acceptodds
Withdrawn from ICLR 2027

CrysPFlow: Physics-Supervised Flow Matching for Crystal Relaxation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Crystal structure relaxation seeks stable, low-energy structures from unrelaxed configurations. First-principles methods achieve high accuracy through repeated energy and force evaluations, but incur substantial computational cost, while direct learning-based approaches largely rely on relaxed-structure supervision and underuse the local directional information encoded by the potential-energy surface (PES). We introduce CrySPFlow, a physics-supervised Flow Matching framework that combines geometric endpoint supervision with PES-derived force feedback during training. To make physical supervision reflect the model-induced relaxation dynamics, we propagate terminal force feedback through a short differentiable manifold rollout. We further introduce an endpoint-preserving gradient injection rule that incorporates physical descent information while preserving the geometric objective’s first-order descent rate in raw-gradient space. The interatomic potential is used only during training, so inference remains purely flow-based without additional potential evaluations. On , CrySPFlow reduces coordinate and bond-length MAEs by 40.6% and 48.5%, respectively, relative to the strongest baseline for each metric. On X–Mn–O, it reduces coordinate MAE by 9.26%, while experiments with adapted FlowMM, RG-VFM, and CrystalFlow backbones show consistent benefits of physics supervision across the tested architectures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal Relaxation, Flow Matching, Physics-Guided Learning, Geometric deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Z3WvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces CrysPFlow, a flow model for crystal structure relaxation. The authors introduce endpoint-preserving physics injection using the MLIP forces to adjust the computed geometric gradients. Specifically, they add the orthogonal component of the physics loss gradient to the geometric gradient, which is passed to the optimizer. The results show that their model has lower structural error on the SrTiO3 and X-Mn-O datasets than the baselines and that their physics guidance helps not only their model but also the adapted flow matching baselines.

Critical strengths

Motivation for endpoint-preserving physics injection is reasonable.

Critical weaknesses

While the proposed method appears technically sound, I have substantial concerns about the implementation and the rigor of the experimental evaluation.

1. Insufficient methodological details and reproducibility

The paper lacks several important details needed to reproduce the work. These omissions raise substantial concerns about the reproducibility of the experiments, the fairness of the baseline comparisons, and the reliability of the resulting conclusions.

  • Model architecture: The authors do not describe their own backbone, despite Table 4 showing that it already outperforms the other backbones even without physical guidance. The model architecture, number of layers, and parameter count are not specified.
  • Baseline adaptation: The paper refers to an RG-VFM [1] backbone, although RG-VFM primarily introduces a training objective, and its material model builds on MOFFlow [2], which uses an IPA-based architecture modeling building-block translations, rotations, and lattice parameters. It is unclear how this architecture was adapted to inorganic crystals without such building blocks. More broadly, the paper provides insufficient details on how the other flow matching baselines were adapted for crystal relaxation. Furthermore, all adapted baselines use 20 Euler inference steps, deviating from their original inference configurations, without justification or a step-count ablation.
  • Dataset construction: The datasets lack important details about their source/curation, preprocessing, and number of data points. In particular, the SrTiO3 dataset uses ground-truth atomic annotations as Y1Y_1, but how is the initial structure Y0Y_0 determined?
  • Reproducibility: Although some training hyperparameters are provided, GPU specifications, training time, and inference latency are missing. No reproducible code is provided, making it difficult for practitioners to reproduce the results.

2. Limited evaluation of crystal relaxation

  • Missing MLIP baselines: Crystal structure relaxation is regularly performed using iterative optimization with MLIPs. While the authors motivate direct prediction models by arguing that MLIP-based relaxation remains computationally expensive, they do not compare against iterative MLIP relaxation to establish the accuracy–compute trade-off. A matched-budget comparison (e.g., with a few MLIP steps) would also be helpful.
  • Limited benchmarks: The evaluation is restricted to SrTiO3 and X-Mn-O. A related structure relaxation task, IS2RS, was introduced in OC20 [3] and also studied in OC22 [4]. Although these benchmarks focus on adsorbate–surface systems rather than bulk crystals, the paper does not discuss this related work or evaluate the proposed method on broader, established relaxation benchmarks.
  • Robustness across MLIPs: The proposed method is evaluated using only MACE for physical supervision. Experiments with other MLIPs would help establish whether the improvements generalize across different learned energy landscapes.

