acceptodds
Under review as a conference paper at ICLR 2027

Duplicated-Graph Training for Machine Learning Force Fields

Abstract

Machine learning force fields now approach quantum-mechanical accuracy at a small fraction of its cost, yet every training label still requires an electronic-structure calculation, so how much a model learns from each label determines how far it reaches. Nearly all such models share one interface to their data: a labeled conformation is converted into a single cutoff graph, and that graph is everything the network ever sees of it. We revisit that interface with duplicated-graph training, which pairs each labeled conformation with a rigidly translated copy of itself and applies one cutoff rule to both, so that atoms interact within each copy and across the two. Rigidity leaves every intramolecular distance and angle intact, so the reference energy and forces remain exact labels, while the geometry between the copies varies with the translation. Resampling the translation therefore lets a single label constrain the model on a family of interaction geometries, with no additional electronic-structure calculations and no change to the architecture. Applied to MACE, duplicated graphs lower force error on every system we study: by 16.4% on average across the ten rMD17 molecules and by 9 to 20% on the seven MD22 systems. The gains are largest where the test conformations depart from the training distribution, growing on 3BPA from 4.57% at 300 K to 16.55% at 1,200 K. Distillation transfers the improvement to a model that runs on an ordinary graph at unchanged inference cost, retaining a 15.6% average force reduction on rMD17.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine learning force fields, duplicated-graph training, molecular graph augmentation

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_1pXVOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper (1) proposes 'duplicating the input molecular graph' as a way to augment the input information used by GNNs for molecular property prediction, and introduces (2) distillation approaches to help transfer the learned representations done in this way to a regular network. This work is not bad, but the idea and extent to which it is developed in the current form would rather suit a workshop paper as a 'neat idea' than a full ICLR submission. For a full submission, there should be reasonable evidence that this is not just a neat idea, and that doing this actually provides practical benefits for the materials modelling community. So I would recommend the following:

(1) Full training on large datasets, to show that the improvements hold in the regimes where people actually train models for MD. When it comes to MLIPs, I think this is almost required to show impact, because there are many architectural changes which show improvements on small-scale datasets but ultimately disappear at scale.

(2) Computational ablations to show that the accuracy gained remains under fixed compute cost. Eg, as it is, adding additional edges (in architectures like MACE) increases the training cost almost linearly with the number of added edges. This additional cost can go into extended training strategies, which makes this not a direct comparison.

^ If the authors manage to show the above, I would be impressed and consider it a suitable contribution as an ICLR main conference paper.

(Note: I did not read the appendices for my review)

Critical strengths

It's a very straightforward idea, and clearly explained. Small scale studies show a reasonable improvement.

Critical weaknesses

Size transferability seems to already be breaking down at the scales the authors test (MD17->MD22), where the duplicated or distilled models no longer show a consistent improvement over the baseline. I think that it's unclear how these benefits would transfer to even larger graphs, where the adjacency matrix becomes highly sparse.

The main weakness about the current form, though, which is also addressed, is the fact that duplicating the graph ultimately increases the training cost. I can think of other ways to improve the accuracy of MLIPs, if I could just increase the training cost.

Minor points:

There are a few sentences which simply don't say anything, like "Our results are consistent with a simple reading of the construction". What does this mean.

Questions for authors

Clarification requests:

(1) What is the practical conclusion that people training MLIPs can learn from this? The conclusion is somewhat vague about this and doesn't make any concrete recommendations. Should we be training on duplicated and translated versions of input graphs?

Other questions (just from curiosity):

(2) What is the relationship between the overall translation 't' and the improvements in accuracy, and was this studied empirically? In the extremes, 't' being too large doesn't add any additional information (because there's no more connections with the original graph), and t=0 is the same.

(3) How would this work for periodic materials (or can it work at all)?

4 comments

Response to Reviewer 1pXV (1/2)

We thank the reviewer for recognizing our work. We understand the concerns about large-scale applicability and computational cost, but disagree with some of the assessments. Below, we clarify and respond to them one by one using the original paper and the additional experimental results.

R1. Full training on large-scale datasets

A1. To examine whether the method remains effective beyond single-molecule, small-data benchmarks, we further performed experiments on SPICE (Eastman et al., 2023; approximately one million molecular configurations) and OMat24-1M (Barros-Luque et al., 2026; approximately one million periodic crystal configurations).

On SPICE, we compare the test results of the MACE-OFF23 Large model and MACE (DupG), as follows.

