acceptodds
Under review as a conference paper at ICLR 2027

CrystalJTF: Jacobian-Pulled Gradient Guidance with Atom-Adaptive Gradient Scaling for Crystal Diffusion Models

Abstract

Diffusion models offer a promising route to crystal discovery, but generating thermodynamically stable and novel structures remains challenging. Integrat- ing physical signals from pretrained interatomic potentials into sampling tra- jectories could improve generation quality without additional training. Effec- tive feedback, however, must address two challenges: the coordinate-space mis- match between predicted-endpoint forces and current sampling coordinates, and differences in local response scales across atoms in periodic systems. We pro- pose CrystalJTF, a plug-and-play, training-free guidance framework for crystal diffusion models. Point Q-JTF pulls endpoint forces back to the current sam- pling coordinates through the predicted clean endpoint map’s Jacobian trans- pose. Relative-BB uses sampling history to adapt atom-wise relative update scales, while a shared periodic geometry cap limits displacement magnitudes. CrystalJTF freezes both the generator and potential and retains native lattice and element-type sampling rules, supporting composition-conditioned and un- conditional generation. Across six benchmark settings involving CrysLLM- Gen, DiffCSP, and MatterGen, CrystalJTF consistently improves the fractions of stable or metastable, unique, and novel structures (SUN/MSUN). Under CHGNet 0.3.0 evaluation, SUN and MSUN increase by an average of 1.28 and 4.78 percentage points on MP-20, respectively, while MatterGen’s MSUN in- creases by 13.8 percentage points on Alex-MP-20. Sensitivity analyses across evaluation-potential versions and matching tolerances further support the robust- ness of these results.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal Structure Generation, Diffusion Models, Physics-Guided Sampling, Jacobian-Transported Force Guidance, Atom-Wise Adaptive Step Sizes, Training-Free Guidance.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QdDAOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes CrystalJTF, a training-free way to use a frozen MLIP as feedback during sampling from pretrained crystal diffusion models. The method reweights per-atom forces from the MLIP on the clean structure estimate and uses them to update fractional coordinates while keeping lattice and species intact. Experiments on a wide range of base models and training data show higher SUN and MSUN in all settings, with CHGNet used for relaxation and scoring. An ablation suggests that atom-wise scales improve results over uniform scaling and over using POINT Q-JTF instead of Endpoint forces.

Critical strengths
  1. The method really is a plug-in and does not require retraining. One controller with the same hyperparameters improves SUN/MSUN on three different generators, including an LLM+diffusion.
  2. The evaluation is more careful than is common in this area. I would like to point out the analysis of matcher tolerances and the dependency on the number of generated samples.
Critical weaknesses

The paper is generally clear. My concerns are mostly about how well the specific design choices are supported.

  1. The 2x2 ablation switches two components on top of the full controller, but simpler alternatives are not tested:
  • a sweep over the guidance interval.
  • a sweep over ρ\rho.
  • a constant step size aa instead of the cap-derived aka_k.
  • plain Point Q-JTF with several guidance strengths.

Without these, it is hard to say whether Relative-BB and the cap are needed, or whether a tuned DPS-style step would do the same.

  1. The isolated effects are also small. Experiments in Table 3 differ by 0.2-0.6 SUN. CIs over several independent sets of 10,000 generations are needed, and even more so for the MatterGen rows with n = 1,000.

  2. Guidance uses a single MACE MLIP, and all evaluation uses CHGNet. So the "two evaluation potentials" are two versions of the same, by now fairly old, model. Three additions would address this:

  • Separate the effect of in-loop guidance from simple pre-relaxation, e.g., by relaxing baseline samples post hoc with the same MACE model before the CHGNet step.
  • Recompute relaxation and E_hull with a more recent model from Matbench Discovery (https://matbench-discovery.materialsproject.org).
  • Try 2-3 other MLIPs as the source of guidance gradients to check that the gains are not specific to MACE.
Questions for authors
  1. How often does the pair term in Eq. (8) bind rather than the atom term? What fraction of proposals is rejected by the 0.5A check?
  2. GSS (Qin et al., 2026, arXiv:2604.27636), which uses MatterSim forces inside a MatterGen-style sampler, is missing from related work. Could you discuss it in the relevant works?
  3. De Lambilly et al. (2026) report that MatterSim energy guidance on MatterGen performed about the same as unguided sampling. Could you discuss their results in comparison to the proposed approach?

