acceptodds
Under review as a conference paper at ICLR 2027

Electron-Density-Informed Molecular Property Prediction Using Surrogate Functionals

Abstract

Machine learning methods are increasingly used to predict chemical properties directly from the molecular structure to bypass the often prohibitive cost of quantum mechanical calculations. Recent work has shown that including the ground-state electron density as input to the machine learning model can improve accuracy in this setting. This work employs the recent surrogate functional framework originally developed for orbital-free density functional theory and extends it to predicting molecular properties. We show that passing the ground-state density as input to the model, while simultaneously enforcing it to be the minimum of the surrogate functional, surpasses state-of-the-art performance for six out of seven extensive properties of the QM9 dataset. Additionally, the approach shows improved generalization to larger molecules taken from the QM40 dataset. On all extensive properties of both datasets, our approach outperforms models that are informed by the ground-state density but do not enforce the surrogate functional constraints.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Electron Density, Molecular Properties, Surrogate functionals, Graph Neural Network, Equivariance, Molecules, Density Functional Theory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xwgvOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper extends surrogate functionals, a framework developed for machine-learned orbital-free DFT, from the electronic energy to arbitrary scalar molecular properties. The motivation is the Hohenberg–Kohn existence argument: since the ground-state density determines all ground-state properties, each property is in principle a functional of the density. For each target property, the authors train a dedicated network using the same pipeline as the energy surrogate, with the ground-state energy label replaced by the property label. On QM9, the surrogate models outperform geometry-only and density-input versions of the same architecture on all seven extensive properties, beating the prior state of the art on six. They also extrapolate substantially better to larger QM40 molecules.

Critical strengths

A clean controlled comparison. The three arms (geometry-only, density-input, surrogate) share architecture, training settings, splits, and at inference the same predicted density. The only difference is the training procedure. So the central finding is clear: surrogate training improves the model without giving it any extra information at inference.

Strong in-distribution results. On QM9, U0 error drops from 4.9 meV (density input) to 2.8 meV, below GotenNet-L (3.37 meV), with similar gains on U, H, G, ZPVE and Cv. Meaningful improvements on a heavily optimized benchmark, though from single runs.

Practical inference cost. Replacing ~100 minimization steps with two forward passes makes inference comparable to a standard model. Training cost is likely higher but not reported.

Size extrapolation. Per-atom U0 error stays roughly flat with molecule size for the surrogate, while it grows for all other models (Figure 4). This extends the earlier result of Wagner et al. for energies to other extensive properties.

The paper is well written and is easy to read and understand.

Critical weaknesses

The physical motivation doesn’t transfer. Only the energy obeys a variational principle. For other properties, Hohenberg–Kohn guarantees that a functional of ρ* exists, not that it is minimized at ρ*. A minimum can be imposed on any property (S[ρ] = Q* + ‖ρ − ρ*‖² is enough), but then the value at the minimum is just a learned function of the molecule, and information about the property still comes only from ground-state labels. Since the optimization loop is removed at inference, the model is effectively a density-input model trained with an extra regularizer (a hypothesis). The paper would be stronger framed this way.

Calibration is borrowed from the energy. The energy-bound loss was motivated by real Kohn–Sham energies rising quadratically around ρ*. No such data exists for other properties; the energy-calibrated bounds and GDI parameters are simply transferred by rescaling labels. For intensive properties there is also a conceptual mismatch: ‖p − p*‖² grows with molecule size, and the fixed GDI step needs per-coefficient gradients of order one, while the mean-pooled readout (Appendix A.4.1) makes each atom’s gradient scale as 1/N. No clear cost shows up in the results (on QM40 the surrogate is best for HOMO and LUMO), but the mismatch is not discussed.

The benefit is property-specific, and regressions are not discussed. For properties that are essentially linear in the density, the surrogate is clearly worse than the density-input model on QM9: µ 12 vs 7.9 mD, ⟨R²⟩ 480 vs 140 a₀². Table 2 reports this, but the text does not address it. A possible explanation is that gains follow energy-likeness rather than extensivity: all improved properties are energies or second energy derivatives (ZPVE, Cv, α). In QM9 the two groups almost coincide, so the claim “beneficial for all extensive properties” cannot be separated from this alternative. QM40 is mixed here (⟨R²⟩ is best with the surrogate).

No mechanistic analysis. The training constraints improve a model with identical inputs at inference, but nothing explains why. Possible reasons: the cache acting as constrained data augmentation, the stationarity and curvature constraints forcing a more structured use of the density, or the extensive bound suppressing size-dependent drift. Also, density input alone helps only a little (5.3 → 4.9 meV on U0 with beSEN) and not at all with eSEN (Table 3). The density seems to become useful mainly with the surrogate constraints, which deserves explanation.

Ablations are hard to interpret. In Table 6, results range from 2.8 meV (full setup) to 5.7 meV for one partial configuration, worse than geometry-only. The benefit seems to depend on specific combinations of components, with little discussion of which interactions matter.

Some QM40 baselines look anomalous. GotenNet’s HOMO errors (6300 and 4700 meV) are about ten times its LUMO errors (540, 520) and beSEN’s HOMO error (300 meV). Geometry-only and density-input beSEN have identical ZPVE errors (150 meV) with the same per-atom drift (~3 meV/atom, Figure 5), for a property that is roughly a sum of local contributions. If this is a systematic readout bias that the surrogate constraints suppress, part of the large QM40 margins (ZPVE 4.9 vs 150, Cv 170 vs 3100) comes from fixing a specific failure, not from better generalization in general. The QM9 comparison is the cleaner evidence.

