acceptodds
Under review as a conference paper at ICLR 2027

Physics-Aligned Electronic Ground-State Learning Improves Generalization

Abstract

Machine-learned interatomic potentials (MLIPs) excel at in-distribution tasks, accelerating drug and material development, yet they struggle to generalize out-of-distribution. We propose to push the cost-accuracy Pareto frontier by designing observable-agnostic electronic ground-state descriptor models (GSMs) with computational costs situated between MLIPs and Kohn-Sham density functional theory (KS-DFT). We align the learning objectives and architectures of GSMs with the governing equations of KS-DFT by enforcing physical constraints and removing optimization pressure on unphysical or irrelevant degrees of freedom. In our size-extrapolation experiments from QM9 to QM40, our combined contributions OrthoNormal-Loss (ON-Loss) and Grassmann Restricted Occupied-Orbital Training (GROOT) reach a 79.1% energy and 83.4% force mean absolute error (MAE) reduction over previous state-of-the-art density GSMs. For Hamiltonian GSMs, ON-Loss and Residual Optimal-gauge Conditioning-aware KS-Eq. Training (ROCKET) together reduce the energy and force MAEs of the strongest baseline by 99.8% and 95.9%, respectively. Using a self-consistency rejection criterion, we filter out extrapolation errors on QMugs, rejecting fewer than 0.4% of predictions while reaching an energy MAE of 0.07 mHa. Finally, we demonstrate the efficiency of label-free self-consistency fine-tuning and transfer GSMs to reactive chemistry in Transition1x, reaching energy errors below chemical accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4Science, Kohn-Sham, Density Functional Theory, DFT, KS-DFT, Ground-State, Electronic Structure, Quantum Chemistry, Computational Chemistry

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_qzEwOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
3(+1)
Read review
Paper summary

This paper introduces strategies to improve the downstream utility of predicted Hamiltonian or Density matrices. The paper is formulated as kind of a 'bag of tricks', so I'll comment on them separately. Note that I did not read the appendices unless mentioned.

(1) The problem addressed with the ON-loss has been treated quite extensively in the wavefunction alignment loss paper (I read Appendix F), and the additional contributions made here amount to minor reformulations and pointing out possible inconsistencies in the released code implementation, which are only moderately useful. The core idea behind the wavefunction alignment loss has been used in other work (QHFlow). I would disagree with the statement that supervision in the "ON basis" is poorly understood, at least for Hamiltonian prediction. The NON-basis is used for supervision in models which treat larger systems (eg, DeepH-pack, HELM) due to lower computational cost, with the limitations in mind.

(2) The issue of gauge invariance (here ROCKET) - for molecules, this should not be an issue in theory, because they share a common energy reference (the vacuum level). Maintaining this common reference allows for element-wise referencing procedures to be applied, as was done in MACE-H and HELM. The paper states, "a standard matrix loss would strictly penalize gauge variation between molecules, while Eq. (9) permits arbitrary gauge variation ", however I would expect to see no arbitrary gauge variation in this training data, and am confused why including this degree of freedom results in performance improvements during training. I would like to know the exact details behind how c- and c* are computed over a given dataset, whether element-wise referencing is included in the baselines, and more intuition behind why this helps, beyond ablations. This very interesting point is just skimmed over.

(3) The last point (GROOT) mentions that constrained learned density matrices to the manifold of 'valid' projectors improves learning. This makes a lot of sense, and I've seen similar ideas very recently elsewhere. However, it should be clear that this involves diagonalizing the predicted matrix, which limits scalability.

Critical strengths

I don't see a single, substantial strength in the paper. This might just be because the message is diluted into multiple smaller contributions which are confusingly written, so it's hard to come away with a clear 'takeaway'.

Critical weaknesses

The first significant weakness of this paper is that it's incredibly difficult to read. There is very little signposting, and walls of text which switch between different points. The frequent use of bold font and italic font is distracting. The contributions, ablations, and experiments are not in order or nicely differentiated. It takes a long time of staring at the figures to understand what is being shown, as the way the data is presented is overly 'synthesized'. It feels like this manuscript is not designed with a human reader in mind.

Questions for authors

(Please see the summary section)

I think satisfying the following would improve the overall volume of contribution in the paper to be a strong contribution:

(1) Showing evaluations on open-shell systems and heavier elements than CHNOF. The paper states, "Although a generalization to unrestricted systems is straightforward, this has not been covered by our evaluations. " so this should not be too difficult to evaluate.

(2) Quantifying evaluation runtimes, time complexity, and how they change with system size for each of the optimizations presented. Currently, the paper states that only a single Fock build is used, thus positioning this method in-between MLIPs (zero Fock builds), and multi-build Hamiltonian models. I think the paper should quantify this, to better understand any tradeoffs with scalability. This could be gathered from the existing data collected. It would also be good to understand how this compares with Hamiltonian models which do not require an initial matrix construction.

(3) Data scaling: What would take these contributions to the 'next-level', is if the performance improvements demonstrated are maintained even when larger amounts of training data is used, for example when training (not just evaluation) on fractions of the larger dataset used. This would place these observations into the regime of 'contributions which cannot be effectively replaced with sufficient training data'.

Minor points:

  • Number of electrons should be replaced with numbers of atoms. Since heavy elements are already omitted, using # electrons doesn't add anything but makes the sizes difficult to compare with other work.
17 comments
AuthorsOpenReview ↗

Response to question (1) from summary

We thank the reviewer for their response.

What we claim is:

  1. Why it works. We connect the supervision basis to observable errors through worst-case bounds in the overlap spectrum (Sec. 4, Table 1; App. C) and through a directional sensitivity analysis showing that LON\mathcal{L} _ {\mathrm{ON}} weights matrix directions approximately as the observables do, while LNON\mathcal{L} _ {\mathrm{NON}} does not (Sec. 5, Fig. 2a–d; App. D).
  2. The consequence. The NON terms should be dropped entirely, and the model should also predict in the ON basis, instead of adding an alignment term to NON supervision.
  3. Densities. The same analysis applies to density matrices, which WALoss does not cover.

The revised manuscript tests the second and third claim by separating the readout basis from the loss basis (App. R, Fig. 27). On QM9 →\to QM40, at one Fock build and averaged over all molecules:

ModelReadoutLossEE [mHa]FF [mHa/Å]
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})NONLNON\mathcal{L} _ {\mathrm{NON}}707254
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})ONLNON\mathcal{L} _ {\mathrm{NON}}2.455.14
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})ONLON\mathcal{L} _ {\mathrm{ON}}1.09‾\underline{1.09}3.26‾\underline{3.26}
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}NONLNON\mathcal{L} _ {\mathrm{NON}}2.177.51
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}ONLNON\mathcal{L} _ {\mathrm{NON}}34,6896,668
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}ONLON\mathcal{L} _ {\mathrm{ON}}0.621.59

For Hamiltonians, predicting in the ON basis gives most of the gain, and LON\mathcal{L} _ {\mathrm{ON}} roughly halves the remaining error. For densities, an ON readout under LNON\mathcal{L} _ {\mathrm{NON}} fails completely, and only the combination of ON readout and LON\mathcal{L} _ {\mathrm{ON}} improves on the NON baseline.

