acceptodds
Under review as a conference paper at ICLR 2027

Bonding-Aware Representation Learning for Machine-Learning Force Fields

Abstract

Machine-learning force fields (MLFFs) provide a promising framework for modeling molecular energies and forces from atomic configurations. However, most existing MLFFs construct representations from local geometric neighborhoods defined by a fixed cutoff, treating nearby atom pairs primarily through their relative positions. Such representations do not explicitly encode bond connectivity or continuous changes in bond state during dynamics, e.g., reactive dynamics. Consequently, they may struggle to capture the associated variations in energy and force, particularly when the accompanying charge redistribution must satisfy global charge conservation. To address this challenge, we introduce a bond-aware representation framework that models local atom pairs as nodes in a bipartite bond-incidence graph. Each pair is assigned a structured state with orientation-invariant and orientation-equivariant components: the former captures latent bonding information, while the latter encodes signed charge-transfer flow. Aggregation through a signed incidence matrix produces orientation-independent atomic charges while exactly conserving the prescribed total charge. We evaluate our method on rMD17, MD22, Chignolin, Transition1x, and HORM, covering both non-equilibrium and reactive molecular dynamics. Across these benchmarks, our bond-aware model improves energy and force prediction accuracy. Stability analyses and assessments of energy drift further demonstrate the performance of our model in molecular dynamics simulations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine learning force field, molecular dynamics, atomistic simulation

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

This paper proposes BondFF, which augments a geometric graph with an atom–bond-state bipartite graph to explicitly represent continuously varying bonding states and charge redistribution. The method uses orientation-invariant and orientation-equivariant bond states, constructs strictly conserved atomic charges through a signed incidence matrix, and decomposes the residual energy into bond, charge-response, and electrostatic energies. The authors report improved prediction accuracy on several molecular and reaction datasets and provide dynamics tests.

Critical strengths
  • Incorporating bonding changes and charge conservation into machine-learning force fields is well motivated.
  • The reported evaluations cover systems of different sizes and types.
  • The numerical results are generally strong.
Critical weaknesses
  • I appreciate the authors for uploading the code, but it does not appear to match the method described in the paper. Sections 3.1–3.2 describe virtual bond-state nodes, along with the aggregation, propagation, and updating of bond states through atom nodes. However, the submitted PairChargeModule primarily feeds radial features and optional features such as element types into a standard MLP, with two hidden layers in the default configuration. I do not see the atom–bond-state bipartite message passing described in the paper in this branch. Section 3.3, Equations (11)–(13), and Proposition 3.1 present directed charge-transfer flows and a charge-conserving readout as core contributions, but in the provided BondFF branch I could not find predictions of relevant properties. The energy readout also does not match the definition in Section 3.4. The README states that the root directory contains setup.cfg, pyproject.toml, MANIFEST.in, and LICENSE.md, and instructs users to run pip install -e .. However, all four files are missing from the provided directory, preventing installation as instructed.

  • The latent bond strengths and charges are not supervised or physically validated, so the chemical interpretation is currently unsupported.

  • The dynamics results do not establish the claimed reliability. The tests in Section 4.2 and Appendix E.3 focus mainly on a single held-out Transition1x molecule. In an NVT trajectory with a target temperature of 300 K, BondFF produces an average temperature of 395.5 K. Although this is closer to the target than the other listed methods, it is still a substantial deviation and undoubtedly indicates an error in the simulation. I do not consider this an acceptable result to present in a paper.

  • I do not think the existing ablations adequately explain why the method works. Appendix E.4 contains parameter sensitivity analyses, and Appendix E.5 and Table 7 report results from removing each of the three energy terms and from two structural variants. These comparisons are valuable. However, removing energy terms one at a time mainly establishes their roles within the full configuration; it does not adequately distinguish the benefit of an additional energy correction from the benefits of explicit bond-strength and charge modeling. The two structural variants also change several design choices at once. The orientation-independent variant removes the orientation initialization and all associated components, while Full BondFF removes the geometric backbone and randomly initializes atom features. The current results therefore make it difficult to separate the contributions of bipartite graph propagation, bonding representations, and the charge-conservation mechanism. A directly relevant control would compare a standard pairwise MLP energy correction with the full method, using the same backbone and similar parameter counts. This comparison is particularly necessary because the provided code closely resembles the former model, while several core mechanisms ablated in Table 7 are absent from that implementation.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer YPAX

We thank the reviewer for the thoughtful assessment. We appreciate the recognition that incorporating bonding changes and charge conservation into machine-learning force fields is well motivated, as well as the positive assessment of the evaluation across diverse systems and the numerical results. Below, we address each concern and provide clarifications and additional results, which have been incorporated into the revised manuscript and highlighted in blue.