[1] Zaghen, Olga, et al. "Riemannian variational flow matching for material and protein design." International Conference on Learning Representations. Vol. 2026. 2026.
[2] Kim, Nayoung, et al. "Mofflow: Flow matching for structure prediction of metal-organic frameworks." International Conference on Learning Representations. Vol. 2025. 2025.
[3] Chanussot, Lowik, et al. "Open catalyst 2020 (OC20) dataset and community challenges." Acs Catalysis 11.10 (2021): 6059-6072.
[4] Tran, Richard, et al. "The Open Catalyst 2022 (OC22) dataset and challenges for oxide electrocatalysts." Acs Catalysis 13.5 (2023): 3066-3084.

Questions for authors

Mainly, my questions are about implementation and experimental design, as stated in the weaknesses section -- e.g.,

  1. What model architectures do the authors use? What GPU resources did the authors use?
  2. What do the authors mean by the RG-VFM backbone? The backbone RG-VFM uses for metal-organic frameworks assumes that crystals can be decomposed into building blocks, which isn't obvious for inorganic crystals.
  3. How did the authors construct the dataset? Did the authors consider evaluating on more established benchmarks like OC20/22?
  4. How do the model's speed and accuracy compare to MLIPs?
  5. Does the endpoint-preserving physics injection method show benefit across other MLIPs?
4 comments
AuthorsOpenReview ↗

We thank the reviewer for the specific suggestions!

We thank the reviewer for the specific suggestions on reproducibility and experimental rigor. We agree that the original manuscript did not sufficiently distinguish the training objective, prediction network, and task adaptation. Below, we provide the network architecture and computational resources for X–Mn–O and clarify data construction, inference budgets, and comparisons with iterative MLIP relaxation.

1 Network architecture, GPU resources, and reproducibility (Weakness 1, Question 1)

We provide the network architecture and computational resources for the X–Mn–O periodic endpoint model below.

For X–Mn–O periodic crystals, we use an atom-level message-passing endpoint network with a hidden dimension of 512, a time-embedding dimension of 256, and 128 periodic Fourier frequency terms per coordinate axis. The network comprises 7 blocks: 1 initial context block and 6 subsequent message-passing blocks, with LayerNorm, SiLU, and residual connections. The atom-level prediction head outputs fractional-coordinate corrections, followed by removal of global translation and periodic wrapping. The pooled prediction head outputs corrections to the 6 independent components of the symmetric logarithmic cell metric. The predictor has 14,269,190 trainable parameters, approximately 14.27M, excluding the frozen potential model.

The original X–Mn–O training used a single NVIDIA RTX PRO 5000 Blackwell GPU and took approximately 48 hours in total. Single-step endpoint prediction has an average latency of 5.347 ms/structure, requires only one prediction-network call, and makes no MACE queries at inference. The timing scope is detailed in Sec. 5.

We agree that the backbone is already strong without physical supervision, indicating that the network itself also contributes to performance. We therefore assess the contribution of physical supervision through same-backbone comparisons, such as Table 3 of the original manuscript, rather than attributing all performance differences between architectures to physical feedback. In the camera-ready version, we will consolidate the descriptions of the network, computational resources, and baseline adaptations, and provide training/inference implementations, fixed configurations, and details of data processing and evaluation.