The idea is indeed known for Hamiltonians. What we add is to follow it through, which current practice does not yet do. WALoss adds its alignment term to elementwise terms in the NON basis (their App. K, Eq. 19). QHFlow trains in the NON basis and uses only the orbital-energy (diagonal) part of the alignment term for post-hoc fine-tuning (their Eq. 7).

It is a fair point that the NON Hamiltonian is more localized than its Löwdin counterpart, which matters for linear-scaling and periodic applications. It also explains the deliberate choice by HELM etc. to use the NON basis for their training. For density matrices the NON basis has no such advantage: both representations decay at nearly the same rate with atom-pair distance (App. G, Fig. 6).

The practical question is whether the weaker localization of the ON Hamiltonian hurts at the cutoffs the model uses. We keep the cutoffs of QHFlow2 for all models (5 Å backbone, 15 Å readout). At 15 Å the entries of the ON target have already dropped about four orders of magnitude below their typical size (Fig. 6), so the readout cutoff removes almost nothing. In line with this, the ON models keep extrapolating up to 319 heavy atoms at these cutoffs, while the NON baselines break (App. T).

On cost, the Fock build dominates one-shot molecular observables: setup and model inference together take about 1–2% of the time of an SCF with forces (App. N.3), and training time is the same for all losses and readouts. The ON-to-NON transform adds one symmetric eigendecomposition of S\mathbf{S} per molecule, which has the same cubic cost as a single eigen-solve of an SCF cycle.

Changes in the manuscript: readout vs. loss ablation (App. R, Fig. 27; Table 37), QHNet validation on QH9 (App. Q), cost breakdown (App. N).

AuthorsOpenReview ↗

Response to question (2) from summary

The reviewer is right that molecules share the vacuum level as a common energy reference. For absolute Hamiltonians this means that composition determines the constant shift of a molecule almost completely, which is why element-wise referencing as in MACE-H and HELM works there (App. H, Fig. 8, top).

Our Hamiltonian models, in contrast, learn the correction to the SAP guess, ΔHSAP=H−HSAP\Delta\mathbf{H} _ {\mathrm{SAP}} = \mathbf{H} - \mathbf{H} _ {\mathrm{SAP}}. The constant shift is then the offset of this Δ\Delta-target, c∗=Tr(ΔH~SAP)/Bc ^ {\ast} = \mathrm{Tr}(\Delta\widetilde{\mathbf{H}} _ {\mathrm{SAP}})/B, which is the mean orbital-energy level of the true Hamiltonian relative to that of the guess. Measured from the guess instead of the vacuum, this offset is a property of the whole molecule: it varies between molecules of the same composition and grows with molecule size (Fig. 8, bottom). This is the gauge variation between molecules that the ROCKET loss permits, and ROCKET removes exactly this molecule-specific offset of ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}, on which no observable depends.

In detail, the per-molecule shift c∗c ^ {\ast} is the closed-form minimizer of the loss, computed for every molecule at every training step:

c∗=arg⁡min⁡c∥ΔH~−cI∥F=1BTr(ΔH~).c ^ {\ast} = \arg\min _ c \Vert\Delta\widetilde{\mathbf{H}} - c\mathbf{I}\Vert _ F = \frac{1}{B}\mathrm{Tr}(\Delta\widetilde{\mathbf{H}}).

The reference cˉ\bar{c} starts at the mean of c∗c ^ {\ast} over the training split and follows the model as an exponential moving average of the batch means:

cˉ←γ cˉ+(1−γ) ⟨c∗⟩batch,γ=0.99.\bar{c} \leftarrow \gamma\thinspace\bar{c} + (1 - \gamma)\thinspace\langle c ^ {\ast}\rangle _ {\mathrm{batch}}, \qquad \gamma = 0.99.

The second-stage loss combines both:

LROCKET(2nd stage)=∥ΔH~−c∗I∥FB+αROCKET (c∗−cˉ)2.\mathcal{L} ^ {\mathrm{(2nd\ stage)}} _ {\mathrm{ROCKET}} = \frac{\Vert\Delta\widetilde{\mathbf{H}} - c ^ {\ast}\mathbf{I}\Vert _ F}{B} + \alpha _ {\mathrm{ROCKET}}\thinspace(c ^ {\ast} - \bar{c}) ^ 2.

The gradient of the first term is traceless and supervises everything except the shift; the gradient of the second term is a multiple of I\mathbf{I} and moves the shift alone. The weight αROCKET\alpha _ {\mathrm{ROCKET}} is set once from the spread of c∗c ^ {\ast} over the training split (App. H, Eq. 80; Table 17).

The revision includes element-wise referencing among the baselines in two forms (Tables 38–39): HELM-style per-element normalization of the scalar node blocks on absolute targets, and a per-element on-site reference on top of ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}. On QM9 all variants are on par with ROCKET. On QM40 the restrained per-molecule shift of ROCKET is the most accurate, with 0.4 mHa energy MAE compared to 6.6 mHa for the per-element reference on ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}.

The intuition for why this helps is that the offset of ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}} is a global quantity that a local model would have to assemble from per-atom outputs, and it changes with molecule size. With the offset removed, the scalar on-site blocks of H, C and O stay close to their training averages across molecule sizes and datasets (Fig. 9). The remaining target is local and transfers to larger molecules. The weak restraint towards cˉ\bar{c} keeps the constant in the model output fixed during training (Fig. 10), which makes the per-atom outputs agree on one level also for molecules larger than those seen in training.

App. H makes this connection between ROCKET and element-wise referencing quantitative and explains it in more detail.

QM9 →\to QM40 held-out test splits (16 ep., all molecules [kept]; Tables 38–39)QM9 Test (2,000) EKSE _ {\mathrm{KS}} [mHa]QM40 (1,000) Rej.QM40 EKSE _ {\mathrm{KS}} [mHa]QM40 EHFE _ {\mathrm{HF}} [mHa]
Abs. H\mathbf{H} + per-elem. mean / +σZ+\sigma _ Z, NON readout0.15 / 0.1584.4% / 84.2%5,407 [3.00] / 6,585 [2.78]3,534 / 4,178
Abs. H\mathbf{H} + per-elem. mean / +σZ+\sigma _ Z, ON readout1.36 / 0.9993.3% / 86.7%45.9 [5.75] / 47.2 [5.57]0.59 / 0.69
ΔHSAP+cZ\Delta\mathbf{H} _ {\mathrm{SAP}} + \mathbf{c} _ Z per-element reference (ON)0.04212.5%6.72 [3.18]6.57 [2.47]
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S}) single constant (ON, 3 seeds)0.1320.97%3.07 [2.91]1.09 [0.90]
ROCKET stage 1 (unrestrained c∗c ^ {\ast})0.0388.50%6.84 [3.61]5.07 [2.16]
ROCKET stage 2 (EMA-restrained c∗c ^ {\ast}, 3 seeds)0.0400.07%1.71 [1.70]0.41 [0.40]
Values in brackets are averaged over the molecules kept after self-rejection.
AuthorsOpenReview ↗

Response to question (3) from summary

It is correct that GROOT requires one eigendecomposition of a B×BB \times B matrix to place the prediction on the manifold of valid projectors. This is the same cubic-cost operation that a conventional SCF calculation performs in every cycle. For the system sizes we tested, the cost of a ground-state descriptor model is set by the Fock build that turns the predicted density into observables: with direct four-center integrals, setup and model inference together take about 1–2% of the time of an SCF with forces, up to 182 heavy atoms and 3,378 basis functions (App. N.3, Fig. 20). Training time is the same as for the unconstrained density models (App. N.1, Fig. 17).