Table R1: SPICE test MAE. E is in meV/atom, and F is in meV/Å.

ModelE MAEF MAE
MACE-OFF23 Large0.7713.39
MACE (DupG)0.5312.40

We also observed positive test results on OMat24-1M.

Table R2: OMat24-1M test MAE. Both models were trained for 10 epochs. E is in meV/atom, and F is in meV/Å.

ModelE MAEF MAE
MACE78.69161.25
MACE (DupG)64.95154.46

Both MACE and MACE (DupG) completed 10 epochs on OMat24-1M. On the full official test set, DupG reduces Force MAE and Energy MAE by 4.21% and 17.45%, respectively, relative to MACE.

These additional experiments directly address the reviewer's concern that gains might disappear as the training-data scale increases: at the million-configuration training scale tested so far, DupG still achieves lower energy and force prediction errors, rather than being effective only on single-molecule benchmarks with a few hundred training configurations. The SPICE test results and the completed OMat24 test results further show that the method can extend from cross-molecular potential training to periodic crystals.

We have revised the manuscript to include the SPICE and OMat24-1M experiments in Appendix F.1, Tables A9–A10, and their training settings in Appendix F.1, Tables A11–A12.

R2. Accuracy improvements under a fixed computational budget

A2. We performed a cumulative edge processing budget-matched ablation on rMD17 malonaldehyde, directly examining whether additional edge-message-processing computation is sufficient to explain the gains of DupG.

Ordinary MACE processes approximately 360 directed edges per update, whereas DupG processes approximately 1,530. We therefore extended ordinary MACE training to 4.25 times the original number of updates, making the cumulative number of processed directed edges approximately equal for the two models.

Extending ordinary MACE training by 4.25 times leaves the post-SWA test Force MAE unchanged at 5.2 meV/Å, compared with 3.8 meV/Å for MACE (DupG).

The accuracy gains of DupG therefore cannot be attributed simply to additional edge-processing computation; spending the corresponding budget on longer ordinary MACE training does not yield the same improvement.

We have revised the manuscript to include this edge-processing-budget control in Section 5.5 and Appendix F.2, Table A13.

W1. Size generalization to larger molecular systems

A3. Table 3 of the original paper shows that MACE (DupG) reduces Force MAE on all seven MD22 systems, with relative improvements of 9.20%–20.43%, averaging 16.00%. MACE (Distill) also achieves lower Force MAE than the MACE baseline on all seven systems.

The seven MD22 systems span a broad range of 42–370 atoms, differ in molecular size by almost 9 times, and include complex systems containing hundreds of atoms. This is already a sufficiently broad range. Within this size range, the force-prediction gains do not show a systematic decline as molecular size increases.

In addition, we focus primarily on Force MAE because atomic forces directly determine accelerations during MD integration and are an important measure of MLFF applicability to dynamics. Energy MAE is also meaningful, so Table 3 reports both energy and force errors and faithfully presents their different improvement trends.

The existing MD22 experiments in the original paper therefore do not show that the force-prediction gains disappear as molecular size increases.

References

Eastman, P. et al. (2023). SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials. Scientific Data, 10, 11. https://doi.org/10.1038/s41597-022-01882-6

Barros-Luque, L. et al. (2026). The Open Materials 2024 (OMat24) inorganic materials dataset and models. Nature Computational Science, 6, 642–652. https://doi.org/10.1038/s43588-026-00996-w

Response to Reviewer 1pXV (2/2)

W2. Additional training cost and the method's practical value

A4. We agree that DupG increases training computation, but believe that this issue should be evaluated from three perspectives: label efficiency, the method's complementarity, and practical deployment cost.

First, our core objective is to improve Label Efficiency, rather than simply optimize Compute Efficiency. DupG uses additional training computation to improve accuracy under a fixed DFT-reference-label budget, without requiring additional expensive electronic-structure labels.

Second, increasing training computation to improve model generalization is a widely used strategy in machine learning, for example, Data Augmentation and Consistency Regularization. Its practical value depends on the resulting accuracy gains.

Furthermore, DupG is orthogonal and complementary to other accuracy-improvement strategies by design. Unlike methods that improve representation capacity through better equivariant network architectures, such as NequIP and MACE (Batzner et al., 2022; Batatia et al., 2022), or select labeled configurations through active learning, such as DP-GEN (Zhang et al., 2020), DupG changes the interaction-graph representation through which existing labeled configurations enter the network, without requiring changes to the backbone or additional reference labels. It therefore has the potential to be combined with other architectures and data-utilization strategies.