Limited reproducibility. There is no reproducibility statement and no link to code or data. This matters because the density labels are new: they were computed by the authors at B3LYP/6-31G(2df,p) and fitted to an even-tempered basis (Appendix A.3), and small differences there may affect results. The QM40 test set cannot be reconstructed: binning and filters are given, but not how molecules were selected within each bin, and no identifiers are listed. Results are single runs without seeds or error bars, so it is unclear which differences in Table 2 are meaningful.

Unsupported claim. Section 3.6 says variational minimization and the direct density shortcut give “comparable performance,” but no results are shown.

The weaknesses that matter most for my decision come down to three points. First, a method that is well motivated for the energy is transferred to other properties without a physical reason why it should work there, and without another explanation of why it works in practice. Second, several issues with the results (the anomalous QM40 baselines, the missing seed variance, and the regressions on µ and ⟨R²⟩) leave me not fully convinced by the reported gains. Third, no code or data is provided for reproduction, which is especially relevant here because the density labels were computed by the authors and the QM40 test set cannot be reconstructed from the description. I would be happy to update my recommendation if the authors provide an explanation of why the approach works, and with it a clearer picture of its limitations, along with answers to the questions above.

Questions for authors
  1. Did you train a density-input model on perturbed densities (similar to the cache) but without the surrogate losses? If it reaches similar accuracy, the gain comes from seeing more densities, not from the variational structure.

  2. How were the GotenNet checkpoints evaluated (units, atomic references)? HOMO errors of 6300 and 4700 meV look like a mismatch. And why do geometry-only and density-input beSEN have identical ZPVE errors on QM40?

  3. How large is the run-to-run variance for the results in Tables 2 and 3? Several differences (e.g., surrogate vs. GotenNet-L on U0, 2.8 vs 3.37 meV) may be within noise.

  4. Why does the surrogate degrade µ and ⟨R²⟩ on QM9 (12 vs 7.9 mD, 480 vs 140 a₀²)? Do you have results on any extensive property that is not energy-derived, to test whether the benefit follows extensivity or energy-likeness?

10 comments
AuthorsOpenReview ↗

Answer to Reviewer xwgv

We thank the reviewer for the detailed and constructive review. In the revision, all results are now reported as mean and standard deviation over three seeds, we added control experiments for the data-augmentation explanation, a new auxiliary surrogate variant that resolves the regressions on the non-extensive properties, an analysis of why the approach works, and results on a second level of theory (PBE). A summary of all new results can be found in the general response.


Q1. Density-input model trained on perturbed densities, without the surrogate losses

"Did you train a density-input model on perturbed densities (similar to the cache) but without the surrogate losses? If it reaches similar accuracy, the gain comes from seeing more densities, not from the variational structure."

We trained the density-input model on perturbed densities with the ground-state label and without any surrogate loss, with three seeds each. As perturbed densities we use the MINAO initial guess, which replaces the ground-state density as input for half of each batch.

This control does not improve over the density-input model: U0 on QM9 is 4.15 ± 0.05 meV, compared to 4.14 ± 0.06 meV for the density-input model and 2.18 ± 0.09 meV for the surrogate model. On QM40 it is even worse than the density-input model (190 ± 10 meV compared to 150 ± 10 meV), and HOMO does not improve either. We also tested weight decay as a conventional regularizer, which does not reach the surrogate either. The gain of the surrogate model therefore does not just come from seeing more densities. The results are shown in the table below and in Appendix A.6.

Table. Mean absolute error of U0 and HOMO on the QM9 test split and on QM40, mean and standard deviation over three seeds. All models use the predicted ground-state density as input. The controls are density-input models trained on perturbed densities or with weight decay, without any surrogate loss. Best value per column in bold.

ArmU0, QM9 (meV)U0, QM40 (meV)HOMO, QM9 (meV)HOMO, QM40 (meV)
Density input4.14 ± 0.06150 ± 1013.7 ± 0.1310 ± 20
+ perturbed densities4.15 ± 0.05190 ± 1013.8 ± 0.2320 ± 30
+ weight decay 1e-24.01 ± 0.08170 ± 3013.55 ± 0.05310 ± 30
+ weight decay 1e-16.5 ± 0.4102 ± 115.41 ± 0.08300 ± 10
Surrogate2.18 ± 0.0968 ± 414.0 ± 0.3220 ± 20
Auxiliary surrogate2.51 ± 0.0263 ± 211.1 ± 0.6280 ± 30

Q2. QM40 baselines: GotenNet evaluation, identical ZPVE errors

"Some QM40 baselines look anomalous. GotenNet’s HOMO errors (6300 and 4700 meV) are about ten times its LUMO errors (540, 520) and beSEN’s HOMO error (300 meV). Geometry-only and density-input beSEN have identical ZPVE errors (150 meV) with the same per-atom drift (~3 meV/atom, Figure 5), for a property that is roughly a sum of local contributions. If this is a systematic readout bias that the surrogate constraints suppress, part of the large QM40 margins (ZPVE 4.9 vs 150, Cv 170 vs 3100) comes from fixing a specific failure, not from better generalization in general. The QM9 comparison is the cleaner evidence."

"How were the GotenNet checkpoints evaluated (units, atomic references)? HOMO errors of 6300 and 4700 meV look like a mismatch. And why do geometry-only and density-input beSEN have identical ZPVE errors on QM40?"

We evaluated the released GotenNet checkpoints with the code provided by the authors. To check our evaluation (units and atomic references), we evaluated the checkpoints on GotenNet's own QM9 test split and reproduce the published numbers. The large HOMO errors on QM40 are real and follow from GotenNet's readout.

GotenNet predicts both HOMO and LUMO as a sum of per-atom contributions. For an intensive property, the network therefore distributes the mean value of the property over the atoms, which no longer adds up for the larger molecules of QM40: the HOMO errors grow linearly with the number of heavy atoms, whereas for LUMO, whose mean value is close to zero, this drift is much smaller.