The projection also speeds up the Fock build when density fitting is used. Because GROOT outputs a rank-OO projector, the exchange contraction (RI-K) runs over the OO occupied orbitals instead of the full B×BB \times B density matrix. In our timings up to 101 heavy atoms, this makes the Fock build 1.6–5.4x faster than for the unconstrained density model Pbase\mathbf{P} _ {\mathrm{base}} trained with LNON\mathcal{L} _ {\mathrm{NON}} (Fig. 19). In conlcusion we do not see limited scalability of our method withing the size regimes tested.

Reviewer_5s8qOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how the basis and loss function, along with output constraints, affect ground-state generalization on closed-shell molecules. It proposes ON-Loss in an orthonormal basis and GROOT for idempotent densities via rotations (GROOTrotGROOT_{\text{rot}}) or projections (GROOTprojGROOT_{\text{proj}}), plus ROCKET for Hamiltonians with scalar shifts and subspace masking. At test time, it applies a one-step Harris-Foulkes correction with filtering and fine-tuning on transfer from QM9 to QM40, QMugs, 3BPA, and Transition1x at B3LYP/def2-SVP.

Critical strengths

Building physical constraints into the readout and loss makes mathematical sense. Eqs. (3) and (7) with Appendix C.1 show how the overlap spectrum distorts standard matrix losses, while Appendix C.4 shows that keeping GROOT on the projector manifold makes the KS energy error quadratic rather than first order. Table 26 also has the exact ablation I wanted, showing that training through the projection beats test-time purification of unconstrained outputs.

To be fair, the transfer gains on a shared backbone are large. On QM9 to QM40 in Fig. 1 and Table 26, GROOTrotGROOT_{\text{rot}} cuts unfiltered energy MAE from 2.172.17 to 0.104 mHa0.104\text{ mHa} and force MAE from 7.517.51 to 0.889 mHa/A0.889\text{ mHa/A} over NON across three seeds. On QMugs in Table 28, GROOTprojGROOT_{\text{proj}} and ROCKET reach 0.2090.209 and 0.557 mHa0.557\text{ mHa} with under 0.4%0.4\% rejection, and Appendix O shows 0.700.70 and 0.720.72 median SCF time ratios against MINAO.

Critical weaknesses
  1. Calling ON-Loss a new metric is an overstatement, and its effect is not isolated. In Appendix F, Eqs. (67) and (68) show that for complete orbitals with C∗⊤SC∗=IC^{\ast\top}SC^{\ast}=I, the full matrix WA term satisfies ∥C∗⊤ΔHC∗∥F=∥S−1/2ΔHS−1/2∥F\Vert C^{\ast\top}\Delta H C^{\ast}\Vert_F=\Vert S^{-1/2}\Delta H S^{-1/2}\Vert_F, matching uniformly weighted WA up to normalization. I also doubt the gains come from ON-Loss alone. In Fig. 14 and Table 9, NON predicts in the AO basis while ON predicts in the Löwdin basis with different scaling. Changing readout and loss together confounds the two: Table 17 tests an AO readout for GROOTprojGROOT_{\text{proj}}, but the unconstrained density model lacks this control and the Hamiltonian run stopped before QM40.
  2. My main worry with ROCKET in Secs. 4 and 5 is that parts of the math are unclear, so I lean toward a weak reject. ROCKET is motivated by shift invariance: a uniform shift ΔH~=cI\Delta\widetilde H=cI leaves PP unchanged and should cost nothing. Stage 1 handles this, but I am not sure Stage 3 does. As I read Eq. (12), the Eq. (11) mask is applied first and the best constant is subtracted afterwards. If so, the mask turns cIcI into cQcQ with Q=Q~O+kQ=\widetilde Q_{O+k}, and no single constant cancels it, which would leave a penalty of ∣c∣r−r2/B/B>0\vert c\vert\sqrt{r-r^2/B}/B > 0 for c≠0c\ne 0 and 0<r=rank⁡(Q)<B0 < r=\operatorname{rank}(Q) < B. I would like the authors to clarify the order of these steps, and whether a Hamiltonian that is exact up to a shift is still penalized in Stage 3. The masking also has a gap. Stage 3 uses the mask MQ(A)=A−(I−Q)A(I−Q)\mathcal M_Q(A)=A-(I-Q)A(I-Q) with Q=Q~O+kQ=\widetilde Q_{O+k}, so errors in the high virtual block (I−Q)ΔH~(I−Q)(I-Q)\Delta\widetilde H(I-Q) are not graded. This assumes those virtual levels stay above the occupied ones. In theory, a virtual level could drop below an occupied level at zero masked error, and then the aufbau filling would pick the wrong orbitals and change PP.
  3. Several equations and details do not line up. Eqs. (2) and (16) use a per-spin density and Appendix C.3 writes ∇E[P]=2H[P]\nabla E[P]=2H[P], yet Sec. 5 and Eq. (57) drop the factor of 2. Table 22 lists self-consistency losses by residuals without explaining solver gradients, and Appendix K on training costs is blank.
  4. The paper is hard to read. A lot is packed into a few pages: the ON basis, ON-Loss, three ROCKET stages, the EMA term, the masking, GROOT, and the energy correction all show up close together, and key definitions are left to the appendix. I had to go back and forth between Secs. 3 to 5 and the appendices just to work out how the pieces fit, and it is still not clear which component is meant to solve which problem. A short overview at the start, one running example, and moving some details out of the main text would help a lot.
Questions for authors
  1. With basis and scaling fixed along with readout, how much QM40 gain comes from swapping NON for ON loss on unconstrained density and Hamiltonian models?
  2. In third stage ROCKET, do you subtract the scalar shift before masking or after as in Eq. (12), and how do you stop virtual states from crossing below occupied ones?
  3. Why does Eq. (57) drop the factor of 2 from ∇E[P]=2H[P]\nabla E[P]=2H[P] in Appendix C.3, and do Table 3 forces at Paufbau(F[P^])P_{\mathrm{aufbau}}(F[\widehat P]) match corrected energy finite differences?
  4. Can you test original QHNet or QHNetV2 on standard QH9 splits with matched compute, rather than 2,000-molecule subsets on your backbone in Tables 10 and 19 (and Table 25)?
  5. Because ROCKET removes the uniform shift c∗c^\ast, the absolute eigenvalues of H^\widehat H are only fixed up to a constant. How reliable are quantities that depend on absolute levels, such as the HOMO or LUMO energy, when read directly from H^\widehat H, and do you ever report them that way?
11 comments
AuthorsOpenReview ↗

Response to weakness (1) and question (1)

Thank you for the detailed review. We read two questions in this point. The first is whether ON-Loss is new, given its close relation to WALoss. The second is whether the gains come from the loss itself, since our comparison changed the readout basis and the loss basis together.