2 Meaning of “RG-VFM backbone” and baseline adaptation (Weakness 1, Question 2)

The reviewer is correct that RG-VFM defines a training framework rather than a unique backbone architecture, and our original terminology was imprecise. Its original MOF experiments indeed use MOFFlow's rigid-building-block representation. In contrast, our inorganic-crystal experiments adapt the endpoint objective to an atom-level periodic network, without using MOFFlow's IPA architecture or requiring crystals to be decomposed into MOF building blocks. Atomic species are given as known inputs, and the network directly predicts atomic coordinates and the cell. We will use the term “atom-level adaptation of the RG-VFM objective” and describe the objective and architecture separately. This distinction is consistent with the original RG-VFM and MOFFlow methods.

3 Data sources, construction, and Y₀ for SrTiO₃ (Weakness 1, Question 3)

For SrTiO₃, Y₁ consists of reference atomic annotations, whereas Y₀ is constructed from the observed input. Lattice spacing is estimated solely from input-side Sr/Ti observations in the training region. Missing sites are completed using a lattice fitted to the input, while observation indicators are retained. Target annotations are used for supervision, evaluation, and target-coverage filtering, but do not supply the coordinates of Y₀. After spatial separation and valid-region filtering, no atomic identifiers are shared across splits. Supervised interpolation during training uses the target structure, consistent with Section 5.1 of the original manuscript; inference conditioning remains based solely on the input.

4 Single-step inference accuracy, network-call efficiency, and sensitivity to step count (Weakness 1)

We thank the reviewer for drawing attention to the inference configuration and fairness of the comparisons. The 20-step setting is the predefined, uniform network-call budget for the three externally adapted baselines; the main CrySPFlow model uses a single endpoint prediction, with NFE = 1. Section 5.1 of the original manuscript states that the main model uses one forward evaluation, while the cross-backbone experiments specify: “The external adapters used 20 Euler steps at inference.” We will present this accuracy–inference-cost comparison more clearly: CrySPFlow achieves higher structural accuracy with fewer network calls than the externally adapted models using multi-step inference.

The table below directly reports the prediction errors and network-call counts for the “+MACE+Unrolling” configuration in Table 4 of the original manuscript.

SrTiO₃ modelCoordinate MAE ↓ (Å)Bond-length MAE ↓ (Å)Network calls (NFE) ↓
CrySPFlow0.0080210.0054151
FlowMM0.0140000.00682320
CrystalFlow0.0131370.00744820
RG-VFM0.0148140.01392020

Specifically, in Table 4 of the original manuscript, the full CrySPFlow configuration achieves a coordinate MAE of 0.008021 Å and a bond-length MAE of 0.005415 Å with 1 network call. Both errors are lower than the respective best results reported for the three external adapters with 20 network calls. Relative to the best coordinate MAE among the external models (CrystalFlow, 0.013137 Å) and the best bond-length MAE (FlowMM, 0.006823 Å), these errors are reduced by 38.94% and 20.64%, respectively, with 95% fewer network calls. This demonstrates the accuracy and network-call efficiency advantages of single-step endpoint prediction under the evaluated configurations.

For each external backbone, the geometric baseline, local-feedback variant, and unrolled-feedback variant use the same sampler, NFE, and sampling seed. None of the physical-supervision variants calls MACE at inference. The uniform 20-step setting therefore also allows same-backbone comparisons to assess the contribution of physical supervision without changes in the number of inference calls.

5 Accuracy and computational-cost comparison with iterative MLIP relaxation (Weakness 2, Question 4)

We have added a MACE+FIRE iterative relaxation baseline, evaluated on all 2,859 X–Mn–O test structures with budgets of at most 1, 5, 20, and 200 steps. These budgets cover both a small number of MLIP iterations and more extensive optimization.

Table R5. MACE+FIRE results on the X–Mn–O test set. Coordinate, bond-length, and lattice errors are in Å, and volume errors are in ų. Atomic positions and the cell are optimized jointly, with no computation or evaluation failures at any budget.

