acceptodds
Under review as a conference paper at ICLR 2027

Curvature-Aware Flow Matching for Molecular Structure Generation

Abstract

Generative models produce three-dimensional molecular structures with high geometric accuracy, yet they are trained on distributions of atomic coordinates without explicit access to the underlying physics. The potential energy surface (PES) describes how the energy changes around a molecular geometry, and the RMSD between a generated structure and its reference captures this local landscape only indirectly. Supervising the PES during training could close this gap, but conformers and transition states (TSs) are stationary points at which the forces vanish, so first-order quantities carry little information about them. What remains is the local curvature, encoded in the Hessian of the PES. Quantum-chemical Hessians are too expensive to be available at the scale required for training, but recently proposed machine-learned predictors make curvature inexpensive enough to supervise directly. We introduce CurvFM, a flow matching (FM) formulation whose training objective constrains the curvature at the generated structure. A frozen Hessian prediction model is evaluated at the generated endpoint and at the reference structure, which removes the need for quantum-chemical Hessians and keeps the residual cheap enough to impose at every training step. Across TS and conformer generation, CurvFM substantially improves physical fidelity at unchanged global structural accuracy, reducing the median deviation in maximum force by 63% on GEOM-QM9 and raising the fraction of valid TSs by up to 10% across three reaction datasets.

Primary area:
generative models
Keywords:
flow matching, generative models, physics-informed machine learning, machine-learned interatomic potentials, Hessian, potential energy surface, AI for science, molecular structure generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_79FVOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CurvFM, a curvature-aware flow matching framework that improves the physical fidelity of generated molecular structures through Hessian-based supervision. The authors argue that geometric accuracy alone does not guarantee physical correctness, as geometrically similar structures can exhibit different energies, forces, and PES curvatures. Unlike energy- or force-based approaches that rely on zeroth- or first-order information, CurvFM incorporates second-order curvature information to better capture the local PES characteristics of conformers and transition states (TSs).

CurvFM fine-tunes pretrained flow matching models without modifying their architectures. During training, the predicted velocity field is used to estimate the clean molecular structure, and a pretrained, frozen Hessian predictor (HIP) computes Hessians for both the estimated and reference structures. An auxiliary loss minimizes their Hessian difference, weighted by a time-dependent schedule. Direct Hessian prediction avoids the expensive higher-order derivatives required by energy-based potentials. The authors also investigate low-eigenmode subspace matching but ultimately adopt full Hessian matching, which performs better in their ablations. The framework introduces no additional inference-time computation.

Experiments cover TS generation using React-OT on three Transition1x datasets and conformer generation using ET-Flow on GEOM-QM9. For TS generation, CurvFM largely preserves geometric accuracy while reducing barrier-energy errors and maximum forces, improving Hessian eigenvalue accuracy, and increasing Hessian-based TS validity rates. For conformer generation, it maintains comparable coverage and RMSD while substantially reducing energy and force deviations before relaxation.

Critical strengths

1. Simple and Efficient Integration of Physics-Based Supervision: The paper addresses an important limitation of geometry-based molecular generation: low structural error does not necessarily imply physical accuracy. Incorporating PES curvature provides a physically motivated direction for improving generated molecular structures. Additionally, the proposed approach is simple and can be incorporated into existing flow matching architectures without modifying the generator or introducing additional inference-time overhead.

2. Improvements in Physical Fidelity Across Two Generation Tasks: The experiments demonstrate improvements in physical metrics across conformer and TS generation. In particular, the improvements in QM-evaluated energy, force, and Hessian properties suggest that the method provides benefits beyond merely matching the predictions of the pretrained Hessian surrogate.

Critical weaknesses

1. Insufficient Experimental Validation

• Insufficient Comparison with Existing Physics-Based Baselines: Comparisons with energy-guided generation (e.g., EnFlow) or Hessian-assisted post-generation refinement should have been provided. Also, CurvFM achieves comparable performance to DiTMC but does not demonstrate clear, consistent improvements in physical metrics (Table 16, 17). The authors' argument that improving a weaker backbone to a comparable level is sufficient is not fully convincing.

• Limited Evaluation of Sampling Efficiency and Scalability: The authors should provide comparisons with existing methods in terms of physically valid structures generated under matched computational budgets. Also, conformer experiments are limited to small GEOM-QM9 molecules, without validation on larger drug-like systems.

• Incomplete Validation of Physical and Chemical Validity: Valid TS (%) only checks for one negative Hessian eigenvalue, without ensuring near-zero forces or correct reactant–product connectivity. Reporting joint force–Hessian validity rates and the Chemical Accuracy Rate used in React-OT would provide stronger evidence of the physical and chemical validity of the generated structures.