Finally, it is necessary to distinguish the one-time Training Cost from the recurring Inference Cost. MLFFs are usually trained offline and then repeatedly called over many MD time steps and trajectories. MACE (Distill) uses an ordinary N-atom graph without cross-copy edges, so offline training costs can be amortized over repeated MD inference calls.

In summary, the value of DupG lies not only in improving accuracy with additional training computation, but also in providing a training method that requires no new DFT labels, complements other training strategies, and avoids duplicated-graph inference overhead through distillation. We believe that label efficiency, training computational cost and long-term MD deployment efficiency should be considered together, rather than judging the method's practical value solely on the basis of additional training cost.

We have revised to make this training and deployment recommendation explicit in Section 7 (Conclusion).

Minor Comment

We have revised the manuscript and removed this sentence from Section 6.

Q1. Regarding the practical conclusion

A5. Our practical recommendation is that, if the goal is to train an MLFF with more accurate force predictions without adding DFT-reference labels, DupG is a training strategy worth adopting.

When duplicated-graph inference is unaffordable, we recommend deploying the ordinary-graph distilled Student, without incurring the additional duplicated-graph inference overhead. DupG can therefore concentrate the additional computation primarily in training, improving force-prediction accuracy while maintaining the inference efficiency of ordinary-graph models.

We have revised the manuscript to include this practical recommendation in Section 7 (Conclusion).

Q2. Relationship between the translation and the accuracy improvement

A6. During training, the translation vector tt is not fixed, but is randomly resampled according to the following distribution:

t∼N(0,σ2I3).t \sim \mathcal{N}(0,\sigma^2 I_3).

The relevant tunable hyperparameter is therefore the scale of the translation distribution, rather than a particular fixed translation vector. In Appendix C.1 (Figure A1) of the paper, we have already conducted an empirical ablation of σ\sigma on six rMD17 molecules.

The experiments show that:

  • When σ≤0.003\sigma \leq 0.003 Å, all six molecules perform worse than the baseline.
  • When σ\sigma is within 0.03–0.1 Å, Force MAE is lower than the baseline for all six molecules.
  • When σ\sigma is increased further, performance on some molecules begins to deteriorate.

When the translation becomes too large, cross-copy interactions disappear; when t=0t=0, the geometric diversity introduced by translation no longer exists.

Q3. Applicability to Periodic Materials

A7. The duplicated-graph construction can be extended to periodic systems by consistently handling periodic boundary conditions and the corresponding cell shifts in neighbor searches within and between copies.

As shown in Table R2 of response (1/2), after completing 10 epochs, MACE (DupG) achieves lower test energy and force errors than the MACE baseline.

We have revised the manuscript to include the periodic construction and experiment in Appendix F.1, Table A10, and the training settings in Appendix F.1, Table A12.

We believe these results warrant a reconsideration of the overall assessment of this work.

Dear Reviewer,

As the discussion period is coming to an end, we would like to kindly follow up on our rebuttal.

We have conducted substantial additional experiments addressing your main concerns, including million-configuration evaluations on SPICE and OMat24-1M, a compute-matched ablation, and further evidence of the method's practical applicability. We have also incorporated these results into the revised manuscript.

We would greatly appreciate it if you could let us know whether these additional results address your concerns, or if there are any remaining issues you would like us to clarify.

Thank you again for your time and consideration.

Dear Reviewer 1pXV,

As the rebuttal period is coming to an end, I kindly remind you to check the authors' responses. Please indicate whether your concerns have been adequately addressed. Many thanks!

Best,

AC

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

This paper proposes DupG, a method that improves the training of machine-learning potentials by duplicating molecular graphs. The authors place a molecule and a rigidly translated copy within the same cutoff graph, allowing the two to interact during message passing while retaining the original energy and force labels for supervision. Experiments with MACE show improved force prediction on rMD17, MD22, and 3BPA. Feature distillation also allows a student model operating on ordinary molecular graphs to retain some of these gains.