Regarding the identical ZPVE errors: on QM40 the geometry-only and the density-input model share a size-dependent drift in their ZPVE predictions, which the surrogate models largely reduce. We agree that part of the QM40 margin for ZPVE and Cv comes from removing this drift. However, the surrogate also improves properties without such a drift, e.g. the HOMO energy on QM40, and it shows clear gains on QM9.

AuthorsOpenReview ↗

Answer to Reviewer xwgv (continued)

Q3. Run-to-run variance

"Results are single runs without seeds or error bars, so it is unclear which differences in Table 2 are meaningful."

"How large is the run-to-run variance for the results in Tables 2 and 3? Several differences (e.g., surrogate vs. GotenNet-L on U0, 2.8 vs 3.37 meV) may be within noise."

We retrained all models with three seeds and now report mean and standard deviation in almost all tables (Tables 2–6, 10, 11 and 13; the block ablation of the beSEN components in Appendix A.8 uses one seed per combination). The seed standard deviations are small compared to the differences we discuss. Differences that are within the seed variance are now marked as ties.

Table. MAE on the QM9 test split, mean and standard deviation over three seeds. All models that use a density take the predicted ground-state density as input. GotenNet-L is reproduced as published. The best value per row is bold, ties are bold together.

PropertyGeometry onlyDensity inputSurrogateAuxiliary surrogateGotenNet-L
U0 (meV)4.23 ± 0.094.14 ± 0.062.18 ± 0.092.51 ± 0.023.37
U (meV)4.20 ± 0.064.07 ± 0.022.12 ± 0.022.4 ± 0.13.41
H (meV)4.17 ± 0.044.07 ± 0.022.12 ± 0.062.4 ± 0.13.30
G (meV)5.52 ± 0.065.47 ± 0.033.6 ± 0.14.00 ± 0.064.98
ZPVE (meV)1.022 ± 0.0041.0152 ± 0.00070.903 ± 0.0070.901 ± 0.0021.08
α (m a0^3)33.9 ± 0.333.3 ± 0.623.1 ± 0.323.2 ± 0.228
Cv (mcal/mol/K)19.94 ± 0.1019.5 ± 0.116.1 ± 0.216.0 ± 0.319
HOMO (meV)13.7 ± 0.213.7 ± 0.114.0 ± 0.311.1 ± 0.613.4
LUMO (meV)13.5 ± 0.713.2 ± 0.513.7 ± 0.812.2 ± 0.612.2
gap (LUMO − HOMO) (meV)20.5 ± 0.520.1 ± 0.220.7 ± 0.717.6 ± 0.819.8
μ (mD)6.8 ± 0.56.52 ± 0.088.3 ± 0.55.2 ± 0.46.7
⟨R²⟩ (m a0^2)15 ± 113.1 ± 0.511.91 ± 0.1011.0 ± 0.524

Q4. µ and ⟨R²⟩ regressions; extensivity vs energy-likeness

"The benefit is property-specific, and regressions are not discussed. For properties that are essentially linear in the density, the surrogate is clearly worse than the density-input model on QM9: µ 12 vs 7.9 mD, ⟨R²⟩ 480 vs 140 a₀². Table 2 reports this, but the text does not address it. A possible explanation is that gains follow energy-likeness rather than extensivity: all improved properties are energies or second energy derivatives (ZPVE, Cv, α). In QM9 the two groups almost coincide, so the claim “beneficial for all extensive properties” cannot be separated from this alternative. QM40 is mixed here (⟨R²⟩ is best with the surrogate)."

"Why does the surrogate degrade µ and ⟨R²⟩ on QM9 (12 vs 7.9 mD, 480 vs 140 a₀²)? Do you have results on any extensive property that is not energy-derived, to test whether the benefit follows extensivity or energy-likeness?"

With three seeds and the updated training setup, the surrogate model no longer degrades ⟨R²⟩ (11.91 ± 0.10 compared to 13.1 ± 0.5 m a₀² for the density-input model), but it still degrades µ (8.3 ± 0.5 compared to 6.52 ± 0.08 mD). We attribute this degradation to surrogate losses contradicting the actual behavior of the property.

We therefore added the auxiliary surrogate model, where the surrogate is trained on a separate head of the same network, while the property head is trained only on the labels. The auxiliary surrogate resolves the regressions and gives the best results for µ (5.2 ± 0.4 mD), ⟨R²⟩ (11.0 ± 0.5 m a₀²), HOMO, LUMO and the gap, while it is on par with the surrogate for ZPVE, Cv and α (Table 2 and Section 4.1 of the revised manuscript).

Regarding extensivity vs. energy-likeness: QM9 contains no size-extensive property that is not derived from the energy, so the two hypotheses cannot be fully separated on this dataset. However, the auxiliary surrogate improves on all properties compared to the density-input model, also the non-energy properties. The benefit of the surrogate training therefore does not require the property to be energy-like, what fails for non-energy properties is placing the surrogate loss on the property output itself.

AuthorsOpenReview ↗

Answer to Reviewer xwgv (continued)

Q5. Physical motivation does not transfer; reframe as a regulariser

"The physical motivation doesn’t transfer. Only the energy obeys a variational principle. For other properties, Hohenberg–Kohn guarantees that a functional of ρ* exists, not that it is minimized at ρ*. A minimum can be imposed on any property (S[ρ] = Q* + ‖ρ − ρ*‖² is enough), but then the value at the minimum is just a learned function of the molecule, and information about the property still comes only from ground-state labels. Since the optimization loop is removed at inference, the model is effectively a density-input model trained with an extra regularizer (a hypothesis). The paper would be stronger framed this way."