Response to Weakness 1 [Code–paper consistency and installation]

I appreciate the authors for uploading the code, but it does not appear to match the method described in the paper. Sections 3.1–3.2 describe virtual bond-state nodes, along with the aggregation, propagation, and updating of bond states through atom nodes. However, the submitted PairChargeModule primarily feeds radial features and optional features such as element types into a standard MLP, with two hidden layers in the default configuration. I do not see the atom–bond-state bipartite message passing described in the paper in this branch. Section 3.3, Equations (11)–(13), and Proposition 3.1 present directed charge-transfer flows and a charge-conserving readout as core contributions, but in the provided BondFF branch I could not find predictions of relevant properties. The energy readout also does not match the definition in Section 3.4. The README states that the root directory contains setup.cfg, pyproject.toml, MANIFEST.in, and LICENSE.md, and instructs users to run pip install -e .. However, all four files are missing from the provided directory, preventing installation as instructed.

Response: Because the paper had not yet been published, we initially chose not to release the full implementation and uploaded an incomplete code branch. We have now uploaded complete scripts implementing the framework described in the paper. We have also added the missing installation files and updated the README.

Response to Weakness 2 [Physical validation of latent quantities]

The latent bond strengths and charges are not supervised or physically validated, so the chemical interpretation is currently unsupported.

Response: We agree that the latent bond strengths and charges are not physically validated, and we do not claim that they reproduce physical bond strengths or atomic charges. Our objective is to construct a predictive force field; the reported evaluations therefore focus on energy and force predictions.

To illustrate how the learned quantities respond to changing geometries, we have added an exploratory analysis of pair-level responses along Transition1x test reaction paths. In the illustrated examples, several responses vary for pairs with large distance changes, while those of pairs with small distance changes remain comparatively stable. These results suggest that the learned representations capture diverse responses to geometric changes, without establishing a physical interpretation of the latent quantities. The revised manuscript includes the corresponding figures and discussion.

AuthorsOpenReview ↗

[2/2] Response to Reviewer YPAX

Response to Weakness 3 [Molecular dynamics reliability]

The dynamics results do not establish the claimed reliability. The tests in Section 4.2 and Appendix E.3 focus mainly on a single held-out Transition1x molecule. In an NVT trajectory with a target temperature of 300 K, BondFF produces an average temperature of 395.5 K. Although this is closer to the target than the other listed methods, it is still a substantial deviation and undoubtedly indicates an error in the simulation. I do not consider this an acceptable result to present in a paper.

Response:

We would like to clarify that the single held-out Transition1x molecule was used for the illustrative MD trajectory. The quantitative evaluations in Tables 1–3 were conducted on the complete test sets, not on that single molecule.

The reported average was calculated over a 100 ps trajectory. We conjecture that the deviation may be due to insufficient equilibration. We therefore re-evaluated BondFF and the baselines with an explicit equilibration phase.

The new figure shows that BondFF has a mean temperature of 298.9 K over the displayed 100 ps trajectory, closest to the target among the methods compared. We have included this figure in the revised paper and moderated our claims accordingly.

To assess stability across a broader set of held-out systems, we evaluate the stability metric with MD simulation on AIMD-Chig, 3BPA, th Transition1x testing set, and HORM testing set. The table reports stability time in ps.

MethodAIMD-Chig3BPATransition1xHORM
MACE59.8710045.8051.05
EquiformerV281.6510052.4462.24
PaiNN39.7948.8734.8639.08
GemNet-OC23.4959.3239.3836.03
Neural P³M74.0510038.2441.96
BondFF89.1210058.4168.50

BondFF has the longest reported stability time on AIMD-Chig, Transition1x, and HORM, and matches the longest reported time on 3BPA.

