- 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 (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 is not fixed, but is randomly resampled according to the following distribution:
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 on six rMD17 molecules.
The experiments show that:
- When Å, all six molecules perform worse than the baseline.
- When is within 0.03–0.1 Å, Force MAE is lower than the baseline for all six molecules.
- When is increased further, performance on some molecules begins to deteriorate.
When the translation becomes too large, cross-copy interactions disappear; when , 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
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.
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.
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