We agree with the reviewer and adopted this framing in the revised manuscript (introduction, Section 3.3 and discussion in Section 5). For properties other than the energy, the surrogate does not follow from a variational principle, and we now describe it as a regularizer of the shared network. This view is supported by our new results: the auxiliary surrogate, which does not touch the property output at all, gives similar or larger improvements for most properties, and perturbed densities without the surrogate losses do not help (see Q1). The energy remains the case where the surrogate is physically motivated.


Q6. Calibration borrowed from the energy

"Calibration is borrowed from the energy. The energy-bound loss was motivated by real Kohn–Sham energies rising quadratically around ρ*. No such data exists for other properties; the energy-calibrated bounds and GDI parameters are simply transferred by rescaling labels. For intensive properties there is also a conceptual mismatch: ‖p − p*‖² grows with molecule size, and the fixed GDI step needs per-coefficient gradients of order one, while the mean-pooled readout (Appendix A.4.1) makes each atom’s gradient scale as 1/N. No clear cost shows up in the results (on QM40 the surrogate is best for HOMO and LUMO), but the mismatch is not discussed."

We agree that the calibration of the surrogate losses (the energy bound, the GDI step and the target scale) is taken from the energy and transferred to the other properties by rescaling the labels. Following the reframing in Q5, we view these settings as hyperparameters of the regularizer rather than as a physical model of the property away from the ground state.

Regarding intensive properties, we agree with the reviewer: with the mean-pooled readout the per-atom gradients decrease with molecule size, so the fixed GDI step and the bound act differently on small and large molecules. The auxiliary surrogate avoids this mismatch by construction: it uses no bound, applies GDI to a separate head, and trains the property head only on the labels, so no calibration taken from the energy is imposed on the intensive property. As shown in Q4, it gives the best results for the intensive properties on QM9.

Reviewer_4qk8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a model that predicts various properties of a given molecular structure via the electron density. While previous work uses the electron density only as an input, the authors additionally impose a constraint that the predicted value attains its minimum at the ground-state density. The proposed model, which is called beSEN, is based on eSEN with modifications to part of its architecture for improved performance. Experiments are conducted on the QM9 and QM40 datasets, using electron densities computed by the authors with PySCF. As a result, the proposed method improves prediction accuracy on several QM9 properties, with even larger improvements on QM40. However, the paper provides limited information for understanding the details of the model implementation and experiments.

Critical strengths
  • The improvement in property prediction accuracy is clear.
  • Since the proposed method predicts molecular properties through the electron density, a certain degree of interpretability could be expected.
Critical weaknesses
  • The manuscript requires substantial revision overall for the proposed method and evaluation to be reproducible, which makes it difficult to verify the reported results.
  • The experiments on QM9 and QM40 alone appear somewhat insufficient, and additional experiments on datasets from other material classes (e.g., crystalline materials) are needed. Moreover, the paper lacks in-depth analysis beyond quantitative results, as detailed below.
Questions for authors

I have the following questions and suggestions:

  • Even including the Appendix, the specific details needed to reproduce the method and experimental setup cannot be verified from the paper alone. Key details, such as how the electron density labels were generated, are not described in the main text, and the distinction between the surrogate and density-input models is also unclear. Could the authors supplement this information and clarify whether they plan to release the code?
  • The paper does not analyze why the surrogate model predicts better than the density-input model. Could the authors provide analysis or qualitative examples beyond the quantitative results? In addition, the ablation study in the Appendix is presented only as a table without interpretation. Could the authors discuss the contribution of each component?
  • The paper does not explain why each modification in beSEN is necessary, nor does it provide ablations. In particular, is there a specific reason for using the Gram matrix instead of the Gate activation?
  • Figure 3 only shows the number of parameters per module, which does not support the claim that the additional parameters incur little computational cost. Could the authors report the FLOPs of each module?
  • Could the authors discuss whether the proposed method can be extended to other material classes, such as crystalline materials?

Overall, the proposed method shows good results on several properties. However, even including the Appendix, the details of the method and experiments cannot be fully understood from the paper alone, which undermines the reliability of the reported results. Since this requires substantial revision, I recommend rejection in its current form.

7 comments
AuthorsOpenReview ↗

Answer to Reviewer 4qk8

We thank the reviewer for the careful review and the constructive suggestions. In the revision, we clarify the distinction between the surrogate and the density-input models and the label generation, and we will release the code. We added an analysis of why the surrogate outperforms the density-input model, including control experiments and a linear-probe analysis of the learned representations, and an ablation of the individual beSEN modifications. All results are now reported as mean and standard deviation over three seeds; a summary of all new results can be found in the general response.


Q1. Reproducibility: label generation, surrogate vs density input, code release

"The manuscript requires substantial revision overall for the proposed method and evaluation to be reproducible, which makes it difficult to verify the reported results."

"Even including the Appendix, the specific details needed to reproduce the method and experimental setup cannot be verified from the paper alone. Key details, such as how the electron density labels were generated, are not described in the main text, and the distinction between the surrogate and density-input models is also unclear. Could the authors supplement this information and clarify whether they plan to release the code?"

We agree that reproducibility is important and therefore plan to release the code upon acceptance. Details about the label generation process can be found in Appendix A.3. We also added additional clarification about the distinction between the surrogate and density-input models in Section 4.1, introducing the labels used in the result tables verbatim. Additionally, we made the following changes to the manuscript to ensure that every detail needed for reproducibility is provided:

  • The generation of the density labels (Kohn–Sham calculations at B3LYP/6-31G(2df,p) and the density fitting to the even-tempered basis) is now summarized in the main text, with all details in Appendix A.3.
  • The surrogate and the density-input models are now defined side by side in Section 4.1: both are the same network, receive the same directly predicted density as input and need one forward pass at inference; they differ only in the training procedure.
  • All hyperparameters of the models and of the surrogate losses are listed in the appendix.
  • The selection rule of our QM40 test set is provided.
  • A reproducibility statement has been added.