Response to Weakness 4 [Ablations and matched pairwise control]

I do not think the existing ablations adequately explain why the method works. Appendix E.4 contains parameter sensitivity analyses, and Appendix E.5 and Table 7 report results from ...

A directly relevant control would compare a standard pairwise MLP energy correction with the full method, using the same backbone and similar parameter counts ...

Response: Our existing ablations in Appendix E.5 and Table 7 remove each of the bond, charge, and electrostatic energy terms in turn. Relative to the default model, each removal worsens the reported errors across HORM, Transition1x, and Chignolin, with the size of the effect depending on the metric and dataset. The orientation-agnostic and Full BondFF variants also test broader structural changes, while Appendix E.4 examines parameter sensitivity.

To alleviate the reviewer's concern, we added a capacity-matched pairwise MLP energy correction using the same MACE backbone. Full BondFF achieves lower errors across all seven metrics. We also evaluated a generic pair-state control and variants retaining only the bond, charge-response, or electrostatic term. Full BondFF achieves the lowest error on every reported metric across these comparisons. Together, these controls show that the full combination of bonding and charge-related components yields the strongest performance and provide a more focused assessment of their contributions.

ModelDescriptionHORM energyHORM forceHORM HessianTransition1x energyTransition1x forceChignolin energyChignolin force
Backbone only (MACE)No correction branch0.1480.0980.1920.0710.0831.0270.389
Backbone + pairwise MLPGeneric pairwise energy correction, matched in capacity to BondFF0.1490.1020.1980.0830.0851.2900.498
Generic pair-state controlGeneric pair-state control0.8390.5820.9830.4820.4281.2930.375
Bond onlyBond term only0.1650.0910.2250.1250.0990.9040.322
Charge response onlyCharge-response term only0.1530.1130.1790.1130.0960.9070.341
Electrostatic onlyElectrostatic term only0.1530.1310.2070.1260.0870.6090.241
Full BondFF (default)All proposed components0.1290.0510.0970.0630.0810.5640.153
Reviewer_QkaJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes BONDFF, a plug-in branch for machine-learning force fields. A bipartite bond-incidence graph introduces latent pair states with orientation-invariant and orientation-equivariant components. Signed incidence aggregation yields atomic charges with exactly conserved total charge, and bond, charge-response, and screened-electrostatic terms correct the backbone energy. Evaluation covers six molecular benchmarks, paired backbone comparisons, cross-dataset transfer, and molecular-dynamics diagnostics, including energy, force, and Hessian prediction.

Critical strengths

The paired backbone comparisons provide useful evidence that the branch improves prediction. The comparisons use the same splits and training budgets, with two fewer backbone layers in the augmented models. All three backbones with available paired accuracy results improve on the reported targets; vanilla ViSNet fails during training. These comparisons are more informative than the heterogeneous main-table baselines. Total charge is conserved by construction. The signed incidence matrix gives exact conservation and independence from arbitrary edge orientation without a soft conservation penalty. The evaluation also covers reactive geometries, temperature shifts, transfer, and MD diagnostics.

Critical weaknesses

The physical motivation is not sufficiently connected to the learned representation. The proposed bond-state bottleneck is not directly diagnosed, and the method combines latent pair features with physically motivated energy terms. The paper appropriately states that the learned bond-strength variable is not an identified population-analysis bond order. Appendix C.7 additionally notes non-uniqueness of the flow representation on cyclic graphs; this concerns flow identifiability, not the usefulness of every latent feature. Appendix G also acknowledges that the learned charges have not been independently validated as physical charge estimates. My concern is that the experiments do not establish what specifically is gained from bond-state semantics beyond an additional interaction pathway and energy readouts. Reference-based analysis of the latent variables or reaction-coordinate-stratified errors would make the interpretation more convincing. The empirical gains are not sufficiently isolated from architectural and protocol choices. The paired backbone results and Table 7 provide useful evidence: removing individual energy terms or orientation-aware components worsens accuracy. However, these removals do not replace a generic pair-state branch matched in parameters, receptive field, and energy-readout capacity, leaving the specific value of the proposed semantics unresolved. Appendix D.2 explains that broader baselines retain heterogeneous settings and that some electronic models are adapted without their original auxiliary inputs or supervision. Those comparisons characterize the restricted common task, rather than the full published capabilities of every model. The paper reports means over ten runs without corresponding variability, limiting assessment of smaller differences. Energy/force/Hessian errors, transfer results, and MD diagnostics establish predictive benefits, but do not resolve this attribution question.