What could increase my score is the set of experiments listed under weaknesses:

  • CIs over independent 10k runs.
  • Sweeps over ρ\rho and the guidance interval.
  • Constant-step and plain Point Q-JTF baselines.
  • A newer evaluation MLIP and other guidance MLIPs.
  • The post-hoc pre-relaxation control.
4 comments
AuthorsOpenReview ↗

Response Part 1: Geometric Constraints and Related Work

We thank the reviewer for the careful assessment and constructive suggestions, and for recognizing the plug-in nature of CrystalJTF and our evaluation efforts. We organize our response into two parts. This first part addresses the three Questions for Authors; Part 2 will address the experimental concerns and requested additional analyses.

  1. Constraint binding and proposal rejection. Across two independent DiffCSP runs of 10,000 samples each, the pair term was the binding constraint in 99.9798% of 3,611,616 proposed updates, versus 0.0202% for the atom term, with no ties. The post-proposal geometry check rejected only 7 updates (0.000194%).
  2. Relationship to GSS. GSS is closely related, and we will include it in Related Work. GSS combines diffusion and physical-force updates through a time-dependent mixing schedule for composition-conditioned structure search. CrystalJTF develops an atom-resolved, geometry-constrained feedback controller for de novo crystal generation. Its distinctive design couples Relative-BB scaling across atoms with atom/pair displacement caps. The atom-wise weighting can also alter the joint update direction of the entire structure, rather than merely rescaling its magnitude. The controller guides coordinates while leaving composition and lattice evolution to the pretrained generator. We evaluate this design across multiple generators using SUN/MSUN.
  3. Comparison with De Lambilly et al. De Lambilly et al. report that MatterSim energy guidance produced no significant improvement in their tested chemical systems. Their framework already uses clean-structure estimates and gradient normalization. However, a physically meaningful guidance direction alone does not guarantee effective sampling: its impact also depends on how updates are distributed across atoms and how their magnitudes are controlled. CrystalJTF explicitly addresses this update-control problem by coupling atom-wise Relative-BB scaling with atom/pair displacement caps. In our MatterGen MP-20 comparison with 10,000 samples per method, CrystalJTF increases SUN from 7.68% to 8.57% and MSUN from 52.83% to 54.55%. These results support the effectiveness of our atom-resolved, geometry-constrained feedback design under the evaluated settings. We will clarify this methodological distinction and discuss their findings in the revised manuscript.
AuthorsOpenReview ↗

Response Part 2: Additional Experiments and Validation

  1. Confidence intervals across three 10,000-sample runs

    CrystalJTF improves SUN and MSUN over unguided sampling in all three reported DiffCSP MP-20 runs, each generating 10,000 structures per method. We summarize the mean metrics and 95% Student-t confidence intervals of the run-level paired improvements below. Values are percentages; Δ denotes CrystalJTF − Baseline, with Δ and its confidence interval expressed in percentage points.

    MatcherMethodNUSMSSUNMSUN
    StrictBaseline90.3197.2311.4358.758.7148.28
    CrystalJTF91.3397.3812.7563.5510.1354.08
    Δ+1.02+0.15+1.32+4.80+1.42+5.80
    95% CI of Δ[0.560, 1.480][−0.012, 0.319][1.201, 1.446][3.539, 6.068][1.395, 1.452][4.701, 6.899]

    SUN and MSUN increase in every run, with mean paired gains of 1.423 and 5.800 percentage points, respectively. Both confidence intervals lie entirely above zero, and the positive gains across all three runs demonstrate that the observed improvement is not confined to a single run.