Critical strengths
  • Improving the use of existing labels by changing the training graph is a novel idea, and it requires neither additional quantum-chemical labels nor changes to the backbone architecture.
  • The improvements in force prediction span molecular systems of different sizes, with particularly clear gains on the high-temperature 3BPA test sets.
  • The distillation experiments further strengthen the method's practical value.
Critical weaknesses
  • All experiments concern molecular potentials. Applicability to complex systems such as periodic crystals, condensed phases, reactive systems, or universal potentials remains unclear.
  • Some improvements are relatively small, especially for MD22 in Table 3, and the paper does not report results over multiple random seeds.
  • No actual MD evaluation is provided. Static force errors alone cannot fully reflect the performance of MLFFs.
  • Although the author mentioned that this method requires additional cost, I believe that time cost, especially inference speed, is a very important metric for MLFFs.
Questions for authors
  • In Table 1-3, why can the distilled model generally outperform DupG?
  • Will the author release the code?
5 comments

Response to Reviewer TsZv (1/2)

We thank the reviewer for the constructive comments. We have added the following experimental evidence addressing the relevant concerns.

W1. Applicability beyond molecular potentials to complex systems

A1. To examine the applicability of DupG beyond isolated molecules, we further performed experiments on the official subset of the periodic inorganic materials dataset, OMat24-1M(Barros-Luque et al., 2026; approximately 1 million periodic crystal configurations).

Table R1: OMat24-1M test MAE. E is in meV/atom, and F is in meV/Å.

ModelE MAEF MAE
MACE78.69161.25
MACE (DupG)64.95154.46

The test Force MAE and Energy MAE of MACE (DupG) are 4.21% and 17.45% lower than those of MACE, respectively. These results provide empirical support for extending the method to periodic crystals.

We have revised the manuscript to include the periodic construction and experiment in Appendix F.1, Table A10, and the training settings in Appendix F.1, Table A12.

W2. MD22 improvements are small and multiple-random-seed experiments are missing

A2. We would like to further clarify the results in Table 4 of the revised manuscript: MACE (DupG) reduces Force MAE on all seven MD22 systems, with relative improvements of 9.20%–20.43%, averaging approximately 16.00%. These systems contain up to 370 atoms, and the measured force-prediction gains do not exhibit a systematic decline as molecular size increases.

We focus on Force MAE because MD integration directly uses atomic forces to determine atomic accelerations, making force-prediction accuracy an important measure of MLFF performance in dynamics applications. Energy MAE is also a necessary evaluation metric, and Table 4 faithfully reports its less uniform improvements.

In addition, we performed experiments with additional independent random seeds on three representative MD22 systems.

Table R2: Multi-seed test MAE (mean ± std) of MACE (DupG) on three MD22 systems. E is in meV, and F is in meV/Å.

MoleculeE MAE ↓F MAE ↓
Ac-Ala3-NHMe13.58 ± 0.227.49 ± 0.15
DHA24.90 ± 0.145.96 ± 0.09
AT-AT32.17 ± 0.996.08 ± 0.07

These results additionally illustrate the error variation of MACE (DupG) across independent training initializations.

We have revised the manuscript to include a summary of the training-initialization experiments and their evaluation settings in Appendix F.3, Table A14.

References

Barros-Luque, L. et al. (2026). The Open Materials 2024 (OMat24) inorganic materials dataset and models. Nature Computational Science, 6, 642–652. https://doi.org/10.1038/s43588-026-00996-w

Response to Reviewer TsZv (2/2)

W3. Lack of actual molecular dynamics evaluation

A3. To address this concern, we performed downstream MD evaluation on malonaldehyde, aspirin and ethanol, comparing MACE with MACE (Distill).

The experiments include 54 NVE/NVT and time-step-control trajectories, totaling 8.64 million integration steps. We extracted 180 configurations from the NVT trajectories generated by both models, independently computed their DFT-reference energies and forces, and cross-evaluated MACE and MACE (Distill) on the same configuration pool.

Table R3: DFT-reference Force MAE on the shared MD configuration pool. Units are meV/Å.

MoleculeMACEMACE (Distill)Reduction
Malonaldehyde4.072.9328.00%
Aspirin7.616.2318.16%
Ethanol2.101.6521.23%

This shows that the force-prediction accuracy gains from DupG can be retained after distillation and extend to configurations visited by the models during actual MD simulations, rather than being limited to the original static test set. The additional experiments therefore provide further evidence of the method's usability and prediction accuracy in practical molecular dynamics applications.

We have revised the manuscript to include the MD evaluation in Section 5.6, Table 6, with paired confidence intervals and simulation settings in Appendix F.4, Tables A15 and A16, respectively.

W4. Inference efficiency

A4. We agree that inference cost is an important consideration in practical MLFF deployment.

