acceptodds
Withdrawn from ICLR 2027

StrainPair: Learning Elastic Response in Pretrained Interatomic Potentials

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Elastic moduli measure how stress changes under deformation. Yet interatomic potentials are typically fine-tuned to match stresses one configuration at a time. We find that these two objectives can diverge: an adapted potential can predict elastic moduli more accurately even as its common stress-offset error increases. Motivated by this distinction, we introduce StrainPair, which augments conventional fine-tuning with a simple paired loss that matches stress differences between deformations of the same material. The loss uses existing labels and leaves the potential architecture and inference procedure unchanged. In our evaluation, StrainPair reduces shear- and bulk-modulus errors by 20.00% and 17.52% over matched pointwise fine-tuning across 10,987 materials. Three-seed experiments on 224 external materials confirm the gains. Stronger pointwise supervision, cross-material pairing, and affine output calibration do not recover the same improvements. Instead, the gains accompany a reduction in deformation-response error, despite larger common stress offsets. Our analysis explains this separation: stress differences cancel shared offsets and constrain variation along deformation paths. These results identify response accuracy as a distinct target for adapting pretrained potentials. They also reveal a tradeoff: better elasticity does not necessarily preserve energy, force, and stress accuracy on other configurations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
pretrained interatomic potentials, fine-tuning, paired supervision, elastic properties, materials machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_gNHSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes StrainPair, a fine-tuning objective for adapting pretrained interatomic potentials to elastic properties. In addition to the regression loss on energy, force, and stress, StrainPair adds an L1 loss on the stress differences between deformed configurations of the same material along the same strain axis, using the existing stress labels of Materials Project elasticity calculations. The authors argue that this pairing cancels a common stress offset and constrains how the stress changes with strain, and they support this with an analysis for affine residual paths. Starting from a pretrained Orb-v3 model, StrainPair gives lower shear and bulk modulus errors than fine-tuning on the same data without the paired loss, both on Matbench and on a separate set of materials. The authors attribute the gain to a more accurate stress changes along each strain path.

Critical strengths

The method is simple and easy to adopt, as it reuses stress labels that already exist in elasticity datasets.

Critical weaknesses

Limited methodological novelty. The main contribution appears to be an application of established training ideas to elasticity. Learning output differences has precedents in GAR [1], derivative supervision in Sobolev training [2], and supervision of energy curvature in interatomic potentials in PFT [3]. StrainPair’s main distinction is to pair deformations of the same material along the same strain axis using existing stress labels. The analysis explains offset cancellation and its connection to stress response, but I find the additional insight beyond these expected properties limited. I therefore view the methodological contribution as incremental. Experiments beyond a single Orb-v3 variant would help establish its broader value.

Missing baseline and unclear scoring criteria. Table 2 omits the pretrained model, so it is unclear whether either method improves external performance over its starting point. The criteria for accepting an elastic estimate are also not fully specified. Invalid estimates receive a median replacement when calculating MAE, where in Table 1, StrainPair produces 84 more valid estimates than Single. The comparison therefore includes materials for which one method uses its prediction and the other uses the replacement value. Please define the validity criteria and report MAE on materials for which all models produce valid estimates, alongside the existing scores. This would show whether the accuracy gains persist without median replacements.

Unclear practical advantage. Compared with Single, StrainPair has higher force and stress errors and larger deviations under rotation. On the 1,500 frames excluded from training, force MAE increases by 17.76% over the pretrained model, compared with 10.73% for Single. Both satisfy the authors’ stated error limit, but energy errors mentioned in Section 2 are not reported. The improvement in modulus MAE also yields only a small screening gain: eight additional materials with $ G \geq 100$ GPa among 1,099 selections, with no increase in bulk threshold hits. These results leave unclear when the modulus improvements justify the accompanying losses in other predictions.

Questions for authors
  • Could you add the pretrained model’s results on the external materials to Table 2, using the same evaluation procedure? This would make it clear whether either method improves over its starting point.

  • What are the criteria for accepting an elastic estimate? Please also report MAE on materials for which all models produce valid estimates, alongside the existing scores, to clarify the effect of median replacements.

  • Could you report the energy errors mentioned in Section 2? How do modulus accuracy and force and stress errors change as the weight of the pair loss is varied?

  • Do the improvements over matched Single persist with another Orb-v3 pretraining variant or a different potential architecture?

References

[1] Dixian Zhu, Tianbao Yang, and Livnat Jerby. Gradient Aligned Regression via Pairwise Losses. ICML, 2025.

[2] Wojciech M. Czarnecki, Simon Osindero, Max Jaderberg, Grzegorz Swirszcz, and Razvan Pascanu. Sobolev Training for Neural Networks. NeurIPS, 2017.

[3] Teddy Koker, Abhijeet Sadashiv Gangan, Mit Kotak, Jaime Marian, and Tess Smidt. PFT: Phonon Fine-tuning for Machine Learned Interatomic Potentials. ICML, 2026.

Reviewer_91BmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an objective function called StrainPair for fine-tuning pre-trained conservative interatomic potentials. The authors add an L1 loss function to measure the stress difference between different configurations of the same material along the same strain axis, and claim that the objective function can reuse existing labels while maintaining the model architecture and inference process unchanged.