• Insufficient Evidence for Model-Agnostic Generalization: Although CurvFM is presented as a model-agnostic framework, the experiments employ only one backbone architecture per task: ET-Flow for conformer generation and React-OT for TS generation, which is insufficient to claim the model-agnostic feature of this framework.

2. Unclear Practical Benefits after Geometry Optimization

• Despite improvements in physical accuracy before relaxation, the performance gap largely disappears after geometry optimization, while gains in optimization efficiency remain modest or inconsistent (Table 4). Therefore, the practical advantages of Hessian-based supervision over the original backbones remain unclear.

Questions for authors

Major Questions

  1. While Hessian matching empirically improves physical accuracy, similar Hessians do not necessarily imply proximity to the same stationary point. Could the authors further justify why directly matching reference Hessians is an appropriate objective for molecular structure generation? In particular, could they provide a theoretical explanation or empirical analysis of why minimizing Hessian differences leads to lower force errors, despite not directly constraining the energy gradients?

  2. CurvFM incorporates Hessian supervision as an additive term to the original Flow Matching objective. What motivates this simple additive formulation, rather than a more explicitly coupled geometric–physical learning objective such as transfer learning, multi-task learning? How are potential conflicts between geometric accuracy and curvature accuracy addressed?

  3. Is there a particular reason for focusing on Flow Matching? Given that Flow Matching learns a velocity field rather than directly predicting molecular structures, could the authors clarify what the flow matching model actually learns from Hessian-matching supervision? In particular, how does matching the Hessians of estimated endpoints and reference structures influence the learned velocity field?

  4. The paper frames CurvFM as a molecular structure generation framework, yet much of the evaluation focuses on reproducing reference structures and their physical properties. Could the authors clarify how the proposed task differs from structure prediction, particularly in terms of its generative nature and intended applications?

  5. Could the authors release the reproducible source code?

Minor Questions

  1. Why is Subspace Hessian Matching (Eq. 5) introduced in the main Method section when it is not used in the final model? Furthermore, the ablation considers only k=8 lowest eigenmodes. Would the conclusion favoring Full Hessian Matching remain consistent across different subspace sizes or eigenmode selection strategies?

  2. Table 8 reports negative mode overlap based on the lowest-eigenvalue eigenvectors. Is this metric computed over all generated structures or only those with exactly one negative Hessian eigenvalue? If all structures are included, how should the overlap be interpreted for structures with zero or multiple negative eigenvalues?

4 comments

W1.1: Comparison with physics-based baselines

We thank the reviewer for this suggestion. We have added a comparison with Hessian-assisted post-generation refinement (Section 5.1 line 445-455, Appendix C.2.6). We use HIP as an MLIP that drives a P-RFO optimization, with the Hessian predicted directly by the surrogate, both to refine the generated TSs and to search from scratch. Refinement yields more accurate structures, but even a single step triples the cost per reaction, full refinement raises it by a factor of about 40, and the optimization occasionally fails where the MLIP is unreliable. CurvFM keeps the inference cost of the baseline and produces no such failures, so the two approaches are complementary, with refinement as an optional final step for the most promising candidates.

Regarding DiTMC, we agree that comparable performance alone does not establish the benefit of curvature supervision. We therefore quantify the contribution of CurvFM through controlled comparisons within each backbone. For ET-Flow, both variants start from the same checkpoint and are finetuned for 50 epochs. Adding CurvFM reduces the median maximum-force deviation from 15.91 to 6.02 kcal/mol/Å (62%) and the median ensemble-energy deviation from 0.49 to 0.15 kcal/mol (69%), while leaving the global structural metrics largely unchanged (Section 5.2). The same holds for React-OT and GoFlow (Appendix C.2.7), at no additional inference cost. The DiTMC comparison serves a different purpose, namely assessing absolute performance. Despite its lower structural accuracy, ET-Flow with CurvFM achieves 19% lower mean maximum-force deviations, while DiTMC remains ahead on the ensemble and minimum energies by about 20%. Higher structural accuracy thus does not necessarily imply higher physical fidelity. We have revised Section 5.2 (line 486-493) and Appendix C.3.2 accordingly, distinguishing the improvements attributable to CurvFM from this cross-model comparison.

 

W1.2: Sampling efficiency and scalability

CurvFM samples at exactly the cost of its backbone (91 ms per reaction, Appendix C.2.6), so under a matched inference budget it yields more physically valid structures. Regarding larger systems, we have added Pt(II)-catalyzed methane C–H activation (Pt-cat), a catalysis application with systems of up to 45 atoms (mean 26), roughly twice the size of the Transition1x reactions (Appendix A.1). CurvFM nearly halves the median barrier error and triples the chemical accuracy rate from 8 to 24% (Section 5.1, Appendix C.2.1).