Questions for authors

Adequacy of the motivation: The paper motivates the method through the lack of explicit bond-state and charge-redistribution modeling in existing approaches. Could you provide more direct analysis showing how this limitation manifests in the studied systems and where the proposed mechanisms address it? Improvements in aggregate prediction errors do not yet clearly connect this motivation to the observed benefits.

Novelty and effectiveness of the three mechanisms: The three energy terms draw on established physical forms, making it important to clarify where the methodological novelty lies. Table 7 already provides leave-one-out ablations. Could you further explain why each mechanism helps, whether they provide complementary benefits, and why their combination requires the proposed representation? Targeted combination controls would be helpful where the existing ablations cannot distinguish these factors.

2 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer QkaJ

We thank the reviewer for the careful assessment. We appreciate the recognition of the paired backbone comparisons, the exact charge conservation and orientation independence provided by the signed incidence matrix, and the evaluation across reactive geometries, temperature shifts, transfer, and molecular dynamics. Below, we address each concern and provide clarifications and additional results, which have been incorporated into the revised manuscript and highlighted in blue.

Response to W. 1 and Q. 1 [Physical motivation and latent representations]

The physical motivation is not sufficiently connected to the learned representation. The proposed bond-state bottleneck is not directly diagnosed, and the method combines latent pair features with physically motivated energy terms. The paper appropriately states that the learned bond-strength variable is not an identified population-analysis bond order. Appendix C.7 additionally notes non-uniqueness of the flow representation on cyclic graphs; this concerns flow identifiability, not the usefulness of every latent feature. Appendix G also acknowledges that the learned charges have not been independently validated as physical charge estimates. My concern is that the experiments do not establish what specifically is gained from bond-state semantics beyond an additional interaction pathway and energy readouts. Reference-based analysis of the latent variables or reaction-coordinate-stratified errors would make the interpretation more convincing.

Adequacy of the motivation: The paper motivates the method through the lack of explicit bond-state and charge-redistribution modeling in existing approaches. Could you provide more direct analysis showing how this limitation manifests in the studied systems and where the proposed mechanisms address it? Improvements in aggregate prediction errors do not yet clearly connect this motivation to the observed benefits.

Response: We thank the reviewer for this suggestion. BondFF introduces continuously updated pair states to represent environment-dependent interactions during bond formation and dissociation. These states exchange information through shared atoms and support a charge-redistribution readout that conserves total charge by construction.

This design is motivated by the reactive configurations in Transition1x. Our endpoint analysis shows that all 287 test reactions contain at least one pair with a change in geometric bond status, with a median of three changed pairs per reaction.

SplitReactionsReactions with bond-status changeMedian changed pairsMedian maximum pair-distance change
Train9,561100%32.28 Å
Validation225100%32.41 Å
Test287100%32.29 Å

Bond status is defined using a distance threshold of 1.2 times the sum of covalent radii and serves as a geometric proxy for changing pair interactions. The maximum distance change is computed over all atom pairs between the reaction endpoints.

Besides, we have added an exploratory analysis of pair-level responses along Transition1x test reaction paths. In the illustrated examples, several responses vary for pairs with large distance changes, while those of pairs with small distance changes remain comparatively stable. We would like to clarify that these results suggest that the learned representations capture diverse responses to geometric changes, without establishing a physical interpretation of the latent quantities. The revised manuscript includes the corresponding figures and discussion.