  2. Parameter search and comparison with tuned simple guidance

    CrystalJTF achieves higher SUN and MSUN than the simple-guidance baseline selected through a ten-configuration parameter search in the reported 10,000-sample test. We evaluated five geometric-cap strengths and five constant-step sizes on a separate development set, then froze the selected configuration before testing. CrystalJTF retained its original settings, ρ=0.15\rho=0.15 and κ=2\kappa=2, throughout this comparison.

    Both simple-guidance variants use uniform atom weighting. Uniform-fixed uses a normalized Point Q-JTF direction without Relative-BB or adaptive geometric scaling. Its reference step size, α0≈0.380057\alpha_0 \approx 0.380057 Å, was calibrated independently without using SUN labels.

    Development search: 2,000 structures per configuration. All metric values are percentages.

    MethodConfigurationNUSMSSUNMSUN
    Baseline—90.1599.1012.9559.2010.1049.45
    Uniform-capρ=0.05\rho=0.0590.2099.1012.7060.659.9550.95
    Uniform-capρ=0.10\rho=0.1090.8599.0513.9563.1011.3553.90
    Uniform-capρ=0.15\rho=0.1591.0099.0013.4062.3511.0053.25
    Uniform-capρ=0.20\rho=0.20, selected91.6598.8013.8563.1011.5554.65
    Uniform-capρ=0.30\rho=0.3094.3599.1011.2059.559.7553.40
    Uniform-fixedα=0.25α0\alpha=0.25\alpha_089.9099.0513.2060.6010.2050.45
    Uniform-fixedα=0.50α0\alpha=0.50\alpha_090.5098.9513.3061.8010.4552.25
    Uniform-fixedα=α0\alpha=\alpha_091.2599.0513.1562.6510.8053.70
    Uniform-fixedα=2α0\alpha=2\alpha_094.5099.3010.8558.259.4052.45
    Uniform-fixedα=4α0\alpha=4\alpha_098.6098.809.0541.858.7040.10
    CrystalJTFOriginal settings90.8098.9513.1063.4010.5554.10

    Uniform-cap with ρ=0.20\rho=0.20 achieved the highest development SUN and was selected for the final comparison. The sweep also shows that stronger guidance is not uniformly beneficial: increasing the fixed step from α0\alpha_0 to 4α04\alpha_0 raises novelty but reduces MSUN from 53.70% to 40.10%.

    Final comparison: 10,000 structures per method. Δ denotes CrystalJTF − Uniform-cap, in percentage points.

    MethodNUSMSSUNMSUN
    Selected Uniform-cap (ρ=0.20\rho=0.20)91.8397.3011.9062.839.5253.61
    CrystalJTF91.3897.5312.6563.4210.1754.05
    Δ−0.45+0.23+0.75+0.59+0.65+0.44

    Although the selected simple baseline scored higher on the development set, CrystalJTF obtained 65 additional SUN structures and 44 additional MSUN structures in this independent test. Thus, the development-selected baseline did not match the complete controller’s final SUN/MSUN, despite receiving a ten-configuration parameter search. This comparison supports the value of the complete atom-resolved, geometry-constrained feedback design beyond tuning a global guidance strength within the tested search space.

AuthorsOpenReview ↗

Response Part 2: Additional Experiments and Validation

  1. Cross-MLIP validation and post-hoc pre-relaxation controls

    CrystalJTF improves SUN and MSUN across all four tested combinations of guidance and relaxation MLIPs, demonstrating cross-MLIP transferability within the evaluated settings. We vary the guidance potential between MACE and MatterSim and the final relaxation/evaluation potential between CHGNet and MatterSim. Improvements persist under both matched-potential and cross-potential evaluation, showing that the observed benefits are not restricted to MACE guidance or CHGNet evaluation.

    All experiments use DiffCSP on MP-20 with 10,000 structures per group. Values are percentages. Each Δ row compares the preceding method with the unguided Baseline under the same evaluation MLIP, in percentage points.