As stated in Section 4.3, MACE (Distill) uses the same backbone architecture and ordinary N-atom graph as the MACE baseline, without additional duplicated nodes or cross-copy edges at inference. Feature distillation therefore removes the additional duplicated-graph computation introduced by DupG, allowing the Student to retain the inference computational complexity of an ordinary-graph model.

Table R4: MD Step Runtime of Original MACE and Distilled Students

MoleculeOriginal MACE (ms/step)Distilled Student (ms/step)Student Runtime Change (%)
Benzene28.5129.29+2.73
Malonaldehyde23.7423.71−0.10
Naphthalene25.6125.75+0.54
Paracetamol31.4631.39−0.23
Salicylic acid27.7427.79+0.19
Toluene23.9323.86−0.32
Uracil28.1028.09−0.03

For inference directly using duplicated graphs, Table 5 of the revised manuscript already reports the trade-off between prediction accuracy and latency. The additional Teacher-training and distillation costs occur during offline training and do not need to be paid again at every MD step, so they can be amortized over a large number of downstream force-prediction calls.

We have revised the manuscript to make the ordinary-graph student deployment recommendation explicit in Section 7 (Conclusion).

Q1. Distilled model outperform the DupG Teacher

A5. The distilled Student does not merely imitate the Teacher's final predictions.

As described in Appendix B, MACE (Distill) is initialized from the trained Teacher's weights and learns its aggregated intermediate-layer representations, while still receiving direct supervision from the true DFT Energy/Force labels.

These representation targets may reduce the model's dependence on a single duplicated-graph representation, while further DFT supervision gives the Student an opportunity to correct some of the Teacher's prediction errors. The Student may therefore achieve better generalization than the Teacher.

Q2. Release the code

A6. Yes. We plan to release the source code, experimental configurations and training scripts after the paper is accepted.

Sections 3–5 and Appendices A–E of the current paper already provide detailed method definitions, training objectives, hyperparameters and evaluation protocols to support experimental reproducibility.

We have revised the manuscript's Reproducibility statement to point to Appendix F for the additional experimental details.

We thank the reviewer again for the constructive feedback and sincerely hope that these additional experiments and clarifications address the concerns raised.

I appreciate the authors’ response. However, my concerns have not been fully addressed. In particular, the current results remain insufficient to support the proposed mechanism.

First, the predictive performance for energies and forces does not improve consistently. On rMD17 (Table 1), DupG lowers force MAE on all ten molecules but raises energy MAE on six molecules. Energy degradation also appears on MD22, on 3BPA at 300 K, and in the EquiformerV3, Naph. and Parac. cases, where energy MAEs nearly double. The authors claim that force accuracy is important for molecular dynamics applications does not resolve this concern. The potential energy surface governs important properties and processes, such as energy conservation and transition-state pathways. A method claimed to provide augmentation for force field training should not improve force accuracy at the expense of energy accuracy without a clear account of the source and implications of this trade-off.

Second, the energy and force MAEs depend strongly on the relative weights in the loss function, but these weights are not held constant across the methods being compared (Table A2). This inconsistency complicates the interpretation of the reported performance differences and prevents a direct assessment of whether they arise from the proposed method or from differences in optimization objectives.

Finally, the distilled student model can outperform the DupG teacher model. Although the authors offer possible explanations for this result, these explanations appear applicable to teacher-student fine-tuning procedures in general. They still do not clarify the specific contribution of graph duplication to the improvement.

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

This paper proposes Duplicated-Graph Training, a data augmentation approach for machine learning force fields (MLFFs) that pairs each molecular conformation with a rigidly translated copy in a single cutoff graph. By preserving the original energy and force labels while varying cross-copy interactions, the method improves data efficiency without additional DFT calculations or architectural modifications. Experiments with MACE demonstrate consistent force prediction improvements across rMD17, MD22, and 3BPA, while knowledge distillation preserves these gains with standard inference costs.

Critical strengths
  • Introduces a graph augmentation strategy that creates diverse interaction geometries while preserving exact reference labels.
  • Demonstrates consistent force accuracy improvements across multiple benchmarks, particularly under out-of-distribution conformational shifts
  • Requires no architectural modifications or additional DFT calculations, and supports knowledge distillation for efficient inference
Critical weaknesses
  • Experiments exclusively use MACE, leaving the generalizability to other MLFF architectures unexplored, such as equiformerV2.
  • The paper focuses on energy and force prediction without evaluating downstream applications such as molecular dynamics simulations or geometry optimization. Only evaluating energy and force doesn’t necessarily mean we can have a better MLFF.
  • The benchmark has actually been a bit saturated and small. I recommend evaluating the model on more recent large-scale dataset, such as OMol25 [1] and PubChemQCR [2]

[1] https://arxiv.org/abs/2505.08762
[2] https://arxiv.org/pdf/2506.23008

Questions for authors

please refer to weakness part

4 comments

Response to Reviewer SqJF (1/2)

We thank the reviewer for the constructive comments. We have added new experimental evidence addressing the following concern.