On the first question, we do not claim ON-Loss as a new metric. The full-matrix alignment term of WALoss is the same loss up to normalization (App. F). What we contribute instead is the following.

  1. Why it works. We connect the basis of supervision to the errors of the observables, with worst-case bounds (Sec. 4, Table 1) and a sensitivity analysis (Sec. 5, Fig. 2). LON\mathcal{L} _ {\mathrm{ON}} weights the matrix directions as the observables do, and LNON\mathcal{L} _ {\mathrm{NON}} does not.
  2. The consequence. WALoss adds the alignment term to a NON loss. We drop the NON terms entirely and also predict in the ON basis.
  3. Densities. The same analysis holds for density matrices, which WALoss does not cover.

On the second question, the two changes were indeed confounded. We added an ablation that varies readout and loss separately on QM9 →\to QM40 (App. R, Fig. 27).

ModelReadoutLossEE mean [mHa]EE median [mHa]FF mean [mHa/Å]FF median [mHa/Å]
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})NONLNON\mathcal{L} _ {\mathrm{NON}}7071.632546.93
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})ONLNON\mathcal{L} _ {\mathrm{NON}}2.450.925.144.20
Δ(HSAP+ctS)\Delta(\mathbf{H} _ {\mathrm{SAP}} + c _ t\mathbf{S})ONLON\mathcal{L} _ {\mathrm{ON}}1.090.423.262.83
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}NONLNON\mathcal{L} _ {\mathrm{NON}}2.170.807.516.41
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}ONLNON\mathcal{L} _ {\mathrm{NON}}34,68930,3186,6686,747
ΔPMINAO\Delta\mathbf{P} _ {\mathrm{MINAO}}ONLON\mathcal{L} _ {\mathrm{ON}}0.620.411.591.36

The second and third row of each model share the ON readout and differ only in the loss, which is the comparison you ask for. For Hamiltonians, the ON readout removes the outliers of the NON baseline, and LON\mathcal{L} _ {\mathrm{ON}} then roughly halves the remaining energy error. For densities the loss is decisive. An ON readout under LNON\mathcal{L} _ {\mathrm{NON}} fails completely, and the same readout under LON\mathcal{L} _ {\mathrm{ON}} clearly improves on the NON baseline. QHNet on QH9 shows the same on a different architecture (App. Q).

The detailed treatment is given in App. R for the ablation of readout and loss, and in App. F for the relation to WALoss.

QHNet on QH9-stable-ood (20 epochs, 9,335 test mol., one Fock build)ReadoutLossEKSE _ {\mathrm{KS}} [mHa]
ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}NONLNON\mathcal{L} _ {\mathrm{NON}}55.8
ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}NONLON\mathcal{L} _ {\mathrm{ON}}65.1
ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}ONLNON\mathcal{L} _ {\mathrm{NON}}1.04
ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}ONLON\mathcal{L} _ {\mathrm{ON}}0.444
ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}}ONROCKET0.132
AuthorsOpenReview ↗

Response to question (4)

We trained the original QHNet with 20M parameters on the standard QH9-stable-ood split, from geometry alone and with the published AdamW recipe of QH9, and evaluated it on the full test set. Nothing of our backbone, input or optimizer is used, so only the target, readout and loss differ from the published model.

The results from our backbone carry over. After 20 epochs, the NON readout under LNON\mathcal{L} _ {\mathrm{NON}} has an energy error of 55.8 mHa, while the ON readout under LON\mathcal{L} _ {\mathrm{ON}} reaches 0.44 mHa and ROCKET 0.13 mHa, both without a single self-rejected prediction (Fig. 26).

After the full 95 epochs of the published recipe, QHNet with ROCKET reaches 0.054 mHa at one Fock build, compared to 12.5 mHa for the published QHNet. This is also below the QHNet and QHFlow models that use two Fock builds (Table 22).

QH9-stable-ood (9,335 test mol.; Table 22, Fig. 26)Fock buildsEpochsEKSE _ {\mathrm{KS}} MAE, all molecules [mHa]Would be self-rejected
QHNet Habs(NON)\mathbf{H} _ {\mathrm{abs}} ^ {(\mathrm{NON})} (published checkpoint)165–8012.56.8%
QHNet Habs(NON)\mathbf{H} _ {\mathrm{abs}} ^ {(\mathrm{NON})} (reported by Kim et al., 2026)18015.7–
QHNet Habs(NON)\mathbf{H} _ {\mathrm{abs}} ^ {(\mathrm{NON})} (retrained, full recipe)19578.710.8%
QHNet + ΔHSAP(ON)\Delta\mathbf{H} _ {\mathrm{SAP}} ^ {(\mathrm{ON})}, LNON\mathcal{L} _ {\mathrm{NON}} (ours)1201.040.01%
QHNet + ΔHSAP(ON)\Delta\mathbf{H} _ {\mathrm{SAP}} ^ {(\mathrm{ON})}, LON\mathcal{L} _ {\mathrm{ON}} (ours)1200.4440.0%
QHNet + ROCKET (ours)1200.1320.0%
QHNet + ROCKET (ours)1950.0540.0%
QHNet ΔH[PMINAO]\Delta\mathbf{H}[\mathbf{P} _ {\mathrm{MINAO}}] (reported)2800.484–
QHFlow ΔH[PMINAO]\Delta\mathbf{H}[\mathbf{P} _ {\mathrm{MINAO}}] (reported)2800.323–
QHFlow2-s (12M, reported)280+180.036–
QHFlow2-m (43M, reported)280+180.019–
QHFlow2-l (183M, reported)280+180.011–
QHFlow2-xl (990M, reported)280+180.007–

We did not train QHNetV2, since its code is not available.

A summary is given in [G1] of our global response above, and the full analysis in App. Q.

I thank the authors for their rebuttal and the additional experiments, which address several of my concerns.

I have a few follow-up comments. First, in Table 44, the new ablations suggest that the ON readout accounts for much of the improvement in Hamiltonian prediction. If I understand correctly, the theoretical analysis mainly motivates the choice of loss rather than explaining the benefit of predicting directly in the ON basis. Could the authors clarify this distinction? It seems odd to me that the readout itself plays a more important role than the current presentation suggests.

Second, I am still somewhat confused by the "flow map" terminology used for GROOTrot. When I first read this, I immediately related it to methods such as progressive distillation or MeanFlow. As I understand it, the network predicts a single orbital rotation. It is supervised directly on the final density through the matrix exponential. While this can be interpreted as a flow map mathematically, it seems closer to a regression model with a Grassmann-constrained readout. If the authors want to claim that this is a flow-map model, a baseline with vanilla flow matching seems necessary to me.

I am also somewhat confused by the idea of calling it a flow model. In a conventional flow model, we typically have a source distribution and a target distribution, with the marginal flow constructed from conditional flows. Here, however, the source is a deterministic MINAO density, and for each molecular geometry, there is only one converged target density. This makes the conditional-to-marginal flow construction essentially trivial when conditioned on the geometry. I think the formalism is closer to React-OT, where the transition-state are also deterministic given the reaction.