AuthorsOpenReview ↗

[2/2] Response to Reviewer QkaJ

Response to W. 2 and Q. 2 [Attribution, baselines, and variability]

The empirical gains are not sufficiently isolated from architectural and protocol choices. The paired backbone results and Table 7 provide useful evidence: removing individual energy terms or orientation-aware components worsens accuracy. However, these removals do not replace a generic pair-state branch matched in parameters, receptive field, and energy-readout capacity, leaving the specific value of the proposed semantics unresolved. Appendix D.2 explains that broader baselines retain heterogeneous settings and that some electronic models are adapted without their original auxiliary inputs or supervision. Those comparisons characterize the restricted common task, rather than the full published capabilities of every model. The paper reports means over ten runs without corresponding variability, limiting assessment of smaller differences. Energy/force/Hessian errors, transfer results, and MD diagnostics establish predictive benefits, but do not resolve this attribution question.

Novelty and effectiveness of the three mechanisms: The three energy terms draw on established physical forms, making it important to clarify where the methodological novelty lies. Table 7 already provides leave-one-out ablations. Could you further explain why each mechanism helps, whether they provide complementary benefits, and why their combination requires the proposed representation? Targeted combination controls would be helpful where the existing ablations cannot distinguish these factors.

Response:

Our contribution is the design of incidence-based propagation, orientation-aware flows, and a conserved charge readout combined with three energy terms; the individual physical energy forms are established. The bond term provides a pair-state energy readout, the charge-response term models local charge dependence, and the screened electrostatic term couples charges over a longer finite range. These are intended roles, rather than independently established explanations of the gains. For the variability, as an example, the standard deviations of the HORM energy error over ten runs are:

GemNet-OCPaiNNDimeNet++SpookyNet4G-HDNNPCHGNetAIMNet2-NSEeMLPDPA-2BondFF
Standard deviation0.0030.0020.0040.0030.0030.0040.0030.0220.0030.002

The standard deviations are relatively small for most models. We omitted them from the manuscript tables to preserve readability. We provided the standard deviations for the main results in the appendix.

To address the concern about added pairwise capacity, we compare BondFF with a generic learned pair-state control. This control uses a neural network to construct learned features for candidate atom pairs and maps those features to an energy correction, without assigning them the proposed bond or charge semantics.

We also evaluate single-term and two-term energy configurations. Default BondFF achieves lower MAE than the generic control and all tested energy-term and structural variants on the seven reported metrics. The ablation study shows that the full model consistently achieves the lowest errors across all datasets and metrics, supporting the contributions of its energy terms and structural design.

Model or energy termsHORM energyHORM forceHORM HessianTransition1x energyTransition1x forceChignolin energyChignolin force
Bond + charge response0.1480.0610.1490.0890.0910.8860.284
Bond + electrostatic0.1380.0760.1720.0990.0850.5780.191
Charge response + electrostatic0.1340.0970.1240.0850.0820.5850.201
Orientation-agnostic variant0.2830.0840.1040.1540.1940.7250.283
Full BondFF structural variant0.2420.1240.1930.1840.1180.6190.203
Generic pair-state control0.8390.5820.9830.4820.4281.2930.375
Bond only0.1650.0910.2250.1250.0990.9040.322
Charge response only0.1530.1130.1790.1130.0960.9070.341
Electrostatic only0.1530.1310.2070.1260.0870.6090.241
Default BondFF0.1290.0510.0970.0630.0810.5640.153

Values are MAEs; lower is better. HORM and Transition1x energy and force errors are in eV and eV/Å, respectively; HORM Hessian errors are in eV/Ų. Chignolin energy and force errors are in kcal/mol and kcal/(mol·Å), respectively.

Reviewer_F1LqOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BONDFF, a machine learning force field that augments geometric graph networks with a bipartite graph containing atoms and latent bond states. The pair representations separate invariant bonding features from antisymmetric charge flows, ensuring conservation of the total charge. Bond, charge response, and screened electrostatic terms contribute to a differentiable potential used to predict energies, forces, and Hessians. Experiments across six molecular benchmarks report improvements in prediction accuracy and selected molecular dynamics diagnostics.