    MethodGuidance / pre-relaxation MLIPEvaluation MLIPNUSMSSUNMSUN
    Baseline—CHGNet90.2297.3011.3558.888.7648.41
    CrystalJTFMACECHGNet91.3897.5312.6563.4210.1754.05
    Δ+1.16+0.23+1.30+4.54+1.41+5.64
    CrystalJTFMatterSimCHGNet91.2797.5912.3563.509.9554.07
    Δ+1.05+0.29+1.00+4.62+1.19+5.66
    Baseline + post-hoc pre-relaxationMACECHGNet90.2597.3011.6260.329.0449.86
    Δ+0.03+0.00+0.27+1.44+0.28+1.45
    Baseline—MatterSim90.2297.237.9743.205.7233.01
    CrystalJTFMACEMatterSim91.1297.388.0544.846.0035.50
    Δ+0.90+0.15+0.08+1.64+0.28+2.49
    CrystalJTFMatterSimMatterSim90.9397.468.0645.055.9035.62
    Δ+0.71+0.23+0.09+1.85+0.18+2.61

    Under CHGNet evaluation, MACE-guided and MatterSim-guided CrystalJTF improve SUN/MSUN over unguided sampling by 1.41/5.64 and 1.19/5.66 percentage points, respectively. Under MatterSim evaluation, both guidance potentials retain positive gains: 0.28/2.49 and 0.18/2.61 percentage points, respectively. These results support our central contribution: an atom-resolved, geometry-constrained controller that translates feedback from different MLIPs into effective sampling updates without retraining the generator.

    The post-hoc control further distinguishes this contribution from simply adding an energy-minimization stage. With the same MACE potential and subsequent CHGNet evaluation, in-loop CrystalJTF exceeds post-hoc pre-relaxation by 1.13 percentage points in SUN and 4.19 percentage points in MSUN. Thus, the tested post-hoc procedure does not reproduce the gains of sampling-time feedback, supporting the value of controlling updates during generation rather than relying solely on terminal relaxation.

  2. Expanded MatterGen evaluation from 1,000 to 10,000 samples

    Our original manuscript reported MatterGen results using only 1,000 generated structures per method. To strengthen the evidence for the reliability of CrystalJTF, we expanded the evaluation to 10,000 structures per method on both MP-20 and Alex-MP-20, comparing Baseline and CrystalJTF in four groups comprising 40,000 structures in total. CrystalJTF improves SUN and MSUN on both datasets, with particularly strong MSUN gains on Alex-MP-20.

    Values are percentages. Each Δ row compares CrystalJTF with its corresponding Baseline within the same dataset and evaluation protocol, in percentage points.

    DatasetMethodNUSMSSUNMSUN
    MP-20Baseline91.1997.939.6162.107.6852.83
    CrystalJTF93.0498.3810.1061.868.5754.55
    Δ+1.85+0.45+0.49−0.24+0.89+1.72
    Alex-MP-20Baseline61.2199.4822.5387.7110.8450.36
    CrystalJTF80.5599.7217.0781.4611.4262.75
    Δ+19.34+0.24−5.46−6.25+0.58+12.39

    On MP-20, CrystalJTF produces 89 additional SUN structures and 172 additional MSUN structures at the fixed 10,000-sample budget. On Alex-MP-20, SUN increases from 10.84% to 11.42%, while MSUN increases substantially from 50.36% to 62.75%. These gains correspond to 58 additional SUN structures and 1,239 additional MSUN structures.

    The expanded experiments therefore provide larger-scale evidence that CrystalJTF improves the yield of stable or metastable, unique, and novel crystals on MatterGen. In particular, the Alex-MP-20 results demonstrate a substantial improvement in joint discovery yield, even though not every marginal metric increases.

Reviewer_QdDAOpenReview ↗

Thank you for the comments

I thank the authors for the detailed response. All my questions are addressed. The results look well motivated and supported by the experiments, so I increase the score to 3.