Finally, I have a question about the scalability experiments. I understand that the largest molecules are only used for evaluation, but I am curious how the model handles atom pairs beyond the 15 Å readout cutoff. Are the corresponding predicted matrix blocks simply set to zero, or are those pairs excluded from the decoder computation entirely? Since the matrix size grows quadratically and the Löwdin transformation is not strictly local, it would also be helpful to understand how memory scales for the largest molecules.

Reviewer_R78GOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This article deals with Machine-learned interatomic potentials (MLIPs) models, that supplement or replace Density Functional Theory (DFT) calculations, the standard quantum-mechanical method for computing the energy of a set of atoms. In this article, instead of predicting the energy directly, as an interatomic potential does, the network predicts the matrices that DFT itself solves for: the density matrix P or the Hamiltonian H. The authors call these Ground-State descriptor Models (GSMs). A ‘Fock build’ is a single evaluation of H(P), the most expensive operation in DFT. From the predicted P or H, one Fock build then yields the energy and most other properties, without the repeated iterations of a full DFT calculation. The authors train a network from scratch. For a Hamiltonian model, the network predicts H, and diagonalising it gives P. The Fock build is then evaluated at that P.

  • The DFT equation is HC = SC diag(\epsilon), a generalised eigenproblem. S is the Gram matrix of a non-orthogonal dictionary (the atomic orbitals). H depends on P, so solving it is a fixed-point iteration, the “SCF”, taking about 10–15 cycles.
  • The density matrix is, in an orthonormal basis, the projector onto the O lowest eigenvectors. Only the subspace matters, so P is a point on a Grassmannian.

The paper offers three contributions:

NameIdeaInterpretation
ON-LossMeasure errors on S1/2PS1/2S^{1/2} P S^{1/2} or S−1/2HS−1/2S^{-1/2} H S^{-1/2}A change of metric: error in function space rather than on the coefficients of a redundant frame
GROOTForce the predicted PP to be a valid projectorEither project by eigendecomposition and backpropagate through it (GROOTprojGROOT_{proj}), or rotate an initial guess: exp⁡(X) P0exp⁡(−X)\exp(X)\, P_0 \exp(-X) (GROOTrotGROOT_{rot})
ROCKETDo not penalise changes to HH that leave PP unchangedRemove part of the null space of H↦PH \mapsto P from the loss

They add

  • a one-shot energy correction (Harris–Foulkes);
  • a self-consistency test to reject bad predictions; and
  • label-free fine-tuning. Errors are in mHa; chemical accuracy, 1 kcal/mol, is 1.6 mHa.

The paper's headline results are as follows:

  • QM9 → QM40 (molecules up to 4× larger): energy −79% and force −83% against the density baseline; −99.8% and −96% against the best Hamiltonian baseline.
  • QMugs (trained up to 150 electrons, tested up to 660): 0.07 mHa with under 0.4% of predictions rejected. Against interatomic potentials: energies 5–14× better, but potentials trained on forces still win on forces.
Critical strengths

The ideas are sound and the size-extrapolation benchmark is hard. All arms share one backbone and have large hyperparameter sweeps. Most of the improvement holds up in threshold metrics, not just means. I checked Eqs. 39–43 and 45 and Propositions 1–4, and they are correct ! (hard work)

Critical weaknesses
  1. Eq. 12 is not shift-invariant as written. This is the only concrete mathematical problem I found.
  • The mask (Eq. 11) sends ΔH + aI to ΔH_masked + aQ, where Q is a rank-q projector. Minimising over c with cI then leaves a penalty proportional to ∣a∣|a|, so a pure shift is penalised.
  • The consistent form is minc∥mask(ΔH)−cQ∥min_c \|\text{mask}(ΔH) − cQ\|.
  • Numerically (B = 12, q = 5), the written loss goes from 0.532 to 1.588 as a goes from 0 to 10, while the corrected one stays at 0.527. It may only be loose notation, and the code is not given. It does contradicts the “gauge-free” label in Table 22.
  1. The mean error is dominated by a minority of molecules with enormous errors, so a reduction like −99.9% tells little about the typical molecule.

Eg Harris–Foulkes energy errors from Table 28.

ArmMolecules keptMean on kept (mHa)Mean on all (mHa)
Hamiltonian, NON loss70.4%0.053655
ROCKET99.8%0.4450.557

In addition:

a- The comparison on kept molecules isn’t fair. The NON arm keeps the molecules that passed its own filter, probably the smaller and easier ones. ROCKET might do just as well on those, so this suggests the pattern but doesn’t prove it.

b- However fixing the large failures is a real achievement. A model that is occasionally wrong by 2000 mHa is unusable, so removing that failure may be the paper’s most valuable result. So essentially, I object to the presentation but not the result.

3- The theory is weaker than suggested.

  • The intro says “show mathematically” and related work says “conclusive evidence”, but Appendix C.1 says the bounds “do not establish” that the ON bound is tighter. The bounds are local and per-molecule, so they say nothing about size extrapolation, and Table 1 has no force row.
  • Fig. 2 uses first-order energy sensitivity, whereas the reported energy is second-order.
  • The correlations in Table 24 may be driven by molecule size and by the two-cluster shape of the errors.

4- There may be a better loss, suggested by the authors' own analysis:

Appendix C.5 shows the first-order density error depends only on the occupied–virtual block of ΔH, weighted by inverse gaps. Why not use that loss directly?

5- The cost-accuracy “Pareto” claim isn’t actually shown.

  • Appendix K, “Training Costs”, is empty, and there are no inference times for the interatomic potentials.
  • Reading Fig. 20 at 660 electrons: full SCF about 17 s, one Fock build about 1.1 s, forward pass about 0.3 s. So a one-shot GSM is only about 10× faster than full DFT, with setup time excluded.
  • Forces need an extra nuclear gradient (Table 3), despite Fig. 1’s “exactly one Fock build”.

6- Some weak results stays in the appendix:

From QM9 to QMugs (Table 27), 61–100% of predictions are rejected for every arm. GROOT_proj collapses (mean energy error 26,318 mHa) while GROOT_rot stays at 36.9. On 3BPA, the best energies come from P_base (ON), not GROOT.