Maximum stepsCoord ↓Bond ↓Lat.-R ↓Lat.-A ↓Vol. ↓Rel.V (%) ↓Match (%) ↑Mean MLIP queriesMean runtime (ms/structure) ↓
≤10.3064360.4329410.2170760.21960332.15020619.74465.3381.9959.444
≤50.2635860.3490220.1880400.18746026.66235615.27868.3815.93185.376
≤200.1787420.2103590.1281760.11407411.2544816.45677.12519.24614.411
≤2000.1669280.1796570.1262370.0935458.0396894.38582.58151.731691.211

The additional MACE+FIRE experiments directly cover low-budget iterative relaxation. As the step cap increases from 1 to 200, coordinate error decreases from 0.306436 to 0.166928 Å, Match increases from 65.34% to 82.58%, and mean runtime increases from 59.44 to 1,691.21 ms/structure. Under the 200-step cap, 98.15% of structures meet the MACE force/stress stopping criteria. Further improvements with a larger budget incur additional cost: increasing the step cap from 20 to 200 reduces coordinate error by 6.61% while increasing runtime by a factor of 2.75.

For inference cost, CrySPFlow's independent endpoint timing records 5.35 ms/structure, with only one network evaluation and no MLIP queries at inference. This timing and the above supplementary model-accuracy results come from independent runs. The recorded FIRE runtimes across the four budgets range from 11.12 to 316.31 times this value.

These results directly support the design of using physical feedback during training while retaining low-cost endpoint prediction at inference. We will include this accuracy–runtime comparison in the camera-ready version.

6 Validation on OC20/OC22 and with other MLIPs (Weakness 2, Questions 3 and 5)

The additional OC22 experiments directly address the suggestion to evaluate on an established surface-relaxation dataset, yielding a coordinate MAE of 0.14711 Å and an ADwT of 47.13%. OC22 provides IS2RS initial–relaxed structure pairs, as well as energy- and force-related tasks. We will add OC20 IS2RS and OC22 to the related work and explain the differences between the fixed-atom/fixed-cell constraints in surface relaxation and the joint recovery of bulk atomic coordinates and cells in X–Mn–O. We selected OC22 for this round.

Summary of reported OC22 results.

Method / task adaptationCoordinate MAE (Å) ↓Mean displacement error (Å) ↓Vector RMSE (Å) ↓ADwT (%) ↑
DeepRelax0.1871750.3784820.51697940.36
RG-VFM0.1853940.3762200.51079240.44
CrySPFlow0.1471100.2970600.40772047.13

The CHGNet and MatterSim results (see Table 3 in our response to Reviewer 2) address compatibility with other MLIPs. In the tested endpoint-preserving, terminal-local implementation, both teachers improve coordinate and bond-length errors and the matching rate relative to the geometric baseline. These improvements come at the cost of small increases in lattice/volume errors. The revised manuscript will specify the exact variant, without extending this observation to all gradient paths or all teachers.

We thank the reviewer for the suggestions, which help us present the methodological contributions and supporting evidence more precisely. In the camera-ready version, we will consolidate the implementation and data protocols, correct the RG-VFM terminology, include the additional material/teacher/stress and MLIP-budget experiments, and clarify the scope of the statistical and comparative conclusions.

If these clarifications and additional results have addressed your concerns, we would be grateful if you could consider a more favorable overall assessment of our paper.

I thank the authors for their response. My main concerns about baseline fairness and reproducibility are only partly addressed. Most changes are promised for the camera-ready version, but the manuscript has not yet been revised.

1. RG-VFM baseline The response confirms that "RG-VFM backbone" is the RG-VFM objective applied to the authors' own atom-level periodic network. This model does not exist in prior work, yet the paper presents it as an established baseline. Because the network is not described, it is unclear how it differs from "Our backbone." Its weak performance may therefore reflect the authors' implementation.

2. Baseline inference settings The FlowMM and CrystalFlow adaptations are still not described, and the baselines use 20 Euler steps instead of their original inference settings. The response justifies this by noting that CrySPFlow needs fewer network calls. However, this efficiency does not justify running the baselines with weaker settings than those in their original papers.