W1. Generalization to other MLFF architectures

A1. We further extended duplicated-graph training and feature distillation to EquiformerV3 and GemNet, and evaluated all ten rMD17 molecules. The distilled models of both architectures achieve lower Force MAE than their respective baselines on all ten molecules.

Table R1: Test MAE of EquiformerV3 on ten rMD17 molecules. E is in meV, and F is in meV/Å.

MoleculeEquiformerV3 EDistill EEquiformerV3 FDistill F
Aspirin2.351.517.624.69
Azobenzene0.700.472.561.80
Benzene0.110.070.310.28
Ethanol0.440.282.791.69
Malonaldehyde0.740.565.013.63
Naphthalene0.230.381.240.84
Paracetamol1.312.545.043.42
Salicylic acid0.790.674.272.60
Toluene0.520.521.511.03
Uracil0.530.473.022.02
Mean0.770.753.342.20

Table R2: Test MAE of GemNet on ten rMD17 molecules. E is in meV, and F is in meV/Å.

MoleculeGemNet EDistill EGemNet FDistill F
Aspirin2.892.429.608.01
Azobenzene1.591.135.204.18
Benzene1.080.030.900.35
Ethanol0.810.473.872.85
Malonaldehyde1.421.626.125.07
Naphthalene0.640.302.191.66
Paracetamol2.471.897.235.86
Salicylic acid1.220.795.374.35
Toluene0.450.352.371.72
Uracil1.470.574.003.12
Mean1.400.964.693.71

Overall, EquiformerV3 (Distill) and GemNet (Distill) reduce the mean Force MAE by 34.10% and 20.72%, respectively. These results show that the accuracy gains from duplicated-graph training can be transferred to different MLFF architectures through feature distillation, rather than being limited to MACE.

We have revised the manuscript to describe the EquiformerV3 and GemNet experiments in Section 5.2, Table 2, and added the training configurations to Appendix B, Tables A4–A5.

W2. Lack of downstream application evaluation beyond static energy and force prediction

A2. To evaluate downstream applications in practice, we performed molecular dynamics simulations on malonaldehyde, aspirin and ethanol, comparing MACE with MACE (Distill).

The experiments include 54 NVE/NVT and time-step-control trajectories, totaling 8.64 million integration steps. We further sampled 180 configurations from the NVT trajectories generated by the two models, independently computed their DFT energies and forces, and evaluated both models on the same configurations to assess their accuracy on configurations actually visited during MD.

Table R3: DFT-reference Force MAE on the shared MD configuration pool. F is in meV/Å.

MoleculeMACEMACE (Distill)Reduction
Malonaldehyde4.072.9328.00%
Aspirin7.616.2318.16%
Ethanol2.101.6521.23%

The results show that MACE (Distill) reduces DFT-reference Force MAE by 28.00%, 18.16% and 21.23% on the three molecules, respectively. Therefore, the accuracy gains of DupG are not limited to static benchmarks: they can be retained in ordinary-graph models through distillation and extend to configurations actually visited during MD simulations. These results provide further experimental evidence of the method's usability and force-prediction accuracy in downstream molecular dynamics applications.

We have revised the manuscript to include the MD evaluation in Section 5.6, Table 6, with paired confidence intervals and simulation settings in Appendix F.4, Tables A15 and A16, respectively.

Response to Reviewer SqJF (2/2)

W3. Lack of evaluation on newer, larger-scale datasets

A3. To address this concern, we further extended the experiments to SPICE (Eastman et al., 2023; approximately 1 million molecular configurations) and OMat24-1M (Barros-Luque et al., 2026; approximately 1 million periodic crystal configurations), to examine applicability to large-scale cross-molecular training and periodic materials, respectively.

On SPICE, we compare the test results of the official MACE-OFF23 Large model and MACE (DupG), as follows.

Table R4: SPICE test MAE. E is in meV/atom, and F is in meV/Å.

ModelE MAEF MAE
MACE-OFF23 Large0.7713.39
MACE (DupG)0.5312.40

We also performed additional experiments on OMat24-1M, and their test energy and force errors are reported below.