Reviewer_T2fUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CrystalJTF is a training free guidance method for pretrained crystal diffusion models. In the last 200 sampling steps, it evaluates a pretrained MACE potential on the Tweedie estimate of the clean coordinates. It then pulls the force back to the current coordinates with one vector-Jacobian product, using their method called Point Q-JTF. Per-atom step weights come from Barzilai-Borwein secants between consecutive guided steps (Relative-BB), and a periodic geometry cap sets the step size. Lattice and species updates are left to the base sampler. The method is applied to CrysLLMGen, DiffCSP and MatterGen on MP-20, MPTS-52 and Alex-MP-20 with paired seeds. Structures are relaxed and scored with CHGNet against Materials Project. SUN and MSUN improve in all six settings, with average gains of 1.28 and 4.78 pts on MP-20. A 2×2 ablation on DiffCSP separates the pullback from the atom-wise scaling. My expertise is diffusion models and guidance rather than materials evaluation, so I focus on the guidance method and the evaluation design.

Critical strengths
  1. Plug-and-play and consistent across generators. The generator and potential stay frozen and no retraining is needed. The method works with an LLM-diffusion hybrid, DiffCSP and MatterGen, in both conditional and unconditional generation. SUN and MSUN improve in every setting under both CHGNet versions. The guidance potential (MACE) differs from the evaluation potential (CHGNet), so the method does not directly optimize the metric.
  2. Careful evaluation protocol. Paired seeds isolate the effect of guidance. Fixed denominators count failed relaxations instead of dropping them. Results are checked against matcher tolerances and CHGNet versions. The sample size analysis of uniqueness is a useful note for the field.
Critical weaknesses
  1. Limited novelty and many hand-set choices. Point Q-JTF computes an energy gradient, which is the DPS and Universal Guidance gradient through the Tweedie estimate with energy as the loss. The paper itself credits DPS for evaluating at the endpoint. The pullback view is the chain rule and the descent result follows from it and positive diagonal scaling. Energy and forces as a guidance or reward signal are also common in prior work. The remaining contributions, Relative-BB and the geometry cap, rely on fixed values: clipping to [1/2, 2],15% displacement cap, 0.5 Å rejection distance, and guidance at every one of the last 200 steps. None is ablated. The cap effectively sets the guidance strength but is never varied, so the stability vs novelty trade-off is not shown. The BB secant spans steps where the objective changes and noise is injected, so it has no curvature interpretation, as the authors acknowledge.
  2. No external baselines and no significance tests. The only comparisons are unguided sampling and the internal ablation. Missing comparison methods are DPS with a tuned scale, other training-free guidance methods, MatterGen’s own property-conditioned sampling, and a matched-compute baseline that spends the MACE evaluations on generating and filtering more unguided samples. Compute cost is not reported. No confidence intervals are given despite the paired design, and the SUN gains on MatterGen (1000 samples) and the SUN gaps between guided ablation arms are around one standard error. On MatterGen Alex-MP-20, stability drops while novelty rises by about 21 pts, so the MSUN gain there is novelty-driven, which is unexpected for energy-descent guidance. The ablation covers one generator and is scored with CHGNet 0.2.0, which is not stated. There is no DFT validation.
Questions for authors
  1. Compare against 1) DPS with a tuned guidance scale and other training-free guidance methods 2) MatterGen’s property-conditioned sampling 3) a matched-compute baseline that filters extra unguided samples by MACE energy, all with the same checkpoints. Report wall time per sample.
  2. Report paired significance tests or confidence intervals for the SUN and MSUN gains and the ablation, and a sensitivity sweep over the cap fraction, the clipping range and the guidance window.
  3. Why does energy guidance lower S and MS on MatterGen Alex-MP-20? Can you show that the novel structures are not distorted variants of known ones, for example through symmetry statistics or DFT checks on a subset?
4 comments
AuthorsOpenReview ↗

Response to Question 1

We thank the reviewer for these valuable suggestions. We provide additional comparisons, clarify methodological distinctions, and report computational costs.

1. Baseline Comparisons and Computational Cost

1) Tuned DPS and Other Training-Free Guidance Methods

To test global-strength tuning, we evaluated five Uniform-cap and five Uniform-fixed configurations on a separate development set, then froze the selected configuration. CrystalJTF retained its original settings.