Q2. More datasets / other material classes (crystals)

"The experiments on QM9 and QM40 alone appear somewhat insufficient, and additional experiments on datasets from other material classes (e.g., crystalline materials) are needed."

"Could the authors discuss whether the proposed method can be extended to other material classes, such as crystalline materials?"

We do not see any reason why the proposed method would not be applicable to other classes of chemical systems, such as crystalline materials. The discussion of the manuscript (Section 5) notes this point. However, we do believe that the experiments on QM9 and QM40 are sufficient to demonstrate the effectiveness of the proposed method, and that the extension to other material classes is outside the scope of this work.

AuthorsOpenReview ↗

Answer to Reviewer 4qk8 (continued)

Q3. Why does the surrogate beat the density-input model?

"Moreover, the paper lacks in-depth analysis beyond quantitative results, as detailed below."

"The paper does not analyze why the surrogate model predicts better than the density-input model. Could the authors provide analysis or qualitative examples beyond the quantitative results?"

We believe that the surrogate mechanism acts as an effective regularizer for the model. This is supported by the fact that the auxiliary surrogate model, where the surrogate is trained as an auxiliary task on a separate head, also improves the property prediction performance of the model. This is the case even for non-extensive properties, where the surrogate method failed before (see also the general response).

Additionally, we provide a new analysis of the learned representation of each model type in Appendix A.7. Here we probe the internal representation of the models by evaluating the accuracy of predicting the ground state electron density by linear regression from internal features. The results are shown in the table below. The surrogate and auxiliary surrogate models keep the ground-state density information in their features throughout the network, while the density-input model seems to lose it with depth. This also explains why the density input alone helps only little: without a density task, the network discards most of the information in the density it is given, while the surrogate training forces it to keep this information.

Table. Linear probe of the ground-state density from the node features after each block of the U0 models (48 channels, 6 layers). The density-input and surrogate models are given the label ground-state density as input. The NMAE (in %) of the probed density, on the 1000 molecules of the QM9 test split is shown. The mean and standard deviation over three seeds are reported. Best value per column in bold.

Model (NMAE, %)block 1block 2block 3block 4block 5block 6 (final)
Geometry only (no density input)7.8 ± 1.39.8 ± 1.512 ± 115 ± 217 ± 222 ± 3
Density input1.8 ± 0.23.2 ± 0.25.5 ± 0.68.1 ± 1.611 ± 216.5 ± 0.8
Surrogate0.087 ± 0.0430.48 ± 0.050.88 ± 0.061.5 ± 0.23.9 ± 1.47.2 ± 4.5
Auxiliary surrogate0.12 ± 0.050.71 ± 0.101.4 ± 0.12.23 ± 0.093.1 ± 0.24.0 ± 0.5
AuthorsOpenReview ↗

Answer to Reviewer 4qk8 (continued)

Q4. Interpretation of the ablation

"In addition, the ablation study in the Appendix is presented only as a table without interpretation. Could the authors discuss the contribution of each component?"

Most of the gain seems to come from the gradient-descent-improvement (GDI) loss: GDI alone without the other surrogate losses reaches 2.37 ± 0.05 meV and the full surrogate 2.18 ± 0.09 meV on U0, compared to 4.14 ± 0.06 meV for the density-input model. The energy bound together with the cache gives a smaller improvement (3.68 ± 0.02 meV), and the stationarity loss alone has no effect (4.15 ± 0.04 meV). The only harmful combination is the stationarity loss together with the energy bound and the cache but without GDI (4.5 ± 0.2 meV, and 7.4 ± 1.0 meV when the cache is initialized from MINAO, where one of three seeds diverged). All configurations were repeated with three seeds. We also added a paragraph to the ablation section in the appendix that interprets the results of the ablation table.

Table. Ablation of the surrogate losses on U0 evaluated on the QM9 test split. Mean and standard deviation over three seeds are provided. Density: the predicted ground-state density is given as input; stat: stationarity loss; bound: energy-bound loss; cache: perturbed-density cache; GDI: gradient-descent-improvement loss; minao: cache initialized from MINAO; otherwise each cache entry starts from the ground-state density displaced along a random, electron-count-conserving direction, with the L2 displacement radius drawn log-uniformly from [0.002, 0.3]. Gain: MAE of the geometry-only model divided by the MAE of the configuration. † One of three seeds diverged; mean and standard deviation over the other two.

densitystatboundcacheGDIminaoMAE (meV)gain
4.23 ± 0.091.00×
●4.14 ± 0.061.02×
●●4.15 ± 0.041.02×
●●3.91 ± 0.071.08×
●●●3.94 ± 0.031.07×
●●●3.68 ± 0.021.15×
●●●●4.5 ± 0.20.93×
●●●4.0 ± 0.11.06×
●●●●3.9 ± 0.11.08×
●●●●3.70 ± 0.071.14×
●●●●●7.4 ± 1.0 †0.57×
●●●●2.37 ± 0.051.78×
●●●●●●2.18 ± 0.091.94×

Q5. Justify / ablate each beSEN modification (Gram matrix vs gate)

"The paper does not explain why each modification in beSEN is necessary, nor does it provide ablations. In particular, is there a specific reason for using the Gram matrix instead of the Gate activation?"