Table R5: OMat24-1M test MAE. Both models were trained for 10 epochs. E is in meV/atom, and F is in meV/Å.

ModelE MAEF MAE
MACE78.69161.25
MACE (DupG)64.95154.46

Relative to MACE, DupG reduces test Force MAE and Energy MAE by 4.21% and 17.45%, respectively.

Importantly, on the approximately million-configuration SPICE dataset, MACE (DupG) achieves lower test energy and force prediction errors than the MACE-OFF23 Large model. On the periodic crystal systems in OMat24-1M, DupG also achieves lower test Energy MAE and Force MAE than the MACE baseline.

These additional results therefore directly address the reviewer's concern that the accuracy gains might disappear as training scales up: the advantages of DupG are not limited to the original single-molecule benchmarks with only a few hundred training configurations, but are retained at the million-configuration training scale tested so far, while further demonstrating the potential to extend to periodic materials.

We have revised the manuscript to include the SPICE and OMat24-1M experiments in Appendix F.1, Tables A9–A10, and their training settings in Appendix F.1, Tables A11–A12.

We thank the reviewer again for the constructive feedback and sincerely hope that these additional experiments and clarifications address the concerns raised.

References

Eastman, P. et al. (2023). SPICE, A Dataset of Drug-like Molecules and Peptides for Training Machine Learning Potentials. Scientific Data, 10, 11. https://doi.org/10.1038/s41597-022-01882-6

Barros-Luque, L. et al. (2026). The Open Materials 2024 (OMat24) inorganic materials dataset and models. Nature Computational Science, 6, 642–652. https://doi.org/10.1038/s43588-026-00996-w

Thanks for the additional experimental result, which addresses my concerns. I increased my score to 3.

Thank you for your response and for raising the score to 3. We are glad that our revisions have addressed concerns. We truly appreciate your feedback, which has helped us improve this work. We welcome any additional feedback you may have.

Comments on the paper

3 comments

Dear Reviewers,

The authors have now responded to the questions and concerns raised by everyone. Could I kindly ask you to review the authors' responses, take a look at the comments from the other reviewers, and actively participate in the discussion?

Thank you all very much for your valuable time and contributions!

Best regards,

The AC

AuthorsOpenReview ↗

# Summary of Rebuttal Updates and Major Manuscript Revisions (1/2)

Dear Area Chair and Reviewers,

We sincerely thank the AC for coordinating the review process and for encouraging active discussion, and we thank all three reviewers for their thoughtful feedback. During the rebuttal period, we carefully addressed the concerns raised, conducted substantial additional experiments, and significantly strengthened the manuscript.

We are pleased to receive the replies from Reviewer SqJF, who explicitly acknowledged that our additional experiments addressed the concerns and raised the score from 2 to 3.

To facilitate the AC’s assessment, we summarize below the key strengths recognized by the reviewers, the major concerns raised during review, our rebuttal responses, and the corresponding manuscript revisions.

1. Strengths Recognized by the Reviewers

Our work introduces Duplicated-Graph Training (DupG), a training strategy that improves label efficiency by presenting each DFT-labeled configuration through multiple interaction geometries without requiring additional reference calculations.

The reviewers recognized the following strengths of our work:

  • Novelty and label efficiency [SqJF, TsZv]: Both reviewers recognized novelty of modifying training graphs to better utilize existing labels.
  • Consistent force-accuracy improvements [SqJF, TsZv, 1pXV]: Reviewers acknowledged the consistent gains in force prediction across multiple benchmarks.
  • Practical value of distillation [SqJF, TsZv]: Reviewers appreciated that the benefits of DupG can be transferred to ordinary-graph students without duplicated-graph inference overhead.

2. Major Concerns Raised by the Reviewers and Our Responses

2.1 Architectural generalization, large-scale training, and periodic materials

Concern: Whether the method is only effective on MACE and small molecular benchmarks.

Our additions:

  • Cross-architecture validation: Extended DupG and distillation to EquiformerV3 and GemNet. Distilled students improve Force MAE on all ten rMD17 molecules, with average reductions of 34.67% and 20.66%, respectively.
  • SPICE (~1M configurations): DupG reduces test Energy MAE from 0.77 to 0.53 meV/atom and Force MAE from 13.39 to 12.40 meV/Å.
  • OMat24-1M: After 10 epochs of training for both models, DupG reduces test Energy MAE and Force MAE by 17.45% and 4.21%, respectively.
  • MD22: Force MAE improves on all seven systems spanning 42–370 atoms, and we further added independent-initialization experiments on three representative systems.