Newly added Table 8 (Appendix C.2.1):

ModelRMSD (Å)D-MAE (Å)ΔE (kcal/mol)Fmax (kcal/mol/Å)
React-OT0.195 / 0.1490.090 / 0.07115.29 / 6.9091.26 / 62.70
React-OT + CurvFM0.195 / 0.1520.087 / 0.06715.97 / 3.5588.62 / 32.71

 

W1.3: Physical and chemical validity of the generated TSs

We have added both analyses to Appendix C.2.3. First, we report the joint validity, i.e. the fraction of generated TSs with exactly one negative Hessian eigenvalue and a maximum force below a given threshold, as a function of that threshold. CurvFM yields the higher fraction at every threshold, most clearly on Transition1x-TMC. Second, we report the chemical accuracy rate following React-OT, i.e. the fraction of TSs with a barrier error below 1.58 kcal/mol. CurvFM raises it from 38.9 to 44.9% on Transition1x-2p3p4p and from 13.7 to 22.7% on Transition1x-TMC, and from 8 to 24% on the newly added Pt-cat set (Appendix C.2.1).

 

W1.4: Model-agnostic generalization

To test architecture independence, we have added experiments with GoFlow, a second TS generation backbone, on RDB7 (Appendix C.2.7). We use the public HIP checkpoint as surrogate, although RDB7 uses a different level of theory (ωB97X-D3/def2-TZVP) than HIP was trained on. CurvFM again improves the energetic metrics, reducing the median barrier error by 8% and the median maximum force by 10%, with reductions of about 25% in both means. Together with React-OT and ET-Flow, CurvFM thus improves three different backbones across two tasks.

 

W2: Practical benefits after geometry optimization

We agree that the main value of CurvFM lies before relaxation, which matters most in screening applications, where optimizing every candidate is too costly. Our new comparison with MLIP-driven optimization shows that even a single refinement step triples the cost per reaction (see W1.1). We have added two applications where the generated structures are used directly (Section 5.1, Appendix C.2.1). On Pt-cat, CurvFM triples the fraction of TSs within chemical accuracy from 8 to 24%. In a combustion reaction network, ranking the competing channels by the energies of the raw generated TSs improves the Spearman correlation with the reference barriers from 0.75 to 0.89, so the relevant pathways are identified more reliably before any optimization.

Q1: Why Hessian matching lowers force errors

We have added a theoretical analysis in Appendix B. The residual compares Hessians at the endpoint estimate x̂₁ and the reference x₁, so it acts locally around the reference stationary point. Expanding both the force and the Hessian around x₁ shows that both depend linearly on the same displacement δ = x̂₁ − x₁. Assuming that every displacement other than a translation changes the Hessian to first order, the Hessian error bounds the displacement, and thereby the force, Fₘₐₓ ≲ κ‖ΔH‖ (Frobenius norm). The force thus vanishes with the Hessian error, although it never enters the objective. Empirically, the Hessian error and the maximum force of the generated TSs are strongly correlated (Spearman ρ = 0.76 on Transition1x-2p3p4p, both computed with GFN2-xTB), and both grow with the RMSD to the reference, as predicted (Appendix B, "Empirical verification").

 

Q2: Additive formulation and conflicts between objectives

We chose the additive form to keep CurvFM simple and agnostic to both the model and the generative framework. The residual only requires an endpoint estimate, which flow matching, diffusion, and direct structure prediction all provide (see Q3), so it can be added to any such objective without architectural changes, additional heads, or training stages. CurvFM can thus be seen as multi-task learning with a shared model, with physical knowledge transferred from a pretrained PES surrogate. More explicitly coupled objectives are an interesting direction for future work, but lie outside the scope of this paper. Regarding conflicts, our new analysis in Appendix B shows that near the reference, the gradient of the residual has a strictly positive inner product with the endpoint error that the FM loss corrects. The residual thus reweights the same correction toward curvature-sensitive directions rather than introducing a competing target. Empirically, the two gradients align for 97.8% of the one-shot predictions on Transition1x-2p3p4p (Table 7, Appendix B). Conflicts arise mainly far from the reference, at small t, where the weight tᵖ suppresses the residual, so 99.5% of the training weight falls on aligned samples. We thank the reviewer for this question, which led us to add this analysis of the loss gradients.