Here, α₀ ≈ 0.380057 Å is the independently calibrated reference step size.

Development search: 2,000 structures per configuration. Values are percentages.

MethodConfigurationNUSMSSUNMSUN
Baseline—90.1599.1012.9559.2010.1049.45
Uniform-capρ = 0.0590.2099.1012.7060.659.9550.95
Uniform-capρ = 0.1090.8599.0513.9563.1011.3553.90
Uniform-capρ = 0.1591.0099.0013.4062.3511.0053.25
Uniform-capρ = 0.20, selected91.6598.8013.8563.1011.5554.65
Uniform-capρ = 0.3094.3599.1011.2059.559.7553.40
Uniform-fixedα = 0.25α₀89.9099.0513.2060.6010.2050.45
Uniform-fixedα = 0.50α₀90.5098.9513.3061.8010.4552.25
Uniform-fixedα = α₀91.2599.0513.1562.6510.8053.70
Uniform-fixedα = 2α₀94.5099.3010.8558.259.4052.45
Uniform-fixedα = 4α₀98.6098.809.0541.858.7040.10
CrystalJTFOriginal settings90.8098.9513.1063.4010.5554.10

Uniform-cap at ρ = 0.20 was selected by development SUN. Stronger guidance is not uniformly better: increasing the fixed step from α₀ to 4α₀ raises novelty but reduces MSUN from 53.70% to 40.10%.

Reported independent test: 10,000 structures per method. Δ denotes CrystalJTF − Uniform-cap, in percentage points.

MethodNUSMSSUNMSUN
Selected Uniform-cap (ρ = 0.20)91.8397.3011.9062.839.5253.61
CrystalJTF91.3897.5312.6563.4210.1754.05
Δ−0.45+0.23+0.75+0.59+0.65+0.44

Although the selected baseline scored higher on the development set, CrystalJTF obtained 65 additional SUN structures and 44 additional MSUN structures in this test, despite the baseline’s ten-configuration search. This supports the controller’s value beyond global-strength tuning within the tested search space.

GSS’s crystalline experiments address composition-conditioned CSP and low-energy structure coverage. Our DNG comparison does not fix composition and evaluates cross-composition SUN/MSUN, so reported CSP results are not direct DNG comparisons.

2) MatterGen’s Property-Conditioned Sampling

MatterGen’s property-conditioned sampling is not directly comparable to CrystalJTF under the reviewer’s same-checkpoint, training-free setting.

MatterGen learns conditional generation through property-supervised fine-tuning and uses classifier-free guidance during sampling. Its released chemical-system–energy-above-hull model employs full fine-tuning, changing the generator weights.

CrystalJTF instead keeps the original generator and potential frozen, introducing energy and force feedback only during sampling. It requires no property-specific generator retraining.

A direct comparison would therefore confound the sampling method with differences in property supervision and generator weights, rather than isolate the contribution of sampling-time guidance.

3) Matched-Compute Generation Followed by Filtering

Generate-then-filter ranks or accepts/rejects completed candidates using MACE energies. Our motivation is instead to introduce energy feedback into generation itself.

Learning low-energy crystal distributions provides a structural prior, but does not explicitly supply physical feedback along each trajectory. CrystalJTF combines this prior with energy-informed updates during denoising.

Filtering selects completed outcomes; CrystalJTF influences how candidates are produced. Its controller maps endpoint feedback to the current diffusion state, coordinates atom-wise updates, and constrains displacements to respect periodic geometry.

4) Wall Time per Sample

We acknowledge that the improvements in generation quality come with increased sampling time.

MethodSamplesTotal Sampling TimeTime per StructureRelative Runtime
Unguided Baseline10044.12 s0.441 s1.00×
CrystalJTF100154.48 s1.545 s3.50×

CrystalJTF requires 3.50× the baseline runtime. Timings cover warmed sampling, excluding initialization, serialization, relaxation, and evaluation; per-structure times are batch-amortized.