Critical strengths
  1. Separating symmetric bonding features from antisymmetric charge flows provides a coherent representation of pairwise interactions. Conservation of the total charge and independence from arbitrary bond orientations are enforced by construction.
  2. Experiments cover small molecules, larger molecular systems, temperature shifts, reactive configurations, and Hessian prediction. Paired backbone comparisons and component ablations provide evidence that the proposed branch benefits multiple architectures.
Critical weaknesses
  1. DPA-2 explicitly maintains and updates representations of atom pairs, making it relevant to the paper’s motivation, yet it is neither discussed nor evaluated. A comparison under matched data and training budgets would help establish whether BONDFF’s gains arise from its specific charge conservation design rather than richer pair representations more generally.
  2. The stability analysis primarily considers a single molecule, and the reported mean temperature of 395.5 K remains substantially above the 300 K NVT target. Smaller bond length deviations and a mean energy change near zero are insufficient to establish accurate dynamics. Complete simulation settings, multiple independent trajectories, and comparisons with reference dynamical observables are needed to substantiate the reliability claims.
Questions for authors

Why were Deep Potential models such as DPA-2 omitted from the comparisons and related work discussion? DPA-2 explicitly maintains and updates atom-pair representations, making it relevant to the proposed motivation. Could you clarify the distinction between these representations and BONDFF’s bond states, and provide a comparison under matched training data and computational budgets, or explain any technical limitations preventing such a comparison?

3 comments
AuthorsOpenReview ↗

Response to Reviewer F1Lq

Response to Reviewer F1Lq

We thank the reviewer for the thoughtful assessment. We appreciate the recognition of the coherent separation of bonding features and charge flows, the construction that ensures charge conservation and orientation independence, and the breadth of the evaluation across molecular systems and prediction tasks. Below, we address each concern and provide clarifications and additional results, which have been incorporated into the revised manuscript and highlighted in blue.

Response to W. 1 and Q. 1 [DPA-2 and pair representations]

DPA-2 explicitly maintains and updates representations of atom pairs, making it relevant to the paper’s motivation, yet it is neither discussed nor evaluated. A comparison under matched data and training budgets would help establish whether BONDFF’s gains arise from its specific charge conservation design rather than richer pair representations more generally.

Why were Deep Potential models such as DPA-2 omitted from the comparisons and related work discussion? DPA-2 explicitly maintains and updates atom-pair representations, making it relevant to the proposed motivation. Could you clarify the distinction between these representations and BONDFF’s bond states, and provide a comparison under matched training data and computational budgets, or explain any technical limitations preventing such a comparison?

Response:

We omitted DPA-2 initially because its published results rely on large-scale multitask pretraining, which would confound a direct comparison through differences in training data, supervision, and computational budget.

DPA-2 updates invariant and equivariant pair features as geometric representations used to construct atom-centered energy descriptors. BondFF instead uses explicit virtual bond-state nodes on an atom–bond incidence graph to generate signed charge-transfer flows. The resulting charges, q=qref+Bτq=q^{\mathrm{ref}}+B\tau, satisfy the prescribed total charge exactly.

We have now added DPA-2 (training from scratch with the same training data and computational budgets) results to the benchmark comparison and revised the manuscript accordingly:

ModelHORM energyHORM forceHORM HessianTransition1x energyTransition1x force3BPA 300 K energy3BPA 600 K energy3BPA 1200 K energy
DPA-20.2850.0930.1060.1810.1290.2830.8162.752
BondFF0.1290.0510.0970.0630.0810.1290.6291.903

HORM models use Hessian supervision. Energy, force, and Hessian MAEs are in eV, eV/Å, and eV/Ų, respectively. For 3BPA, models are trained on 300 K configurations and evaluated at the indicated temperatures.

Response to Weakness 2 [Molecular dynamics reliability]

The stability analysis primarily considers a single molecule, and the reported mean temperature of 395.5 K remains substantially above the 300 K NVT target. Smaller bond length deviations and a mean energy change near zero are insufficient to establish accurate dynamics. Complete simulation settings, multiple independent trajectories, and comparisons with reference dynamical observables are needed to substantiate the reliability claims.

Response:

All models were evaluated using Langevin NVT dynamics at a target temperature of 300 K, with a 1.0 fs timestep, 100 ps in total, a friction coefficient of 0.01 fs$^{-1}$, and double precision. Trajectories and observables were recorded every 100 steps.

We acknowledge that the measured temperatures exceed the target. This deviation may reflect incomplete equilibration or finite-sampling fluctuations. Our model yields the temperature closest to the target.