7- There is no chemistry validation

  • “Chemical accuracy” here means agreement with B3LYP/def2-SVP. This is a small basis, not with experiment, and on absolute energies rather than energy differences.
  • There is no molecular dynamics or geometry optimisation. Forces at an unconverged density aren’t the gradient of the reported energy.
Questions for authors
  1. Is the mask in Eq. 12 applied before or after removing the shift? As written the loss is not invariant to H + aI.
  2. What explains the abrupt NON failure, and does ∥dHtilde∥2/gap\|dH_{tilde}\|_2 / gap predict it?
  3. Why not a gap-weighted OV Hamiltonian loss, as C.5 suggests?
  4. Would it not make sense to train the eSCN backbone as an MLIP, tune MLIPs, and give training and inference costs. Then please fill App. K and show the actual cost-accuracy plot, including setup and gradient costs.
  5. What are the criterion and "hard" definition used for model selection? Which WALoss flavour?
  6. Can you show risk-coverage curves for the rejection rule, and explain the 75% rejection of P_base (ON)?
  7. Can you define the density SCT objective and its Fock-build cost compared with labelling?
  8. Why does GROOT_proj fail so badly on QM9 -> QMugs while GROOT_rot does not?
  9. Since a few very large errors dominate the means, can you report medians and quantiles of the error per arm, alongside the percentage reductions, and compare the arms on one common set of molecules that none rejects?
5 comments

Thanks for your thorough review of our manuscript, and checking the math in detail. And we are glad to hear that you appreciate both then ansatz and challenging evaluations. Before addressing your specific questions, we kindly refer your to our new additional experiments in our general response targeting

  • G1. Stresstesting our ansatzes on a different architectures with a standard dataset: QH9 /QHnet
  • G2. Data-Scaling on QMugs, and extension to even larger molecules (new custom ood validation dataset QMugsMax)
  • G3. Broader Chemical Space on SpiceV2
  • G4. Open-Shell (UKS) chemistry (new custom carben based dataset QMugsSpin)

W1 & Q1: ROCKET 3rd-Stage Subspace Gauge Notation (Eq. 12)

Thank you for catching this typo in Eq. 12. Our code already uses the shift-invariant subspace form, and only the shorthand LROCKET(2nd stage)(ΔH~O+k)\mathcal{L} ^ {\mathrm{(2nd\ stage)}} _ {\mathrm{ROCKET}}(\Delta\widetilde{\mathbf{H}} _ {O+k}) in the text was wrong. We updated Eq. 12 accordingly.

W2 & Q9: Outlier Driven Errors and Their Representation

We agree with that point and plotted the energy error distribution using Violin plots and marking median and quartiles of them in the hope to better communicate the overall error distribution. While the errors are indeed outlier driven, our models seem to shift the overall distribution towards lower errors.

can you report medians and quantiles of the error per arm, alongside the percentage reductions?

We observe that close to ID the Hamiltonian baselines are quiet strong on a quartile base, but as the models extrapolate futher OOD ROCKET starts to pick up, while GROOT_proj consistenly outperforms all density and hamitonian models.

QMugs OOD test set with N_e ≤ 300 and no prediction excluded. Best is bold and second best is underlined per quartile.

ArmQ1Red. Q1MedianRed. medianQ3Red. Q3
P_base (NON)0.0966–0.230–0.482–
P_base (ON)0.037361.4 %0.064871.8 %0.10378.6 %
GROOT_rot0.0018698.1 %0.00389‾\underline{0.00389}98.3 %0.00703‾\underline{0.00703}98.5 %
GROOT_proj0.00050599.5 %0.0012999.4 %0.0034599.3 %
H_base + c_t S (NON)0.00175‾\underline{0.00175}–0.00423–0.0112–
H_base + c_t S (ON)0.00343−95.5 %0.00913−116 %0.0241−116 %
ROCKET 3rd stage0.00218−24.1 %0.00515−21.8 %0.0113−0.9 %

QMugs OOD test set with N_e > 300 and no prediction excluded. Best is bold and second best is underlined per quartile.

ArmQ1Red. Q1MedianRed. medianQ3Red. Q3
P_base (NON)1.15–5.72–20.4–
P_base (ON)0.26077.4 %0.39793.1 %0.59297.1 %
GROOT_rot0.00919‾\underline{0.00919}99.2 %0.0162‾\underline{0.0162}99.7 %0.0278‾\underline{0.0278}99.9 %
GROOT_proj0.0037199.7 %0.0084799.9 %0.018499.9 %
H_base + c_t S (NON)0.0300–0.130–225–
H_base + c_t S (ON)0.0996−232 %0.392−201 %2.1299.1 %
ROCKET 3rd stage0.0453−50.9 %0.138−6.1 %0.37699.8 %

QMugsMax, entire set up to 300 heavy atoms (ChEMBL parents beyond QMugs' 100-heavy-atom cap, 101–300 heavy atoms, 770–2,340 electrons, 78 molecules; QMugs-trained on ≤ 150 electrons)

ArmQ1Red. Q1MedianRed. medianQ3Red. Q3
P_base (NON)721–2,409–12,005–
P_base (ON)0.97899.9 %2.4599.9 %7.7999.9 %
GROOT_rot0.170‾\underline{0.170}100.0 %0.591‾\underline{0.591}100.0 %1.57‾\underline{1.57}100.0 %
GROOT_proj0.0556100.0 %0.266100.0 %0.753100.0 %
H_base + c_t S (NON)18,444–45,992–124,616–
H_base + c_t S (ON)64.299.7 %35599.2 %11,36490.9 %
ROCKET 3rd stage3.63100.0 %6.65100.0 %17.8100.0 %

W3: Theoretical Claims, Sensitivity, and Correlations

We agree on all three points and have revised the paper accordingly.

  • Wording and scope of the theory (Sec. 1, Sec. 3, App. C.1, Table 1). The introduction now reads "we motivate mathematically and show empirically", and the related work speaks of "empirical evidence". The bounds in Table 1 are local, per-molecule, worst-case statements, and App. C.1 says so explicitly. They motivate the ON basis; the evidence for size extrapolation is empirical ([G1]–[G4]). Table 1 has no force row, so forces are likewise covered empirically: their sensitivity is now part of Fig. 2a,b (Eq. 64), and their errors are reported in Fig. 2g and Fig. 16.

  • First- versus second-order energy in Fig. 2 (App. D.4, Eqs. 65–66). You are right that the reported energy is second-order in the matrix error. Fig. 2a,b now shows the second-order sensitivity of the Harris–Foulkes energy, which is the energy we report, in place of the first-order response of the standard Kohn–Sham energy. The conclusion is unchanged: under LNON\mathcal{L} _ {\mathrm{NON}} the sensitivity falls with the overlap eigenvalues, and under LON\mathcal{L} _ {\mathrm{ON}} it stays approximately constant for all four observables (density, dipole, force, energy).

  • Correlations in Table 19 (formerly Table 24; App. M.1, Figs. 13–16). We agree that both effects enter pooled statistics, so the table and figures now report rr and II under two joint controls. First, each model keeps only its predictions with an energy error ∣ΔE∣<30|\Delta E| < 30 mHa. This removes the failed cluster of the two Hamiltonian models trained with LNON\mathcal{L} _ {\mathrm{NON}} (36–39% of their predictions, drawn in black in Figs. 13–16). Second, log⁡10Nelec\log _ {10} N _ {\mathrm{elec}} is regressed out of both quantities, so rr is the partial correlation at fixed molecule size. Under these controls the ON residual has the higher rr in all 24 panels and the higher II in 23 of 24.

W4 & Q3: Better Hamiltonian Loss?

There may be a better loss, suggested by the authors' own analysis: Appendix C.5 shows the first-order density error depends only on the occupied–virtual block of ΔH, weighted by inverse gaps. Why not use that loss directly?