3. Architecture and reproducibility Thank you for the X-Mn-O details. However, the SrTiO3 backbone is still not described, although it accounts for most of the gain on SrTiO3. Baseline parameter counts and dataset sources, sizes, and splits are also missing.

4. OC22 evaluation I appreciate the OC22 experiments. However, the baselines again include the self-implemented RG-VFM (see point 1) and omit the standard IS2RS baselines from the OC22 paper (e.g., SpinConv, GemNet-dT, and GemNet-OC). The reported 47.13% ADwT is only slightly above the initial-structure baseline, which involves no learning (44%), and well below GemNet-OC trained on OC22 (58%).

Overall, the remaining concerns about baseline implementations and reproducibility limit the credibility of the results, so I will keep my score.

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

This paper proposes CrySPFlow, a physics-supervised flow-matching framework for crystal structure relaxation that integrates geometric endpoint supervision with force feedback from a frozen interatomic potential. The method propagates physical feedback through a short differentiable manifold rollout and introduces an endpoint-preserving gradient injection mechanism to incorporate physically informed updates while maintaining the first-order descent behavior of the geometric objective in raw-gradient space. The interatomic potential is used only during training, allowing inference to remain purely flow-based. Experiments on SrTiO3 show substantial reductions in coordinate and bond-length errors compared with the strongest respective baselines, while results on X–Mn–O demonstrate more moderate but generally favorable improvements.

Critical strengths
  1. The paper addresses an important limitation of purely endpoint-supervised crystal relaxation. Using forces as training-only supervision is appealing because it incorporates local PES information without requiring potential evaluations during deployment.

  2. The gradient-routing analysis and cross-backbone experiments are valuable. The ablation shows progressive improvement from geometric-only supervision to direct physical supervision and full rollout feedback. Moreover, MACE guidance improves coordinate error across adapted FlowMM, RG-VFM, CrystalFlow, and the authors’ own backbone, supporting the claim that the supervision strategy is at least partially architecture independent.

Critical weaknesses
  1. The experimental results lack uncertainty estimates despite several fine-grained numerical comparisons. The paper reports point estimates, and the main experiments appear to rely on a single random seed, making it difficult to assess the robustness of some of the smaller improvements. This concern is particularly relevant to the cross-backbone experiments, where certain gains are marginal (e.g., the CrystalFlow coordinate MAE improves from 0.013140 with local MACE guidance to 0.013137 with unrolling). Reporting results over multiple independent runs with standard deviations or confidence intervals would provide stronger evidence that these improvements are robust and reproducible.

  2. The empirical validation of the proposed physical supervision remains limited in scope. The method is primarily evaluated on SrTiO3 and X–Mn–O, leaving its generalization to more diverse chemical compositions and relaxation settings unclear. In addition, the physical supervision uses force feedback with the predicted cell held fixed and does not directly incorporate lattice-stress supervision, which may be relevant when cell degrees of freedom play an important role in crystal relaxation. Evaluation on a broader range of materials and further investigation of stress-aware supervision would strengthen the evidence for the general applicability of the proposed framework.

Questions for authors
  1. Could the authors provide results across multiple independent training seeds, particularly for the main CrySPFlow model and the cross-backbone experiments? Since the current results are reported as point estimates and some of the observed improvements are relatively small, uncertainty estimates would help assess the robustness and reproducibility of the reported gains.

  2. How sensitive is CrySPFlow to the choice and accuracy of the frozen interatomic potential? Since the physical supervision is derived from MACE forces, it would be helpful to clarify whether the observed benefits are expected to remain consistent when using alternative interatomic potentials or when the force estimates are imperfect.

  3. The current physical supervision holds the predicted cell fixed during force evaluation and does not directly incorporate stress information. Could the authors clarify how this design choice may affect the remaining lattice errors and whether stress-based supervision has been considered for periodic crystal relaxation?

2 comments
AuthorsOpenReview ↗

We thank the reviewer for the specific suggestions!