To address this concern, we performed an additional simulation after equilibration to 300 K. In the resulting 100 ps trajectory, BondFF has a mean temperature of 298.9 K, compared with the 300 K target. It is also the closest to the target among the methods shown in the new figure. We have included the new temperature figure in the revised paper.

We further expanded our MD simulations to include AIMD-Chig and 3BPA, the Transition1x test set, and HORM. The table below reports the simulation stability time in picoseconds (ps).

MethodAIMD-Chig3BPATransition1xHORM
MACE59.8710045.8051.05
EquiformerV281.6510052.4462.24
PaiNN39.7948.8734.8639.08
GemNet-OC23.4959.3239.3836.03
Neural P³M74.0510038.2441.96
BondFF89.1210058.4168.50
Reviewer_F1LqOpenReview ↗

Thank you for the additional experiments.

Thank you for the additional comparisons and molecular dynamics experiments. These provide useful evidence and partially address my concerns.

The comparison with DPA-2 trained from scratch under matched data and compute budgets is valuable for assessing the architecture in a controlled setting. However, it does not establish BONDFF’s competitiveness with strong pretrained potentials in practical applications. In particular, the results do not show whether BONDFF can match or exceed pretrained DPA-2 through scaling. A comparison under a realistic deployment would help establish the practical significance of the proposed approach beyond the controlled comparison.

The expanded dynamics experiments and the improved temperature control after equilibration alleviate some of my concerns. However, the stability criterion and the number of independent trajectories remain unclear. Moreover, the new results primarily demonstrate numerical stability rather than accurate molecular dynamics, as comparisons with reference dynamical observables are still absent. The concerns regarding the original NVE energy analysis also remain insufficiently addressed.

Overall, I appreciate the additional evidence, but it is not yet sufficient to resolve my concerns about practical competitiveness and dynamical reliability. I therefore maintain my original score.

AuthorsOpenReview ↗

Response to Reviewer F1Lq

Thank you for the follow-up and for recognizing the additional comparisons and MD experiments.

Our comparison evaluates task-specific training under matched data and computational budgets, where BondFF achieves lower errors on the reported metrics. This is a practically relevant setting, including when a force field must be trained on proprietary data that cannot be shared for pretraining. Task-specific training and pretraining are complementary approaches, and task-specific evaluation remains valuable and practical in its own right, with a variety of prior works (most of our baselines are not pretraining models). We did not claim that scaling BondFF would match or exceed pretrained DPA-2, and pretraining analysis is a direction for future work.

To address the concern that the original stability analysis covered only one molecule, we expanded the evaluation to AIMD-Chig, 3BPA, Transition1x, and HORM. Following Fu et al.’s molecular stability diagnostic [1], we report the time until the maximum absolute deviation of a monitored bond length from its reference value exceeds 0.5 Å. Runs that do not cross the threshold within 100 ps are reported as 100 ps. For AIMD-Chig and 3BPA, we performed 10 separate runs initialized from randomly selected test frames; for 3BPA, these were drawn from the 300 K test set. For each of Transition1x and HORM, we initialized simulation runs from 100 randomly selected molecular configurations.

We agree that a mean energy change near zero does not establish energy conservation or accurate dynamics and we did not claim our method produced accurate dynamics. This energy analysis also followed prior evaluation protocol [2]. Our primary contribution is a force-field architecture evaluated through energy, force, and Hessian prediction. Following prior work [1, 2], the MD experiments provide an additional assessment of stability and energy drift, rather than a claim that the simulations reproduce accurate dynamics.

We acknowledge the value of comparisons with reference dynamical observables. However, MLFF test splits generally do not provide complete, sufficiently long reference trajectories. Generating such trajectories at ab initio fidelity would be computationally demanding and is beyond the scope of this work.

To provide an additional reference-based comparison to alleviate reviewer's concern, we evaluated the radius of gyration RgR_g, its distribution, pairwise all-atom RMSD, mean all-atom RMSF, and maximum per-atom RMSF across 25 AIMD-Chig sub-trajectories. Each contains 200 frames, corresponding to approximately 0.2 ps, the maximum continuous length available per trajectory in the dataset. BondFF achieves the lowest RgR_g distribution JSD and the closest RMSD and RMSF values to the reference among the tested models.