The main text already compares eSEN and beSEN for all three model types (Table 3). In the revision we repeat this comparison at 48 channels and 6 layers with three seeds each: beSEN lowers the U0 MAE for all three model types, from 5.7 ± 0.5 to 4.23 ± 0.09 meV for the geometry-only model, from 5.5 ± 0.3 to 4.14 ± 0.06 meV for the density-input model and from 3.8 ± 0.3 to 2.18 ± 0.09 meV for the surrogate model. The improvement is largest for the surrogate model. In addition, we added an ablation of the individual modifications in Appendix A.8.

In this ablation, we vary the seven components in which beSEN differs from eSEN: the residual connections, the node update, the final normalization, the readout, the scaling of the readout output, the radial basis and the edge-degree embedding. We use a fractional factorial design with 16 combinations of eSEN and beSEN components, each trained on QM9 U0 with a single seed. No combination outperforms beSEN: all 15 alternatives are either worse or diverge during training. All diverged runs combine the residual connections of one model with the final normalization of the other, which shows that these two components have to be chosen together. The results of all combinations are shown in Appendix A.8 (Table 12).

Regarding the Gram matrix: both the Gate and the Gram activation use scalar features to gate the tensorial features. The Gate activation computes the gates from the scalar (l = 0) part of the node features only. The Gram activation however computes not only a scaling of the tensorial features but also a weighted mixing of the channels. In the ablation, the Gram node update seems to have one of the largest effects of all components.

Reviewer_6iAbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for extending surrogate functions, originally developed for orbital-free density functional theory, to molecular property prediction. The authors introduces a physically-motivated constraint to enforce the ground-state density to be the minimum of the surrogate functional. Experiments demonstrate state-of-the-art accuracy for six extensive the QM9 properties and improved scale-related extrapolation on the QM40 dataset.

Critical strengths
  1. The method achieves state-of-the-art accuracy on six of seven extensive QM9 properties, demonstrating substantial gains over existing molecular property prediction methods.

  2. Ablation study among geometry-only, density-input, and surrogate models using the same backbone clarified the contribution of variational training.

  3. The QM40 evaluation offers an informative size-extrapolation test, showing marked improvements for larger molecules despite training only on QM9.

Critical weaknesses
  1. The proposed framework enforces the ground-state electron density to be a minimum of the learned surrogate functional. However, the physical justification for imposing such constraints on general molecular properties beyond ground-state energy remains unclear. Moreover, the paper does not establish whether these constraints provide benefits beyond conventional regularization techniques or whether their effectiveness extends to experimentally measured molecular properties. Physical justification and validation on experimentally collected data are strongly required.

  2. Although the proposed method achieves state-of-the-art performance on extensive properties, it fails to demonstrate consistent improvements on non-extensive properties. In particular, the prediction errors for some properties increase substantially compared with the density-input baseline on QM9. However, the authors provide insufficient analysis of these performance degradations or the underlying reasons why the proposed method benefit certain properties while adversely affecting others. This raises concerns about the generality of the proposed framework, particularly given that extending surrogate functionals to general molecular properties constitutes a central contribution.

  3. Both QM9 and QM40 contain molecular properties computed with the same theory. Consequently, the reported extrapolation results primarily demonstrate generalization to larger molecules within a shared computational framework, rather than robustness across different electronic-structure approximations or experimental conditions. Although these results are meaningful, additional evaluations using independently generated computational datasets or experimentally measured molecular properties are strongly required to demonstrate the genuine generalizability of the proposed method.

Questions for authors

See the Weaknesses section.

3 comments
AuthorsOpenReview ↗

Answer to Reviewer 6iAb

We thank the reviewer for the review and the helpful comments. In the revision, we reframe the surrogate for properties other than the energy as a regularization of the shared network, compare it to conventional regularization (weight decay and training on perturbed densities), and add an auxiliary surrogate variant that resolves the regressions on the non-extensive properties. All results are now reported as mean and standard deviation over three seeds, and we add results for a second level of theory (PBE). A summary of all new results can be found in the general response.


Q1. Physical justification for constraining non-energy properties

"The proposed framework enforces the ground-state electron density to be a minimum of the learned surrogate functional. However, the physical justification for imposing such constraints on general molecular properties beyond ground-state energy remains unclear."

"Physical justification and validation on experimentally collected data are strongly required."

We agree with the Reviewer that no strong physical justification for imposing the surrogate constraints on non-energy properties can be given. Instead, the employed surrogate method should in this context be understood as a form of regularization. In fact the newly added experiments utilizing the auxiliary surrogate head (see Q4) can be seen as a form of multitask learning resulting in the regularization of the shared trunk. Using this auxiliary head enables state-of-the-art performance also on intensive properties. We now state this interpretation in the introduction, in Section 3.3 and in the discussion of the revised manuscript.


Q2. Benefit beyond conventional regularisation

" Moreover, the paper does not establish whether these constraints provide benefits beyond conventional regularization techniques"

We concur that a direct comparison to conventional regularization techniques strengthens our analysis. Therefore, we added additional experiments, showing that weight decay and perturbing the input densities do not lead to a similar improvement. As perturbed densities we use the MINAO initial guess, which replaces the ground-state density as input for half of each batch while the label stays the ground-state property. We added these experiments to the revised manuscript; details can be found in Appendix A.6.

Table. Mean absolute error of U0 and HOMO on the QM9 test split and on QM40, mean and standard deviation over three seeds. All models use the predicted ground-state density as input. The controls are density-input models trained on perturbed densities or with weight decay, without any surrogate loss. Best value per column in bold.

