the empirical results in this submission:

  1. 9 SCF steps from pure Gaussian random noise beat 5 of the 6 trained ML baselines by 20–160x: Even starting from untrained random Stiefel noise U0\mathbf{U} _ 0, 9 standard PySCF SCF steps converge to -495.091 Eh (0.020 Eh error), whereas Table 1 reports -491.94 to -494.71 Eh (0.41–3.17 Eh error) for 5 of the 6 trained baselines after 9 SCF steps. Moreover, geometry-to-Hamiltonian models (QHNet, DEQHNet, WANet, SPHNet, PhiSNet) do not even take U0\mathbf{U} _ 0 as input. The baseline numbers in Tables 1–2 therefore reflect either severe mis-training, fundamental misuse of PySCF, or invalid reporting.
  2. Zero-learning baselines (MINAO, HSAP, and previous MD frame) beat all 0-SCF baselines by 10–100x, and a previous MD frame matches or beats "Ours" in 0.3 ms: Because a separate model is trained per single-molecule MD trajectory (8/1/1 split), simply Grassmann-projecting orbitals from a previous MD frame (Cprev(Cprev⊤StCprev)−1/2\mathbf{C} _ {\mathrm{prev}}(\mathbf{C} _ {\mathrm{prev}}^\top \mathbf{S} _ t \mathbf{C} _ {\mathrm{prev}})^{-1/2}) takes 0.0003 s (0.3 ms, 0 Fock builds) and outperforms or matches "Ours" across energy, gap, and DIIS residual with zero learned parameters:
Method (Salicylic Acid, Table 1 setup)SCFE\mathcal{E} (EhE _ h)∥E−E⋆∥\|\mathcal{E} - \mathcal{E}^\star\| (EhE _ h)Sc\mathfrak{S} _ c (%)ϵocc\epsilon _ {\mathrm{occ}} (EhE _ h)vϵv\epsilon (EhE _ h)∥ϵDIIS∥\|\boldsymbol{\epsilon}^{\mathrm{DIIS}}\| (EhE _ h)H\mathbf{H} MAE (mEh\mathrm{m}E _ h)D\mathbf{D} MAE (×10−3\times 10^{-3})State (+E-eval = Total) Time (s)
Paper ML Baselines (Table 1 range)0−220.4-220.4 to −454.4-454.440.7–274.740.7\text{–}274.721.0–67.021.0\text{–}67.00.94–14.20.94\text{–}14.20.062–0.710.062\text{–}0.71————
Paper ML Baselines (Table 1 range)9−491.9-491.9 to −495.0-495.00.08–3.170.08\text{–}3.1794.9–99.994.9\text{–}99.90.018–0.150.018\text{–}0.150.008–0.110.008\text{–}0.111.07–6.501.07\text{–}6.50——2.56–14.62.56\text{–}14.6 (their RTX 5000 + CPU)
Paper "Ours" (Table 1 + Q2 reply)0−495.109-495.109∼0.002\sim 0.00299.9999.993.55×10−43.55 \times 10^{-4}4.54×10−44.54 \times 10^{-4}0.3630.363——0.490.49 (+1.65=2.14+1.65 = 2.14, their RTX 5000 + CPU)
Random Stiefel U0\mathbf{U} _ 0 (untrained noise)0−143.787-143.787351.32351.3218.2018.203.953.950.2100.21041.241.2198.0198.026.126.10.010.01 (+2.48=2.49+2.48 = 2.49, laptop)
Random Stiefel U0\mathbf{U} _ 0 + 9 PySCF SCF steps9−495.091-495.0910.0200.02099.9699.962.10×10−22.10 \times 10^{-2}1.40×10−21.40 \times 10^{-2}0.4800.4800.850.850.310.314.704.70 (+0.25⋆=4.95+0.25^\star = 4.95, laptop)
HSAP (T+VSAP\mathbf{T} + \mathbf{V} _ {\mathrm{SAP}}, 0 Fock builds)0−491.949-491.9493.163.1695.7495.740.3940.3940.1010.1013.443.4428.028.04.274.270.410.41 (+2.46=2.87+2.46 = 2.87, laptop)
MINAO (Grassmann-projected, 0 Fock builds)0−493.233-493.2331.881.8895.5095.500.2270.2270.0730.0733.083.087.647.643.843.840.040.04 (+2.48=2.52+2.48 = 2.52, laptop)
MINAO (1 Fock build F[DMINAO]\mathbf{F}[\mathbf{D} _ {\mathrm{MINAO}}])1−494.841-494.8410.270.2799.5999.590.0940.0940.0200.0201.051.052.382.381.031.032.402.40 (+0.27⋆=2.67+0.27^\star = 2.67, laptop)
MINAO + 9 PySCF SCF steps (no ML)9−495.111-495.1113.1×10−73.1 \times 10^{-7}100.00100.009.69×10−59.69 \times 10^{-5}8.59×10−58.59 \times 10^{-5}1.05×10−31.05 \times 10^{-3}0.00450.00450.00120.00124.654.65 (+0.25⋆=4.90+0.25^\star = 4.90, laptop)
Prev MD Frame (distant δr≈0.03 A\delta r \approx 0.03\,\text{A}, 0 Fock)0−495.099-495.0990.0120.01299.8699.862.36×10−32.36 \times 10^{-3}5.03×10−45.03 \times 10^{-4}0.1710.1710.180.180.750.750.00030.0003 (+2.48=2.48+2.48 = 2.48, laptop)
Prev MD Frame (adjacent δr≈0.01 A\delta r \approx 0.01\,\text{A}, 0 Fock)0−495.110-495.1108.5×10−48.5 \times 10^{-4}99.9899.987.24×10−47.24 \times 10^{-4}1.65×10−41.65 \times 10^{-4}0.0440.0440.0370.0370.220.220.00030.0003 (+2.48=2.48+2.48 = 2.48, laptop)

(⋆^\starIn PySCF, the first Fock build initializes the Becke grid and integral screening (~2.2 s), after which incremental Fock/energy evaluations on the same RKS object take ~0.25–0.32 s. Laptop timings measured on an Apple M4 Pro 14-core CPU, 24 GB RAM, PySCF 2.15.0.)

This submission is a clear example of why we as an academic community need to reconsider how reviewing capacity is allocated and how to prevent the unvetted submission and resubmission of drafts like this to A and A* venues, as we cannot afford to waste volunteer reviewing time on work with such fundamental scientific and ethical issues.

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

Related papers

Discussion (0)

Sign in to comment.