MethodRgR_g (Å)Gyration radius JSDPairwise all-atom RMSD (Å)Mean all-atom RMSF (Å)Maximum per-atom RMSF (Å)
Reference data6.726 ± 1.182–0.463 ± 0.0220.332 ± 0.0171.039 ± 0.117
MACE6.689 ± 1.1950.467 ± 0.1520.349 ± 0.0190.242 ± 0.0120.768 ± 0.166
EquiformerV26.671 ± 1.2090.508 ± 0.1780.346 ± 0.0200.242 ± 0.0130.730 ± 0.181
PaiNN6.663 ± 1.2950.619 ± 0.1630.350 ± 0.0260.242 ± 0.0170.713 ± 0.159
GemNet-OC6.658 ± 1.1730.602 ± 0.1980.347 ± 0.0220.243 ± 0.0170.716 ± 0.150
Neural P3M6.693 ± 1.1910.496 ± 0.1630.353 ± 0.0190.246 ± 0.0130.763 ± 0.151
BondFF6.739 ± 1.2030.456 ± 0.1820.355 ± 0.0230.247 ± 0.0160.778 ± 0.161

[1] Xiang Fu, Zhenghao Wu, Wujie Wang, Tian Xie, Sinan Keten, Rafael Gomez-Bombarelli, and Tommi Jaakkola. Forces are not Enough: Benchmark and Critical Evaluation for Machine Learning Force Fields with Molecular Simulations. Transactions on Machine Learning Research, 2023.

[2] Ishan Amin, Sanjeev Raja, and Aditi Krishnapriyan. Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians. International Conference on Learning Representations, 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, and Area Chairs,

We sincerely appreciate the committee’s time and effort in reviewing our work.

Our paper proposes BondFF, a machine-learning force field that combines bond-state representations with an exactly charge-conserving readout for reactive molecular systems. We thank Reviewers F1Lq, QkaJ, and YPAX for their constructive feedback and their recognition of the model’s physical motivation, architectural design, and broad evaluation.

We summarize the main points of our rebuttal below.

Expanded Discussion and Clarifications

  • Novelty and related work: We clarified the distinction between DPA-2’s geometric pair representations and BondFF’s bond-state nodes and charge-transfer flows. We also clarified that our contribution lies in combining incidence-based propagation, orientation-aware flows, and a conserved charge readout with established energy forms.
  • Physical interpretation: We clarified that the learned bond strengths and charges have not been independently validated as physical bond orders or atomic charges. Our exploratory reaction-path analysis illustrates their responses to geometric changes.
  • Molecular dynamics claims: We acknowledged that the originally reported mean temperature of 395.5 K was substantially above the 300 K target and moderated our claims. The MD experiments assess stability and energy drift.
  • Implementation and reproducibility: We uploaded complete scripts implementing the framework described in the paper, added the missing installation files, and updated the README.

Additional Experiments and Results

  • DPA-2 comparison: We added DPA-2 trained from scratch with matched data and computational budgets. BondFF achieves lower errors on all reported HORM, Transition1x, and 3BPA metrics in this task-specific comparison.
  • Matched controls and ablations: We added a capacity-matched pairwise MLP using the same MACE backbone, a generic pair-state control, and single- and two-term energy configurations. BondFF achieves the lowest errors across the seven reported metrics among these controls and the tested structural variants.
  • Reactive configurations and variability: All 287 Transition1x test reactions have at least one change in geometric bond status between endpoints, with a median of three changed pairs. We also added standard deviations for the main results.
  • Molecular dynamics: After explicit equilibration, BondFF reaches a mean temperature of 298.9 K in the displayed 100 ps trajectory. In expanded stability tests, BondFF achieves the longest reported stability time on AIMD-Chig, Transition1x, and HORM, and ties for the longest on 3BPA.
  • Reference-based observables: On short AIMD-Chig sub-trajectories, BondFF achieves the lowest radius-of-gyration distribution divergence and the closest reported RMSD and RMSF values to the reference among the tested models.

We hope these clarifications and additional results help the committee evaluate the contributions of our work.

Best regards,
Authors of Submission 5788

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.