ArmU0, QM9 (meV)U0, QM40 (meV)HOMO, QM9 (meV)HOMO, QM40 (meV)
Density input4.14 ± 0.06150 ± 1013.7 ± 0.1310 ± 20
+ perturbed densities4.15 ± 0.05190 ± 1013.8 ± 0.2320 ± 30
+ weight decay 1e-24.01 ± 0.08170 ± 3013.55 ± 0.05310 ± 30
+ weight decay 1e-16.5 ± 0.4102 ± 115.41 ± 0.08300 ± 10
Surrogate2.18 ± 0.0968 ± 414.0 ± 0.3220 ± 20
Auxiliary surrogate2.51 ± 0.0263 ± 211.1 ± 0.6280 ± 30

Q3. Experimentally measured properties

"or whether their effectiveness extends to experimentally measured molecular properties."

While we appreciate the importance of agreement with experiment we believe that this is out of scope for this study. Since all properties are trained on generated DFT data, the accuracy of this data is a fundamental limit for any method that solely relies on it for training. This also includes all methods we compare with. Therefore, a comparison with experimental data would confound the accuracy of the underlying DFT method with the accuracy of our trained models.

AuthorsOpenReview ↗

Answer to Reviewer 6iAb (continued)

Q4. Non-extensive properties: no consistent improvement, regressions unexplained

"Although the proposed method achieves state-of-the-art performance on extensive properties, it fails to demonstrate consistent improvements on non-extensive properties. In particular, the prediction errors for some properties increase substantially compared with the density-input baseline on QM9. However, the authors provide insufficient analysis of these performance degradations or the underlying reasons why the proposed method benefit certain properties while adversely affecting others. This raises concerns about the generality of the proposed framework, particularly given that extending surrogate functionals to general molecular properties constitutes a central contribution."

We agree that robustness also for non-extensive properties makes the proposed method much more attractive. By applying the surrogate method on an auxiliary output head we identified a method that achieves state-of-the-art results also for non-extensive properties. Here the gradient descent improvement loss is not applied to the output head that predicts the property, but instead to an auxiliary head. This acts as an effective regularizer for the shared trunk of the model, leading to improvements in accuracy for all properties. This implies that the performance degraded because the additional constraints applied to the property head in the original surrogate setup were too strong or contradicting. The auxiliary head setup fixes this issue. These results are now part of Table 2 of the revised manuscript and are discussed in Section 4.1.

Table. Mean absolute error of the predicted properties on the QM9 test set. Values of GotenNet-L are reproduced from Aykent & Xia (2025). For our models, the mean and standard deviation over three seeds are given. All models that take a density as input are evaluated at the predicted density. Best value per row in bold.

PropertyGeometry onlyDensity inputSurrogateAuxiliary surrogateGotenNet-L
U0 (meV)4.23 ± 0.094.14 ± 0.062.18 ± 0.092.51 ± 0.023.37
U (meV)4.20 ± 0.064.07 ± 0.022.12 ± 0.022.4 ± 0.13.41
H (meV)4.17 ± 0.044.07 ± 0.022.12 ± 0.062.4 ± 0.13.30
G (meV)5.52 ± 0.065.47 ± 0.033.6 ± 0.14.00 ± 0.064.98
ZPVE (meV)1.022 ± 0.0041.0152 ± 0.00070.903 ± 0.0070.901 ± 0.0021.08
α (m a0^3)33.9 ± 0.333.3 ± 0.623.1 ± 0.323.2 ± 0.228
Cv (mcal/mol/K)19.94 ± 0.1019.5 ± 0.116.1 ± 0.216.0 ± 0.319
HOMO (meV)13.7 ± 0.213.7 ± 0.114.0 ± 0.311.1 ± 0.613.4
LUMO (meV)13.5 ± 0.713.2 ± 0.513.7 ± 0.812.2 ± 0.612.2
gap (LUMO − HOMO) (meV)20.5 ± 0.520.1 ± 0.220.7 ± 0.717.6 ± 0.819.8
μ (mD)6.8 ± 0.56.52 ± 0.088.3 ± 0.55.2 ± 0.46.7
⟨R²⟩ (m a0^2)15 ± 113.1 ± 0.511.91 ± 0.1011.0 ± 0.524

Table. Mean absolute error of the predicted properties on 600 QM40 molecules from 10 to 40 heavy atoms. The dataset is split into bins sorted by the number of heavy atoms, with width of 5 atoms. From each of these bins 100 molecules with atom types H, C, N, O and F that are neutral and with even electron count are selected. Each evaluated model was trained only on QM9. Especially for the extensive properties, using the surrogate functional approach enables better generalization to larger molecules. For our models, the mean and standard deviation over three seeds are given. Best value per row in bold.

PropertyGeometry onlyDensity inputSurrogateAuxiliary surrogate
U0 (meV)200 ± 20150 ± 1068 ± 463 ± 2
U (meV)169 ± 6146 ± 1064 ± 867 ± 1
H (meV)190 ± 10150 ± 1065 ± 961 ± 5
G (meV)114 ± 7102 ± 262 ± 462 ± 4
ZPVE (meV)158 ± 10156 ± 838 ± 2073 ± 23
α (m a0^3)2500 ± 3003100 ± 5002600 ± 3002130 ± 50
Cv (mcal/mol/K)4200 ± 10003000 ± 200280 ± 401500 ± 300
HOMO (meV)310 ± 10310 ± 20220 ± 20280 ± 30
LUMO (meV)610 ± 70580 ± 90580 ± 140590 ± 20
gap (LUMO − HOMO) (meV)910 ± 80890 ± 100770 ± 120860 ± 40
μ (mD)124 ± 7124 ± 5125 ± 7128 ± 3
⟨R²⟩ (m a0^2)4100 ± 2003800 ± 3003660 ± 603180 ± 70
AuthorsOpenReview ↗

Answer to Reviewer 6iAb (continued)

Q5. Generalisation across levels of theory / independent datasets

"Both QM9 and QM40 contain molecular properties computed with the same theory. Consequently, the reported extrapolation results primarily demonstrate generalization to larger molecules within a shared computational framework, rather than robustness across different electronic-structure approximations or experimental conditions. Although these results are meaningful, additional evaluations using independently generated computational datasets or experimentally measured molecular properties are strongly required to demonstrate the genuine generalizability of the proposed method."