That is a good point, and we tried it out:

QM40 test set (1,000 molecules; forces on 200). Energy: |ΔE| [mHa]; forces: per-atom force-error norm [mHa/Å]; Excl.: share rejected by the self-consistency rule.

ObjectiveExcl.E meanE Q1E medianE Q3E maxF meanF medianF Q3
Gap-weighted OV (App. C.5), lr 1e-3 (selected)2.6 %2.530.1510.6382.4059.75.294.237.76
Gap-weighted OV (App. C.5), lr 3e-46.3 %5.230.3041.254.811077.735.8311.2
ROCKET, 2nd stage0.1 %0.325‾\underline{0.325}0.0293‾\underline{0.0293}0.121‾\underline{0.121}0.369‾\underline{0.369}14.1‾\underline{14.1}1.94‾\underline{1.94}1.60‾\underline{1.60}2.51‾\underline{2.51}
ROCKET, 3rd stage (ours)0.0 %0.3160.02730.08890.28110.11.731.422.21
ROCKET, 3rd stage, 3 seeds (Table 26)0.03 %0.3770.02810.09380.29826.01.811.412.29

We rationalize these results by noting that this expansion stays accurate only while ΔH is small; once ΔH grows on larger molecules, unconstrained gaps and occupation ordering can drift (so maybe one would need a term that first gets us this close so the first order can further thighten it). ROCKET's objective constrains the occupied subspace and the LUMO as well.

W5 & Q4: Cost-Accuracy Pareto Plots and additional "BackBone-ForceField (BBFF)" MLIP baseline

We added a runtime study in Appendix N (Figs. 17–21, [G5]), and clarified in Fig. 1's caption that FKS\mathbf{F} _ {\mathrm{KS}} adds one nuclear gradient (Table 3):

  • Training & MLIP Pareto front (Figs. 17, 21): ON-Loss, GROOT, and ROCKET add no training overhead over NON baselines. Including setup, MLIP inference (NequIP, MACE, BBFF), and fallback SCF on rejected molecules, GROOT sits on the Pareto front between MLIPs and DFT for energies and energy-only forces (with ROCKET close behind), while E+FE+F MLIPs lead on forces.
  • End-to-end speedups with setup & gradients (Figs. 19–20): Counting setup and gradients, one-shot GSMs are 3–5x faster than cold SCF for energies (~2x with forces) under RI-JK, and 9–11x faster (3.5–5x with forces) under direct J/K—where setup + model inference drop to ~1–2% of runtime on large molecules and fit on one GPU where RI-JK gradients OOM.
  • Projector RI-K speedup (Fig. 19): Because GROOT and ROCKET output rank-OO projectors, RI-K contracts over OO orbitals rather than the full B×BB \times B density, making their Fock build 1.6–5.4x faster than unconstrained Pbase\mathbf{P} _ {\mathrm{base}}.

W6: QM9 →\to QMugs and 3BPA

  • QM9 →\to QMugs (Table 40, formerly Table 27; Figs. 34–35): We placed this split in the appendix because training on 0 K QM9 minima and testing on 300 K QMugs conformers confounds 9x size extrapolation (≤74→660 e−\le 74 \to 660\thinspace e ^ -) with temperature/conformer transfer, whereas Figs. 1 and 3 isolate each regime cleanly. Even on this confounded benchmark, our methods outperform the NON baselines by orders of magnitude: NON baselines reject 100% of molecules (4,417–46,920 mHa EE MAE), whereas GROOTrot\text{GROOT} _ {\mathrm{rot}} (36.9 mHa all, 10.4 mHa kept) and ROCKET (86–152 mHa all, 15 mHa kept) lead every observable in their families, and self-rejection filters GROOTproj\text{GROOT} _ {\mathrm{proj}}'s OOD outliers down to 8.3 mHa kept.
  • 3BPA (Table 42, Section 5, p. 8): 3BPA (144 e−e ^ -) is in-distribution in size (≤150 e−\le 150\thinspace e ^ - train), and Section 5 already highlights our ON-Loss model Pbase\mathbf{P} _ {\mathrm{base}} (ON) here (NON rejects 100%). Moreover, Pbase\mathbf{P} _ {\mathrm{base}} (ON) leads only on Harris–Foulkes EE (0.006–0.008 mHa) while leaving 3.6–4.8 mHa error in EKSE _ {\mathrm{KS}}, whereas GROOTproj\text{GROOT} _ {\mathrm{proj}} reaches 0.01–0.03 mHa on both EE and EKSE _ {\mathrm{KS}}.

W7: DFT Reference Ceiling and Relative Energies

We agree that surrogates cannot exceed the accuracy of their DFT reference ("chemical accuracy" here means 1 kcal/mol vs. B3LYP/def2-SVP) and that one-shot FKS\boldsymbol{F} _ {\mathrm{KS}} is only conservative at self-consistency; we added both points and MD/geometry-optimization benchmarks explicitly to Limitations. Note that relative energies are already evaluated on 3BPA torsional profiles (per-molecule offset removed; Fig. 4, Table 42) and Transition1x reaction barriers (Fig. 4, Table 43).

Comments on the paper

7 comments
AuthorsOpenReview ↗

Presentation, Structure, and Project Context

We thank all reviewers for their thorough feedback. We are genuinely sorry to hear that our presentation made the manuscript difficult to read; as all three reviewers raised readability and structural concerns, we take this consensus seriously and will make presentation and clearer signposting a primary focus for a camera-ready revision.

Our overarching goal was to present our contributions under one unifying principle: aligning ground-state learning objectives and architectures with the governing equations of KS-DFT. To give brief context on why the paper introduces several methods and baselines at once rather than following a single-method storyline, this reflects how the research actually unfolded:

  1. We started from manifold-constrained density transport (GROOT_rot), and during training observed how strongly switching from the NON to the ON metric improved size extrapolation, which prompted our theoretical and empirical investigation of ON-Loss.
  2. Next, we needed a strong, fair Hamiltonian baseline for size extrapolation beyond QH9-ood (where size generalization had been over-estimated). Starting from QHFlow2, we eliminated its costly initial Fock build from PMINAO\mathbf{P} _ {\mathrm{MINAO}} to H(PMINAO)\mathbf{H}(\mathbf{P} _ {\mathrm{MINAO}}), at which point training without Δ\Delta-learning failed completely on larger molecules. Moving to the cheap HSAP\mathbf{H} _ {\mathrm{SAP}} guess helped, yet still lagged far behind density models. Noticing that a training-set gauge shift (ctSc _ t \mathbf{S}) helped on top of Δ\Delta-learning as data normalization led us to test removing per-molecule gauge shifts entirely, and later masking unsupervised virtual-block errors (O+kO+k). By that stage, our pursuit of a competitive baseline had departed far enough from prior work to become a method in its own right (ROCKET).

Because we wanted to report this complete empirical progression faithfully—prioritizing scientific completeness and letting the data speak for itself over trimming results into a neat linear narrative—fitting all of this content into the nine-page limit forced heavy compression in the main text (dense figures, merged paragraphs, and inline emphasis in place of section breaks). With the extra space available in a camera-ready version, we will focus on decompressing the main text, improving signposting, and removing distracting inline formatting.