We thank the reviewer for the specific suggestions on experimental robustness, the scope of applicability, and the design of physical supervision. Below, we address each point using the same-backbone ablations in the original manuscript, additional WBM evaluations, alternative potential models, and stress-supervision experiments.

1 Multiple random seeds and small performance differences (Weakness 1, Question 1)

We agree that results from independent training seeds are important for assessing robustness. Multi-seed training for the main model and cross-backbone experiments has not yet been completed. All current results should therefore be interpreted as single-seed point estimates. As the reviewer noted, the CrystalFlow coordinate MAE decreases from 0.013140 to 0.013137 Å, a difference of only 0.000003 Å. We do not regard this as substantive evidence that unrolling outperforms local feedback. Its bond-length MAE is indeed slightly better with local feedback (0.007444 versus 0.007448 Å).

However, the evidence for the contribution of physical supervision is not limited to this small difference. The same-backbone ablations in Table 3 of the original manuscript directly compare geometric supervision with different physical-feedback paths on the same CrySPFlow network:

Supervision configurationCoordinate MAE (Å) ↓Bond-length MAE (Å) ↓
A: Geometric supervision only0.0088440.006910
B: Direct physical supervision0.0086630.005981
C: Terminal-local correction0.0084620.005883
D: Trajectory correction0.0082620.005785
E: Full terminal feedback0.0080210.005415

Configurations C–E all retain direct physical updates and differ in the gradient path used for the additional terminal feedback. Relative to A, E reduces coordinate and bond-length errors by 9.31% and 21.64%, respectively. These same-backbone results support the contribution of physical feedback more directly than comparisons across architectures. Nevertheless, they remain descriptive evidence from a single seed and do not replace an assessment of statistical uncertainty.

In the camera-ready version, we will explicitly distinguish improvements observed in a particular run from statistical robustness across seeds. We will narrow our interpretation of small differences and plan to report means and standard deviations from independent training runs for the main model and cross-backbone experiments. The additional material and potential-model experiments below address the scope of applicability and interface compatibility, respectively; they do not substitute for multi-seed validation.

2 Coverage of material compositions and relaxation tasks (Weakness 2)

In response to this suggestion, we have extended the evaluation to WBM periodic bulk crystals. These additional experiments use terminal-local physical feedback to assess applicability across broader chemical compositions.

Table R1. WBM results on actual paired initial and relaxed structures. Coordinate, bond-length, and lattice errors are in Å; volume errors are in ų. All methods have zero evaluation failures.

Method / task adaptationCoord ↓Bond ↓Lat.-R ↓Lat.-A ↓Vol. ↓Rel.V (%) ↓Match (%) ↑
DeepRelax0.1352130.1564870.0977770.0948478.4779265.18716189.449523
CrystalFlow0.1757520.2016160.1278650.12450614.1208319.18676889.651351
FlowMM0.1966370.2305470.1665010.15597016.28530910.69063985.040959
Atom-level adaptation of the RG-VFM objective0.1413330.1693530.1046570.10304111.0531956.82019189.726542
CrySPFlow, terminal-local feedback0.0822600.0870850.0609880.0577003.4967072.12710493.509834

On WBM, CrySPFlow achieves the best results across all six geometric error metrics and the matching rate. Compared with DeepRelax, coordinate, bond-length, and volume errors decrease by 39.16%, 44.35%, and 58.76%, respectively. Match reaches 93.51%, exceeding the strongest baseline by 3.78 percentage points. These results extend the validation beyond SrTiO₃ and X–Mn–O to bulk crystals with a broader range of compositions.

3 Dependence on the frozen potential model and force-estimation accuracy (Question 2)

The supervision interface is not restricted to MACE. As described in Section 4.2 of the original manuscript, gradients through the force signal are stopped, while teacher feedback propagates through the prediction network. The interface therefore only requires atomic forces from the frozen potential, without further differentiation of the teacher forces. Under a unified training protocol, we have added two frozen teachers, CHGNet and MatterSim.