 

Q3: Focus on flow matching and effect on the velocity field

We use generative models because both tasks are inherently generative (see Q4). Of the two most prominent frameworks, flow matching and diffusion, we chose flow matching because its straight transport paths yield accurate one-step endpoint estimates during training, on which the residual relies. Both frameworks implicitly contain a structure prediction loss. With the linear path xₜ = (1 − t)x₀ + t x₁, target velocity uₜ = x₁ − x₀, and predicted velocity v(xₜ, t), the endpoint estimate x̂₁ = xₜ + (1 − t) v(xₜ, t) gives δ = x̂₁ − x₁ = (1 − t)(v(xₜ, t) − uₜ), so predicting the velocity is equivalent to predicting the structure, and the FM loss penalizes δ up to a time-dependent factor. The same holds for diffusion, whose noise and data predictions relate linearly to x̂₁, and for direct structure prediction (Appendix B). The residual therefore acts on the same velocity error as the FM loss, but weights its components by how strongly they change the curvature of the PES, so the velocity field learns to correct physically relevant errors most strongly

 

Q4: Generation versus structure prediction

Both tasks are generative in nature. In conformer generation, a molecular graph admits many 3D conformers following the Boltzmann distribution, and our evaluation accordingly measures how well the generated ensemble recovers the reference ensemble, through recall, precision, and ensemble properties (Section 5.2). TS generation is likewise a distribution task. A reaction can proceed along several transition paths, and different conformers of the reactant and product lead to competing TSs. Current TS datasets, however, cover only one minimum energy path per reaction, so the reference TS is the only ground truth available for evaluation, and the TS ensemble cannot be assessed directly. In practice, such models can still recover TS ensembles when provided with ensembles of reactant and product conformers, which is the intended use in reaction network exploration and kinetic screening.

 

Q5: Code release

We will release the full source code upon publication, including the training and evaluation pipelines, the finetuned HIP surrogates, the trained model checkpoints, and the data splits, so that all results can be reproduced.

Q6: Role of the subspace variant and choice of k

We introduce the subspace variant in the Method section for two reasons. First, it illustrates that CurvFM is a general framework rather than a single loss, as curvature can be compared entrywise, projected onto selected modes, or combined with other PES quantities such as forces, and we hope these variants inspire further formulations. Second, it tests the natural hypothesis that constraining only the lowest modes, which define the TS character, suffices. Our analysis shows that the gains stem mainly from the stiff curvature outside any low-lying subspace, so the full residual performs best (Section 5.1). We chose k = 8 to cover the two lowest vibrational modes, as the six rigid-body modes are retained for numerical stability (Appendix C.2.5). As k grows, the subspace variant approaches the full residual, which it covers entirely at k = 3N. We have added ablations with k = 1 and k = 15, using the weights optimized for k = 8 (Appendix C.2.5). Neither changes the picture. With k = 1, the energetic metrics fall slightly below the structural baseline, and k = 15 performs on par with k = 8.

 

Q7: Interpretation of the negative mode overlap

The overlap in the main table is computed over all generated structures. Since the lowest mode of a structure without or with several negative eigenvalues need not correspond to a reaction coordinate, we have added an analysis that resolves the overlap separately for valid TSs and the remaining structures (Appendix C.2.2). Among valid TSs, both models reach virtually the same median overlap (e.g. 98.2 versus 98.5% on Transition1x-2p3p4p), and among the remaining structures they differ by at most 1.5 percentage points. The small overall gain therefore stems not from a more accurate reactive mode, but from the larger fraction of valid TSs under CurvFM.

We thank you for taking the time to review our paper and for your constructive feedback. As the discussion deadline is approaching, we wanted to ask whether there are any remaining concerns or questions that we could address.

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

In this work, authors propose to improve molecule structure generation via matching the Hessian of the generated and reference structures via a residual. They in particular show the implementation in a flow matching-based mol struct gen case, but I believe it can be expanded. The idea is that in the past, using Hessians in this manner was very expensive, as they are expensive to calculate exactly, but now there are relatively accurate ML models that predict Hessians, so they can be used in loss, especially if the systematic errors in predictions can be cancelled by using them also for the reference. The authors choose transition state generation and conformer generation as important tasks that this approach can be useful for. The most important contributions are

  • CurvFM scheme to showcase how to use hessian predictors in mol struct gen, experimenting with a few approaches on residual design, and complementing their results with analysis
  • allowing community to evaluate the accuracy of recent ML based direct hessian predictors -They also extend the reaction datasets with new data, including technologically relevant transition metal dataset, which they commit to sharing upon acceptance.
Critical strengths

Strengths

  • The method is novel, in the sense that, even tho using Hessians to guide generation is not new, they test the relatively new ML surrogates and explore relatively cheap implementations, so it is now a practical approach. Out of the box it can be applied to any flow matching model but it is simple enough to be expanded to other generative schemes with little effort. This may inspire other tasks where hessians can be used to improve generation.
  • The paper, through the results, show that the accuracy of the ML surrogates are good enough to obtain good structure, without having to do qc level of computation, which may be an important validation needed for the penetration of ML based hessian predictors.
Critical weaknesses

The reason why I rated soundness low is also the main weakness of the paper, and am happy to increase my score if authors can point me to counter evidence:

  • The whole premise of the paper is about how Hessians as a second order property hold more information and therefore can steer the flowmatching better than zeroth and first order ones, and hence should yield significantly improved generation due to increased alignment. I have one a priori and one a posteriori reason to challenge that: 1-For the generation of the molecules at the minimum or saddle point, their argument is that forces are zero, and what distinguishes is the curvature, which is true. However there are a lot of points around that zero point conformation where forces can still provide signal. Indeed some of the examples they highlight are about steric clashes, which may be already addressed sufficiently with a good first order steered generation model without going to second order. 2-a posteriori, indeed we do not see a big jump in the results between first order or second order when it comes to accuracy. table 13 (supp mat) shows that forces are already addressing some alignment and there are simply more models and cheaper models for them. It would have been good to have this in the main text.

  • And as a general issue, I would appreciate if the authors tied the very small, like third digit in Angstrom, or second digit in kcal/mol differences between results to some real world application value. Maybe what appears as small gains to me are very consequential in some application? I cannot tell that from the results. The level of impact I see on the results are generally in the order one would see with a few steps in relaxation, a different method than gfn2-xtb, or even a slightly different functional for the much more expensive density functional etc. Perhaps there is a ranking order, binary classification value I am unable to appreciate right now. If the authors could address this that would be great.

  • Since the accuracy improvements I look for speed related benefits to the approach but as table 15 again in supplementary seems to show, the hessian based generation does not bring more gain than what could be achieved with a few steps of relaxation. Consdiering that the method comes with retraining, and may carry the biases of the ML surrogate for Hessians, I am not sure if the added cost is worth it, maybe for the structures where relaxation is not an option.

  • Not severe but the authors could present a few more details about the actual model used, which described only in a single sentence in the appendix and refers to other papers for the architecture and its parameters.

  • One significant increase due to Hessian approach was the % of valid-TS scores. I could not find the total sample number so it was not clear how to judge the difference.

Questions for authors

Addressing the points in weaknesses would suffice. Basically,

  • I would appreciatecritical look at the value of hessians in mol struct gen in their results,
  • I would also appreciate connection of small numbers to any real world applications.
  • It would help to hear authors' opinion on the cost and gain for this approach vs a few more steps of relaxation, considering the effort spent in alignment of the ML surrogate to the data, and retrain of the generative model.
5 comments

W1/Q1: Value of Hessians compared with forces

We sincerely thank the reviewer for this valuable question, which led to a substantial extension of our analysis and improved the revised paper. To our knowledge, CurvFM is the first framework that integrates surrogate-predicted PES information, both Hessians and forces, directly into the training objective of generative models for molecular structures. We agree that one-shot predictions are rarely exact stationary points, so forces can provide a meaningful training signal if the surrogate describes them reliably. We have revised the Introduction and Method accordingly and now benchmark both signals on Transition1x-2p3p4p and Pt-cat (see W2/Q2), covering two surrogates and their combination (Section 5.1 line 412-420, Appendix C.2.5). With a surrogate trained on IRC data, the force residual matches the curvature residual on Transition1x-2p3p4p, and combining both yields the same improvement. On Pt-cat, forces alone improve the barriers considerably less, but combined with curvature they match its energetic accuracy and give the best structural accuracy. Forces and curvature are thus complementary signals whose relative value depends on the system, and curvature improved the model in every setting we considered.

Excerpt of added Table 17 (Appendix C.2.5):

ModelRMSD (Å)D-MAE (Å)ΔE (kcal/mol)FmaxF_\text{max} (kcal/mol/Å)
Transition1x-2p3p4p
React-OT0.196 / 0.1340.083 / 0.0605.85 / 2.5340.96 / 33.50
CurvFM (forces (IRC))0.196 / 0.1340.082 / 0.0574.89 / 1.9130.82 / 25.42
CurvFM (Hessian (eq.))0.196 / 0.1330.082 / 0.0575.10 / 1.9331.48 / 25.76
CurvFM (forces & Hessian (IRC))0.196 / 0.1350.082 / 0.0575.09 / 1.9131.77 / 26.68
Pt-cat
React-OT0.195 / 0.1490.090 / 0.07115.29 / 6.9091.26 / 62.70
CurvFM (forces (IRC))0.193 / 0.1490.089 / 0.07321.33 / 6.06115.65 / 46.72
CurvFM (Hessian (eq.))0.195 / 0.1520.087 / 0.06715.97 / 3.5588.62 / 32.71
CurvFM (forces & Hessian (IRC))0.191 / 0.1470.084 / 0.06317.66 / 3.5297.57 / 34.65

 