Since different electronic-structure approximations result in different approximations to the true properties, a method trained on one level of theory cannot be expected to generalize to another level of theory. Evaluating this generalization would again confound the accuracy of the underlying DFT method with the accuracy of our trained models (as is the case for experimental data). However, to demonstrate that the proposed method is not limited to a single level of theory, we additionally provide results for QM9 with PBE labels with the 6-31G(2df,p) basis set and otherwise the same settings as the used B3LYP data. Here the models are trained on the PBE level of theory and evaluated on the same level of theory. The results show that the proposed method is robust when trained and evaluated on a different level of theory. These results are now included in Appendix A.9 of the revised manuscript.

The surrogate or the auxiliary surrogate outperforms the density input baseline for all evaluated properties.

Table. Mean absolute error on the QM9 test split with PBE labels, mean and standard deviation over three seeds. All models that take a density as input are evaluated at the density predicted by a PBE density prediction model. Best value per row in bold.

PropertyGeometry onlyDensity inputSurrogateAuxiliary surrogate
E_tot (meV)4.63 ± 0.014.50 ± 0.092.2 ± 0.12.59 ± 0.07
HOMO (meV)14.0 ± 0.213.8 ± 0.214.5 ± 0.311.4 ± 0.3
LUMO (meV)12.9 ± 0.212.75 ± 0.0613.7 ± 0.211.6 ± 0.4
μ (mD)7.9 ± 0.57.54 ± 0.099.0 ± 0.25.9 ± 0.4
⟨R²⟩ (m a0^2)16 ± 214.3 ± 0.514 ± 211.1 ± 0.3

In summary, we now present the surrogate for properties other than the energy as a regularizer, show that it outperforms conventional regularization, resolve the regressions on the non-extensive properties with the auxiliary surrogate, which achieves state-of-the-art results also on these properties, and show that the improvements carry over to a second level of theory. We hope that these changes address the reviewer's concerns and would be grateful if the reviewer would reconsider the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Answer to all Reviewers

We thank all three reviewers for their detailed and constructive reviews. Several concerns were shared across the reviews: why the surrogate approach works for properties other than the energy, the missing seed variance, the regressions on the non-extensive properties, the generalization to another level of theory, and reproducibility. Below we summarize the changes and new results of the revision, each is discussed in detail in the answers to the individual reviewers.

  • Larger models. We now report results for larger models (48 channels, 6 layers) instead of the original smaller models (32 channels, 4 layers), improving QM9 results for all properties.
  • Seed variance. We retrained all models with three seeds and now report mean and standard deviation in all main tables.
  • Auxiliary surrogate. A new variant trains the surrogate on a separate head of the same network, while the property head is trained only on the labels. This gives state-of-the-art results also on all non-extensive properties.
  • The surrogate as a regularizer. For properties other than the energy, we now describe the surrogate as a regularizer of the shared network. Control experiments with perturbed input densities and with weight decay show that the surrogate improves over conventional regularization.
  • Understanding the mechanism. A linear-probe analysis of the learned representations shows that the surrogate models keep the ground-state density in their internal features, while the density-input model discards it.
  • Other level of theory. With PBE labels instead of B3LYP, the surrogate or the auxiliary surrogate improves over the density-input model for all evaluated properties.
  • Architecture. We ablate the individual beSEN modifications.
  • Direct density prediction. We show that evaluating the surrogate at the directly predicted density instead of the ground-state density changes the error by at most 0.1 % on all twelve properties.
  • Reproducibility. We will release the code and the QM40 molecule identifiers and added a reproducibility statement.

Table G1. Models are trained on QM9 with three seeds and evaluated on the QM9 test split. The mean absolute error (MAE) and standard deviation over three seeds are reported. Every density-consuming arm is read at the density of a direct prediction model. GotenNet-L as published (Aykent & Xia, 2025, Table 1). Best entry in bold.

QM9 testGeometry onlyDensity inputSurrogateAuxiliary surrogateGotenNet-L
U0 (meV)4.23 ± 0.094.14 ± 0.062.18 ± 0.092.51 ± 0.023.37
U (meV)4.20 ± 0.064.07 ± 0.022.12 ± 0.022.4 ± 0.13.41
H (meV)4.17 ± 0.044.07 ± 0.022.12 ± 0.062.4 ± 0.13.30
G (meV)5.52 ± 0.065.47 ± 0.033.6 ± 0.14.00 ± 0.064.98
ZPVE (meV)1.022 ± 0.0041.0152 ± 0.00070.903 ± 0.0070.901 ± 0.0021.08
α (m a₀³)33.9 ± 0.333.3 ± 0.623.1 ± 0.323.2 ± 0.228
Cv (mcal/mol K)19.94 ± 0.1019.5 ± 0.116.1 ± 0.216.0 ± 0.319
HOMO (meV)13.7 ± 0.213.7 ± 0.114.0 ± 0.311.1 ± 0.613.4
LUMO (meV)13.5 ± 0.713.2 ± 0.513.7 ± 0.812.2 ± 0.612.2
gap (meV)20.5 ± 0.520.1 ± 0.220.7 ± 0.717.6 ± 0.819.8
µ (mD)6.8 ± 0.56.52 ± 0.088.3 ± 0.55.2 ± 0.46.7
⟨R²⟩ (m a₀²)15 ± 113.1 ± 0.511.91 ± 0.1011.0 ± 0.524

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.