Critical strengths

The improvement objective is simple and its physical meaning is clear. By pairing deformations of the same material, a link is established between output difference monitoring and elastic response.

Critical weaknesses
  1. The authors made a comparison of "making the ordinary method pay more attention to stress prediction", but only showed one adjustment method and did not explain what the ordinary method could best achieve after trying multiple reasonable settings on the validation set.
  2. Furthermore, what if we change the pairing objects to different materials? Wouldn't that change multiple conditions simultaneously? This might not fully prove that "pairing within the same material" is the key to gaining an advantage.
  3. Does the test set truly represent an independent evaluation of the method? Currently, this is unclear (as explicitly stated on pages 5-6 of the paper, the authors used materials from Matbench during the early stages of development and also reviewed external test sets).
Questions for authors
  1. The author mentions that historical development involved Matbench-derived materials and that external test sets were reviewed. Specifically, which results were reviewed? Were the methods, training parameters, or evaluation rules modified after reviewing them?

  2. The authors introduce a comparison method for increasing stress training weights. What weights does this method ultimately use? Are there comparison results that select different weights solely based on the validation set?

  3. Table 2 does not list the original, untuned model. How does it perform on these 224 materials?

Reviewer_64oLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper fine-tunes a pretrained interatomic potential with an extra loss on stress differences between two deformed configurations of the same material along the same strain axis. The motivation is that: elastic moduli are derivatives of stress with respect to strain, so supervising differences should target the quantity of interest more directly than supervising absolute stresses in the loss function. The authors report 20.00% / 17.52% reductions in shear / bulk modulus MAE relative to matched pointwise fine-tuning on 10,987 Matbench materials, a three-seed replication on 224 external materials, and three controls (stronger pointwise weight, cross-material pairing, affine readout calibration). The empirical study is broad and several negative results are reported honestly. My concerns are that the methodological contribution is thin, the central mechanism is not fully tested in real applications.

Critical strengths
  1. The design is cheap and well motivated. The loss needs no new labels, no teacher, no architecture change, and no change at inference.
  2. The problem is real. Universal potentials are demonstrably weak on second-derivative properties, so improving elastic constants with a small loss term is a sensible target.
  3. The results seem honestly reported. The statistics look reasonably calibrated, the negative results are not hidden, and the limitations are stated.
Critical weaknesses
  1. The proposed method is not new, and prior work is not fully discussed. Cancelling systematic error by taking differences is well established: Δ-learning between levels of theory (Ramakrishnan et al., 2015), pairwise difference regression between samples (Tynes et al., 2021). These piror works were cited.
  2. Limited methodological contribution. StrainPair adds one regulariser to a standard energy–force–stress fine-tuning setup; architecture, optimisation and inference are unchanged. Useful for an applications venue, but thin for a learning paper, which expects some understanding of why the objective works, not only that it does. The paper does not explain why a difference target should behave differently from a derivative target, nor where the gain should hold.
  3. StrainPair can lead to degraded stress prediction and degraded relaxation. Relaxing with the StrainPair fine-tuned model can lead to wrong cell and therefore evaluating the G and K in the wrong position in the potential energy surface. The authors reported degradation in both force and stress compared to regular fine-tuning. This can lead to many potential issues that are not well discussed in the manuscript.
  4. The method has been only tested on one model and one dataset: a single Orb-v3 checkpoint, one MP elasticity pool, and Matbench only. Frontier potentials such as UMA are far more accurate on stress. If the coherent offset the pair term is meant to cancel is largely gone in a stronger base model, the effect may shrink or vanish. Since the methodology is already simple and transferrable enough, the authors need to show this on multiple different MLIPs, datasets, and applications.
  5. Downstream demonstration too limited. The only end-to-end evidence is Table 3a, which re-ranks 10,987 materials with known moduli and counts threshold hits, and Table 3b, which scores a frozen pool of 400 generated candidates without DFT confirmation. It is unclear whether a 20% MAE gain changes results significantly in realistic applications, especially considering the force and stress accuracy degraded. The manuscript can be improved if the model can be demonstrated in realistic applications.
Questions for authors
  1. Offset-insensitivity. The paper's explanation is that stress differences cancel a shared offset. But a common offset can be removed far more simply: subtract the group mean from the residuals, or centre the predicted stresses, before applying an ordinary pointwise loss. That involves no pairs at all, and under an L1 objective it is close to equivalent to a pairwise difference target. If such a centred pointwise loss reproduces the gain, then the effect belongs to offset removal rather than to pairing, and the "physical pairing" framing is not what is doing the work. Is the gain still there when the comparison is against that baseline rather than against raw pointwise training?
  2. Weight selection. The loss weights are fixed at (10, 2, 5) with λ = 5 and never varied, and I could not find where they come from. This matters more than it might appear: the balance between the pointwise and paired terms decides how much of the paired strss loss actually affects training. How were these values chosen, and do the conclusions survive a sweep over the stress weight and λ?
  3. Relaxation. Does the relaxed cell volume shift between arms? Given that force and offset errors both worsen, I would expect some drift in the optimized cell, and it would help to know how much of the bulk difference is attributable to it.

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

Related papers

Discussion (0)

Sign in to comment.