W2/Q2: Practical significance of the improvements

The barrier height enters the reaction rate exponentially, so seemingly small energy differences matter. At 70 °C, a barrier error of 1.58 kcal/mol changes the rate by an order of magnitude, the common threshold for chemical accuracy. We now report the fraction of generated TSs within this threshold, a binary criterion for direct use in kinetic screening. CurvFM raises it from 38.9 to 44.9% on Transition1x-2p3p4p and from 13.7 to 22.7% on Transition1x-TMC (Appendix C.2.3). We also added two real-world applications (Section 5.1, Appendix C.2.1). On Pt(II)-catalyzed methane C–H activation (Pt-cat), CurvFM nearly halves the median barrier error (6.90 to 3.55 kcal/mol) and triples the chemical accuracy rate (8 to 24%). In a combustion reaction network, ranking competing channels by the energies of the raw generated TSs improves the Spearman correlation with the reference barriers from 0.75 to 0.89, so the dominant pathways are identified more reliably before any refinement.

 

W3/Q3: Cost and benefit compared with a few relaxation steps

We see CurvFM and relaxation as complementary. CurvFM mainly benefits screening applications, such as kinetic modeling or reaction network exploration, where relaxing every candidate is too costly (see W2/Q2). We also added a comparison with MLIP-driven optimization using the IRC surrogate (Section 5.1 line 445-455, Appendix C.2.6). Refinement yields more accurate structures, but even a single P-RFO step triples the cost per reaction, full refinement raises it about 40-fold, and it occasionally fails where the MLIP is unreliable. CurvFM keeps the inference cost of the baseline, with refinement as an optional final step for the most promising candidates.

Regarding training, CurvFM requires no additional retraining. It uses the same number of epochs as the FM baseline, with each step about 30% slower (Appendix C.1), plus a short search over the residual weight and time threshold. If a pretrained Hessian model is available, no surrogate training is needed either. We have added such an experiment, applying CurvFM to GoFlow on RDB7 with the public HIP checkpoint. RDB7 uses a different level of theory (B97X-D3/def2-TZVP) than the one HIP was trained on (B97x/6-31G(d)), yet CurvFM still improves the energetic metrics, so an existing Hessian model transfers across levels of theory (Appendix C.2.7).

W4: Details on the models used

We have substantially extended the implementation details in Appendix A. Appendix A.1 now describes all datasets, including their size, level of theory, molecular size, and splits. Appendix A.2 gives more details on the training of the generative models and the HIP surrogates. Appendix A.3 describes the model architectures, and Appendix A.4 lists all hyperparameters, so that the setup can be understood without consulting the original papers.

 

W5: Number of samples behind the valid TS rate

We thank the reviewer for pointing this out. We have added the number of test reactions to Section 5.1 and Appendix A.1. The valid TS rate is computed over all 1073, 1552, and 3813 test reactions of Transition1x, Transition1x-2p3p4p, and Transition1x-TMC, respectively, and over the 145 test reactions of Pt-cat.

We thank you for taking the time to review our paper and for your constructive feedback. As the discussion deadline is approaching, we wanted to ask whether there are any remaining concerns or questions that we could address.

Reviewer_R8ACOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper observes that molecular generative models are trained on coordinate distributions and evaluated with RMSD-like geometry metrics, which say little about where a structure sits on the potential energy surface (PES). Because the structures of interest (conformers, transition states) are stationary points, first-order signals (forces, energies) vanish there and carry little information; the lowest-order informative quantity is the local curvature, i.e. the PES Hessian. Quantum-chemical Hessians are too expensive to supervise with, so the authors propose CurvFM: during flow-matching training, the predicted velocity is integrated to an estimate of the clean endpoint, and a frozen Hessian prediction model (HIP) is evaluated at both the estimated endpoint and the ground-truth endpoint. The disagreement between the two predicted Hessians is used as an additional residual. Evaluating both sides with the same surrogate is argued to cancel the surrogate's systematic error and to remove the need for reference Hessian labels. The method is evaluated on three transition-state datasets (Transition1x, Transition1x-2p3p4p, Transition1x-TMC) and on conformer generation on GEOM-QM9, with structural metrics (RMSD, D-MAE), energetic metrics (ΔE, maximum force), eigenvalue/eigenvector error, and valid-TS rate.

Critical strengths
  1. The problem diagnosis is precise and the fix is unusually cheap. The argument that curvature is the lowest-order quantity that distinguishes a minimum from a first-order saddle — and therefore the only physically meaningful supervision signal available at stationary points — is a clean reframing of why geometry-only supervision is insufficient. Deriving the residual as a comparison between two evaluations of the same frozen predictor, rather than against reference Hessians, is the key move: it makes the objective applicable to datasets with no Hessian labels at all, which is what makes the idea practical rather than merely conceptually appealing.