Table R3. Comparison of alternative frozen teachers on X–Mn–O. Coordinate, bond-length, and lattice errors are in Å; volume errors are in ų. All models use single-step inference without querying the teacher potential.

Model / force teacherCoord ↓Bond ↓Lat.-R ↓Lat.-A ↓Vol. ↓Rel.V (%) ↓Match (%) ↑
Input baseline (geometric model before fine-tuning)0.1291970.1458250.0760820.0727983.9017702.24592682.324116
+ CHGNet fine-tuning0.1287310.1451420.0762130.0729963.9616202.30092782.884144
+ MatterSim fine-tuning0.1283690.1449880.0760870.0728903.9590842.29911182.674134

Both CHGNet and MatterSim improve coordinate error, bond-length error, and matching rate relative to the geometric baseline. Coordinate error decreases by 0.36% and 0.64%, and Match increases by 0.56 and 0.35 percentage points, respectively. Both teachers also slightly increase lattice and volume errors. The consistent direction of changes in coordinate error, bond-length error, and matching rate indicates that this force-feedback interface can accommodate potential models beyond MACE. However, the benefits depend on the metric.

Regarding force-estimation errors, the teacher supplies an auxiliary estimate of the energy gradient, rather than ground-truth reference forces. Geometric endpoint supervision still targets the reference structures, and the endpoint-preserving rule controls the interaction between the two gradients. However, its first-order property in the raw gradient space does not guarantee that arbitrarily inaccurate forces are harmless or that generalization will improve. The benefits may therefore still depend on the teacher's accuracy within the relevant chemical space. We will make this boundary explicit in the main text. We will not present the alternative-teacher experiments as a guarantee of consistent robustness to arbitrary potential models.

4 Holding the predicted cell fixed and stress supervision (Weakness 2, Question 3)

Holding the predicted cell fixed during force queries does not mean that the periodic model keeps the input cell unchanged throughout prediction. The network predicts both fractional coordinates and the cell, and both are supervised by the geometric endpoint loss (Section 4.1, Eq. 7). It is the direct cell-gradient path during force queries that is stopped. Section 4.2, immediately after Eq. 9, already states: “This supplies coordinate-energy supervision without a direct stress objective.” In other words, the method provides energy supervision for coordinates without a direct stress objective.

The physical limitation raised by the reviewer is valid: atomic force balance does not guarantee cell stress balance. The absence of direct stress supervision may be related to the residual lattice errors, but the current results cannot attribute those errors solely to this factor. The table below presents the main force-only results reported in the paper and the additional force-plus-stress results on X–Mn–O.

Table R4. Force-only and force-plus-stress feedback results on X–Mn–O. Both configurations use single-step endpoint inference. The two rows come from different checkpoints and do not constitute a controlled ablation in which only stress supervision changes.

Physical supervisionCoord (Å) ↓Bond (Å) ↓Lat.-R (Å) ↓Lat.-A (Å) ↓Vol. (ų) ↓Rel.V (%) ↓
Force-only feedback (main results in the paper)0.1150410.1389240.0692020.1179843.6450112.104
Force-plus-stress feedback (additional experiment)0.1333400.1573090.0842520.0808196.0789953.404971

The additional force-plus-stress model achieves a coordinate error of 0.133340 Šand a volume error of 6.078995 ų. Here, stress is estimated by the frozen MACE model from the energy derivative with respect to homogeneous strain. The external pressure is zero, and the stress-term weight is 1. This result does not diminish the importance of stress: atomic force balance and cell stress balance are distinct constraints. Stress feedback still needs to be appropriately balanced with the geometric objective. We will report this additional result and its implications for supervision design in the camera-ready version.

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

See above

Critical strengths

See above

Critical weaknesses

See above

Questions for authors

My weaknesses are qestions for the authors to answer practically (like including informatipn about what they will change in the paper + what experiments they might run to suppor ttheir claims)

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

Related papers

Discussion (0)

Sign in to comment.