Takeaway: These results show that DupG generalizes beyond the original small-scale, single-architecture setting.

2.2 Downstream molecular dynamics evaluation

Concern: Static Energy/Force MAE alone is insufficient to establish practical value in real MD applications.

Our additions:

  • Conducted 54 NVE/NVT trajectories totaling 8.64 million integration steps on Malonaldehyde, Aspirin, and Ethanol.
  • Sampled 180 MD configurations and computed independent DFT references.
  • MACE (Distill) achieves 28.00%, 18.16%, and 21.23% lower DFT-referenced Force MAE on the three molecules, respectively.

Takeaway: The improvements persist on configurations actually visited during MD, not only on static test sets.

2.3 Computational cost and fairness of comparison

Concern: Whether DupG’s gains are due only to additional training computation or different Energy–Force loss weights.

Our additions:

  • Compute-budget ablation: Extending ordinary MACE training by 4.25× still does not reproduce DupG’s Force MAE improvement (5.2 meV/Å vs. 3.8 meV/Å).
  • Matched loss-weight experiment on 3BPA: Under identical 9:1000 E:F loss weights, DupG achieves lower Energy RMSE and Force RMSE at 300 K, 600 K, and 1200 K.
  • Clarification: rMD17 and MD22 already used the same nominal E:F loss-weight ratios in the original comparisons.

Takeaway: DupG’s gains cannot be explained simply by longer training or different loss-weight settings.

Importantly, Reviewer 1pXV explicitly identified full large-scale training and computational ablations as two central requirements for establishing the practical value of this work. We have provided new experimental evidence directly addressing both requests.

2.4 Energy–Force trade-off and distillation mechanism

Concern: Why some energy metrics do not improve uniformly, and why the Student can sometimes outperform the Teacher.

Our response:

  • We clarified the different roles of energy and force errors in MD.
  • We distinguished the independent benefit of the DupG Teacher from the later benefit of distillation.
  • We corrected the Early Stopping issue in the EquiformerV3 distillation experiments on Paracetamol and Naphthalene.

Updated result: The corrected Paracetamol Student achieves 0.78 meV Energy MAE and 3.30 meV/Å Force MAE, improving both over the ordinary EquiformerV3 baseline.

AuthorsOpenReview ↗

# Summary of Rebuttal Updates and Major Manuscript Revisions (Part 2 of 2)

3. Reviewer Discussion Status

  • Reviewer SqJF: Explicitly stated that the additional experiments addressed their major concerns and raised their score from 2 to 3.
  • Reviewer TsZv: Raised further questions after reading our initial rebuttal; we submitted additional clarifications and corrected EquiformerV3 results in response.
  • Reviewer 1pXV: We provided direct responses to all major concerns, including large-scale training, computational cost, size generalization, and practical applicability. Although we have not received a follow-up response, we respectfully hope that the AC will reconsider the reviewer's initial assessment in light of this substantial new evidence. We particularly appreciate the AC's efforts in encouraging the reviewer to examine our rebuttal.

4. Major Revisions in the Manuscript

LocationMain revisions
Section 5.2, Table 2; Appendix BEquiformerV3/GemNet results and corrected distillation settings
Section 5.5; Appendix F.2, Table A14Compute-budget ablation
Section 5.6, Table 6; Appendix F.4MD simulations and DFT-referenced force evaluation
Appendix F.1, Tables A10–A13SPICE and OMat24-1M experiments
Appendix F.3, Table A15Additional MD22 initialization experiments
Appendix C.5, Table A9Matched E:F loss-weight experiment on 3BPA
Section 7Practical recommendations on label efficiency and distillation-based deployment

5. Conclusion

DupG offers a distinct approach to improving the utilization of expensive DFT labels without modifying the backbone architecture, while enabling efficient ordinary-graph deployment through distillation. We believe that the rebuttal has substantially strengthened the experimental support, generalizability, and practical relevance of our work.

Through cross-architecture validation, million-configuration training, periodic-materials evaluation, downstream MD simulations, and compute-budget controls, we have directly addressed the reviewers' main concerns.

Overall, we greatly appreciate the constructive feedback received throughout the review process. These discussions have helped us substantially improve our manuscript, and we hope that the revised version provides a clearer and more comprehensive presentation of our contributions.

Once again, we sincerely thank the AC and all reviewers for their time, thoughtful feedback, and guidance.

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.