AuthorsOpenReview ↗

General Response: Overview of New Experiments and Datasets (`[G1]`–`[G5]`)

Thanks to all reviewers for their feedback. In the revised manuscript (changes in blue), we added five new experiments and two novel datasets (QMugsMax and QMugsSpin). Full tables and protocols are in comments [G1]–[G5]:

  • [G1] QHNet + AdamW on QH9-stable-ood (App. Q, Table 22, Fig. 26): Training 20M QHNet from geometry alone with its published AdamW recipe, ROCKET reaches 0.054 mHa EKSE _ {\mathrm{KS}} MAE at 1 Fock build (vs. 12.5 mHa for published QHNet, 0% self-rejected), improving on 2-Fock-build QHNet and QHFlow (0.32–0.48 mHa).
  • [G2] Training-size scaling & QMugsMax up to 319 heavy atoms (App. T, Figs. 30–31): Doubling the QMugs training cutoff (Ne≤150→≤300N _ e \le 150 \to \le 300) at matched step count only shifts baseline breakdown to larger sizes; on QMugsMax (64 ChEMBL 37 molecules, 101–319 heavy atoms), LNON\mathcal{L} _ {\mathrm{NON}} baselines still fail (44–100% rejected), while GROOT stays near chemical accuracy across the full range (1.3–1.7 mHa, i.e. 0.8–1.1 kcal/mol, vs. 12–20 mHa for MLIPs) and ROCKET reaches 4.2 mHa kept EKSE _ {\mathrm{KS}} MAE (<5% rejected).
  • [G3] 11-element chemistry on SPICEv2 (App. O, Table 20, Fig. 22): Adding B,Si,Br\mathrm{B, Si, Br} (35,774 train structures ≤220 e−\le 220\,e ^ -, 1,409 size-OOD test structures up to 502 e−502\,e ^ -), GROOTproj\text{GROOT} _ {\mathrm{proj}} and ROCKET reach 0.25–0.26 mHa energy MAE across all structures without self-rejection and cut SCF cycles by ~36%.
  • [G4] Open-shell UKS on QMugsSpin (App. P, Table 21, Figs. 23–25): Fine-tuning UKS extensions on neutral CHNOF carbenes (≤22\le 22 heavy atoms) extrapolates to 100 heavy atoms at 0.13–0.27 mHa energy MAE (<0.5% rejected) for GROOTproj\text{GROOT} _ {\mathrm{proj}} and ROCKET, and ROCKET transfers zero-shot to carbenes with held-out P,S,Cl\mathrm{P, S, Cl} and formal charges (6.7 mHa EKSE _ {\mathrm{KS}} MAE).
  • [G5] Wall-time benchmarks & Pareto front (App. N, Figs. 17–21): Single-H200 timings up to 182 heavy atoms (1,380 e−e ^ -) show negligible training overhead, ~30% faster warm-started SCF, and 3.3–4.7x one-shot energy speedups with RI-JK (8.8–10.9x with direct J/K), placing GROOT and ROCKET on the energy Pareto front between MLIPs and KS-DFT.

Upon publication, we plan to release the relaxed 3D geometries and label-generation scripts for QMugsSpin and QMugsMax (Reproducibility Statement; App. I, Tables 10 & 12). Individual replies cover the matched-backbone BBFF control (Fig. 1; App. N.4, T), readout vs. supervision frames (App. R), self-rejection threshold transfer (App. S), per-element node normalization (App. H), and Eq. 12.

AuthorsOpenReview ↗

[G1] Architecture- and Optimizer-Agnostic Validation on `QH9-stable-ood` (`QHNet` + AdamW)

Manuscript references: Appendix I (Table 11); Appendix Q (Table 22, Fig. 26); Appendix U (Tables 35, 44).

To confirm that our gains generalize beyond the eSCNMD backbone or Muon optimizer, we trained the standard 20M-parameter QHNet from molecular geometry alone on QH9-stable-ood (≤20\le 20 atoms in training, 9,335 test molecules with ≥23\ge 23 atoms) using QH9's published 95-epoch AdamW recipe. For 20-epoch ablations, each arm selects its peak learning rate by its own validation objective (Table 35).

  • 20-epoch readout and loss ablations (Fig. 26, Table 44). Predicting ΔHSAP\Delta\mathbf{H} _ {\mathrm{SAP}} in the atomic-orbital (NON) frame breaks down out of distribution under LNON\mathcal{L} _ {\mathrm{NON}} (55.8 mHa EKSE _ {\mathrm{KS}}), WALoss (Table 35), and LON\mathcal{L} _ {\mathrm{ON}} (65.1 mHa). Predicting in the Löwdin (ON) frame removes nearly all size-OOD outliers even under LNON\mathcal{L} _ {\mathrm{NON}} (1.04 mHa), and adding LON\mathcal{L} _ {\mathrm{ON}} (0.444 mHa) and ROCKET (0.132 mHa) lowers errors further with 0% self-rejection.
  • Full 95-epoch comparison & Fock-build cost (Table 22, Table 44). Because HSAP\mathbf{H} _ {\mathrm{SAP}} is a one-electron matrix, evaluating observables requires 1 Fock build total. At 95 epochs, single-pass 20M QHNet + ROCKET reaches 0.054 mHa EKSE _ {\mathrm{KS}} MAE (0.0024 mHa Harris–Foulkes EE MAE, 0% self-rejected)—lowering error relative to published 1-Fock-build QHNet (12.5 mHa) and outperforming 2-Fock-build ΔH[PMINAO]\Delta\mathbf{H}[\mathbf{P} _ {\mathrm{MINAO}}] QHNet (0.484 mHa) and QHFlow (0.323 mHa) at half the Fock-build cost (while recent multi-step flow-matching QHFlow2 models on a larger backbone reach 0.007–0.036 mHa using 2 Fock builds plus 18-epoch fine-tuning):
QH9-stable-ood (9,335 test mol., no self-rejection; Table 22, Fig. 26)Total Fock buildsEpochsEKSE _ {\mathrm{KS}} MAE [mHa]Self-rejected
QHNet Habs(NON)\mathbf{H} _ {\mathrm{abs}} ^ {(\mathrm{NON})} (published / 80-ep. reported)165–80 / 8012.5 / 15.76.8% / –
QHNet Habs(NON)\mathbf{H} _ {\mathrm{abs}} ^ {(\mathrm{NON})} (retrained, full recipe)19578.710.8%
QHNet + ΔHSAP(ON)\Delta\mathbf{H} _ {\mathrm{SAP}} ^ {(\mathrm{ON})}, LNON\mathcal{L} _ {\mathrm{NON}} / LON\mathcal{L} _ {\mathrm{ON}} (ours)1201.04 / 0.4440.01% / 0.0%
QHNet + ROCKET (20M, ours)120 / 950.132 / 0.0540.0%
QHNet / QHFlow ΔH[PMINAO]\Delta\mathbf{H}[\mathbf{P} _ {\mathrm{MINAO}}] (reported)2800.484 / 0.323–
QHFlow2-s to -xl (12M–990M, flow ODE + fine-tune, reported)280+180.036–0.007–

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.