  2. The evaluation is well matched to the claim and reports its own negative results. The gains are reported on the physical quantities the method targets (max force, ΔE, eigenvalue error, valid-TS rate) rather than on the geometry metrics the baseline already optimizes, and the paper explicitly notes that global structural accuracy is unchanged (RMSD 0.188 both with and without the residual on Transition1x) — i.e. the improvement is not a re-labelling of a geometry gain. It also reports where the residual does not help (full Hessian error unchanged on Transition1x), which increases my confidence in the rest of the numbers.

Critical weaknesses
  1. The "systematic error cancels" argument does not by itself establish that the model learns the true PES rather than the surrogate's PES. Minimizing ||H_φ(x̂1) − H_φ(x1)|| makes the generated structure indistinguishable from the reference under the frozen surrogate. If HIP is systematically wrong in a region of configuration space, the residual is smallest exactly where the surrogate is wrong in the same way, so the objective can be satisfied without approaching a genuine stationary point. The paper acknowledges in the Limitations that gains "do not place the generated structures at genuine stationary points", but the main text still over-attributes the reported improvements to "physical fidelity". Two concrete additions would resolve this: (i) evaluate the trained model with an independent Hessian/force provider not used in training (e.g. a different MLIP or DFT on a subset) and show the improvement transfers; (ii) report the residual value on reference structures, to show the floor is not an artifact of surrogate bias.

  2. The cost argument and the optimization path are under-specified, and the gains are inconsistent across datasets. The paper motivates directly-predicting Hessian models by their cheapness, but the residual must be differentiated with respect to the generator parameters, which means backpropagating through the frozen Hessian predictor at every step; the Limitations mention "one surrogate call per training step" but not the backward cost, which dominates. Empirically, the benefit is also uneven: on Transition1x the median ΔE moves only from 1.35 to 1.31 kcal/mol and the full Hessian error is unchanged, while the large gains are on Transition1x-TMC (Fmax median 46.74 → 36.58). Since the three datasets differ in level of theory (ωB97x/6-31G(d) vs. GFN2-xTB) and in the amount of surrogate finetuning, it is unclear whether the effect is driven by the curvature objective or by how well HIP covers the domain. A per-dataset statement of surrogate accuracy (e.g. HIP's own Hessian error on each test set) would let a reader judge this.

Questions for authors
  1. Can you evaluate the CurvFM-trained model with a Hessian/force provider that was not used in training (a different MLIP, or DFT on a few hundred test structures)? If the max-force and valid-TS gains survive, that would be strong evidence against the circularity concern above. If they do not, how should the results be interpreted?

  2. What fraction of the total training step time is attributable to the backward pass through HIP, relative to the base flow-matching step? Please give wall-clock numbers, since the "cheap enough to impose at every training step" claim rests on this.

  3. Why is the effect near-zero on Transition1x (ΔE 1.35 → 1.31 median, full Hessian error unchanged) but large on Transition1x-TMC? Is the difference explained by HIP's per-dataset surrogate accuracy, by molecular size, or by the level of theory used to generate references?

  4. What exactly is the "fraction of valid TSs" criterion (70% → 74% on Transition1x)? Is it the fraction with exactly one imaginary frequency, and is it computed with the same surrogate used in training or with an independent method?

  5. The residual compares Hessians at x̂1 and x1, but x̂1 is a single-sample estimate. How noisy is the estimate, and did you consider averaging over multiple endpoint samples or weighting by the endpoint's uncertainty?

2 comments

W1/Q1: Circularity of the surrogate-based residual

We thank the reviewer for this point and for the opportunity to clarify our evaluation protocol. All energies, forces, and Hessians reported in the paper, including the max-force and valid TS results, are computed with quantum chemistry at the reference level of each dataset, never with the training surrogate. The gains are therefore already measured with an independent Hessian/force provider not used in training. Appendix C.2.4 additionally repeats the Hessian evaluation with HIP. The gains are somewhat larger there, so part of the agreement is surrogate-specific, but the improvements under quantum chemistry as reported in the main text remain substantial. Regarding the residual floor, the residual vanishes at the reference by construction, so the relevant floor is the surrogate error against the quantum-chemical reference Hessians, reported for each dataset in Appendix A.2. We have revised Section 5.1 to state the evaluation protocol more explicitly (line 308-312).

W2/Q2: Cost of the backward pass through the surrogate

Appendix C.1 reports per-step timings on Transition1x-2p3p4p (batch size 16, single A100). A step takes 165 ms for the baseline and 310 ms for CurvFM. Of the additional 145 ms, the HIP forward pass accounts for 81 ms and the backward pass through HIP for 40 ms, i.e. about 24% of the base FM step. Over a full training run, this increases the training time by 30% (23 h to 30 h), and sampling is unaffected. We have revised the Limitations accordingly to state that the residual requires both a forward and a backward pass through the surrogate.

W2/Q3: Uneven gains across datasets

We thank the reviewer for this question and examined each of the three factors. First, results on Transition1x appear largely saturated, with recent models reaching very similar accuracy, which leaves little room for further improvement (Section 5.1 line 323-325). Second, surrogate accuracy is highest on Transition1x (Hessian MAE of 0.68 versus 1.12 and 1.01 kcal/mol/Ų on Transition1x-2p3p4p and Transition1x-TMC, Appendix A.2), so it does not account for the smaller gains there. Third, to test the role of the level of theory, we have added an experiment with a second TS generation architecture, GoFlow, on RDB7 (Appendix C.2.7). RDB7 uses a different DFT functional (ωB97X-D3/def2-TZVP) than the one the public HIP checkpoint was trained on, yet CurvFM still improves the energetic metrics, so the gains transfer across levels of theory. Molecular size and chemical diversity, in contrast, do appear to matter. The gains grow from Transition1x to Transition1x-TMC and are largest on the newly added Pt-cat set, a Pt(II)-catalyzed methane C–H activation dataset with systems of up to 45 atoms (Appendices A.1 and C.2.1).

Q4: Definition of the valid TS criterion

Yes, a valid TS has exactly one negative Hessian eigenvalue, computed with quantum chemistry at the reference level of the dataset and not with the surrogate (Section 5.1). We have added a new analysis under stricter criteria in Appendix C.2.3, reporting the joint validity with an additional maximum-force threshold.

Q5: Noise of the single-step endpoint estimate

For a given xtx_t and tt, x^1\hat{x}_1 is a deterministic estimate whose variance stems from sampling xtx_t and tt and averages out over training like the FM loss. Its uncertainty is accounted for by the weight tpt^p, which suppresses the residual at small tt, where x^1\hat{x}_1 is least accurate. A more accurate estimate further improves the results, as unrolling two ODE steps reduces the median barrier error by another 15% at a 57% higher training cost (Appendix C.2.8).

We thank you for taking the time to review our paper and for your constructive feedback. As the discussion deadline is approaching, we wanted to ask whether there are any remaining concerns or questions that we could address.

Comments on the paper

1 comment

Dear AC, SAC, and reviewers,

We thank all reviewers for their constructive feedback. The reviewers found CurvFM simple, cheap, and applicable to existing flow matching models without architectural changes or inference overhead (R8AC, Vmqs, 79FV), and appreciated the precise problem diagnosis (R8AC) and the gains in QM-evaluated physical metrics (79FV). Below we briefly summarize the main concerns and how we addressed them in the revised paper.

Practical significance and validity of the generated structures (Vmqs, 79FV). We added the chemical accuracy rate (barrier error below 1.58 kcal/mol), which CurvFM raises from 38.9 to 44.9% on Transition1x-2p3p4p and from 13.7 to 22.7% on Transition1x-TMC, as well as joint force and Hessian validity rates, where CurvFM is better at every threshold. We also added two real-world applications. On Pt(II)-catalyzed methane C–H activation with up to 45 atoms, CurvFM nearly halves the median barrier error and triples the chemical accuracy rate from 8 to 24%. In a combustion reaction network, it improves the Spearman correlation of the channel ranking from 0.75 to 0.89.

Value of Hessians compared with forces (Vmqs, 79FV). CurvFM now incorporates both surrogate-predicted forces and Hessians as training signals, making it, to our knowledge, the first framework to integrate surrogate PES information directly into the training objective of molecular generative models. As a new contribution, we benchmark force, curvature, and combined residuals across two surrogates on Transition1x-2p3p4p and Pt-cat. The two signals are complementary, and curvature improved the model in every setting, with a clear advantage on Pt-cat. A new theoretical analysis (Appendix B) shows that the Hessian error bounds the force and that the residual gradient aligns with the FM loss.

Cost and comparison with refinement (R8AC, Vmqs, 79FV). MLIP-driven refinement yields more accurate structures, but a single step already triples the cost per reaction and full refinement raises it about 40-fold. CurvFM keeps the inference cost of the backbone and increases training time by about 30%, with detailed timings in Appendix C.1.

Generality across architectures and levels of theory (R8AC, 79FV). We added GoFlow on RDB7 using the public HIP checkpoint at a different level of theory, where CurvFM again improves the energetic metrics. CurvFM now improves three backbones across two tasks.

In summary, the revision adds two real-world applications (Pt-cat and a combustion reaction network), a second TS architecture (GoFlow on RDB7), a benchmark of force, curvature, and combined residuals, a comparison with MLIP-driven refinement, stricter validity criteria including the chemical accuracy rate, a theoretical analysis of the residual, and extended implementation details.

Reviewer Vmqs acknowledged the new Pt-cat results and raised their score. Following their suggestion, we will present these results more prominently in the main text.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.