AuthorsOpenReview ↗

Response to Question 2

2. Statistical Reliability and Parameter Sensitivity

1) Paired Confidence Intervals Across Three Independent Runs

We thank the reviewer for highlighting statistical reliability. We conducted three independent paired runs on DiffCSP MP-20, each generating 10,000 structures per method. CrystalJTF improves both SUN and MSUN in all three runs.

The table reports mean metrics and 95% Student-t confidence intervals for the run-level paired improvements. Metric values are percentages; Δ denotes CrystalJTF − Baseline, with Δ and its confidence interval expressed in percentage points.

MatcherMethodNUSMSSUNMSUN
StrictBaseline90.3197.2311.4358.758.7148.28
CrystalJTF91.3397.3812.7563.5510.1354.08
Δ+1.02+0.15+1.32+4.80+1.42+5.80
95% CI of Δ[0.560, 1.480][−0.012, 0.319][1.201, 1.446][3.539, 6.068][1.395, 1.452][4.701, 6.899]

The mean paired gains in SUN and MSUN are 1.423 and 5.800 percentage points, respectively, with both confidence intervals entirely above zero. Consistent positive gains across all three runs further support the reliability of the main results and show that the improvement is not confined to a single run.

2) Statistical Analysis of the Ablation Experiments

Our additional experiments currently prioritize the overall benefits of the complete method. The three large-scale paired runs examine whether the SUN and MSUN improvements are reproducible across independent runs.

We agree with the reviewer that repeated runs and paired comparisons of individual ablation arms would help quantify each component’s contribution more precisely. Such analyses would usefully complement the current validation of overall gains, although a complete statistical analysis of every ablation arm has not yet been included in this round of experiments.

3) Parameter Sensitivity and the Guidance Window

Our current focus is on whether the complete method delivers reliable discovery gains and whether those gains can be reproduced simply by adjusting the strength of simplified guidance. We therefore added sweeps over five Uniform-cap and five Uniform-fixed configurations, followed by independent testing with the development-selected parameters frozen.

Restricting guidance to the final 200 steps balances the reliability of structure estimates against computational cost. Early in diffusion sampling, noise levels are high and predicted clean structures remain uncertain; applying physical feedback too early may interfere with structure formation. At later, lower-noise stages, structure estimates become more reliable and energy and force feedback is better suited to local refinement. We therefore retain model-driven exploration in the early stages and introduce physical feedback during the final 200 steps, while limiting additional potential evaluations and gradient computations.

We also recognize the value of the reviewer’s suggestions concerning the cap fraction, clipping range, and guidance window. These analyses would build on the current validation of the main gains by further characterizing parameter dependence and the method’s operating range. Although this round of additional experiments does not cover all these sweeps, we agree that they are important for a more comprehensive analysis.

AuthorsOpenReview ↗

Response to Question 3

3. Stability and Novelty on MatterGen Alex-MP-20

1) Higher Discovery Yield After Expanding to 10,000 Structures

The MatterGen results in our original manuscript were based on 1,000 generated structures per method. To strengthen the evidence, we expanded the Alex-MP-20 evaluation tenfold to 10,000 structures per method. We further classified the generated structures jointly by stability and novelty to examine the change in overall stability. The following are per-sample counts before deduplication:

CategoryBaselineCrystalJTFΔ
Stable and novel1,0841,142+58
Stable and non-novel1,169565−604
Metastable-only and novel3,9565,138+1,182
Metastable-only and non-novel2,5621,301−1,261
E_hull ≥ 0.1 eV/atom and novel1,0341,712+678
E_hull ≥ 0.1 eV/atom and non-novel12571−54
Thermodynamically unknown and novel4763+16
Thermodynamically unknown and non-novel238−15
Total10,00010,0000

The total stable count decreases by 546, but this comprises an increase of 58 stable novel samples and a decrease of 604 stable non-novel samples. Thus, a lower overall stability rate does not imply fewer stable novel discoveries.

The novelty increase is also not confined to high-energy structures: metastable-only novel samples increase from 3,956 to 5,138. After deduplication under the original evaluation protocol, SUN increases from 1,084 to 1,142, and MSUN from 5,036 to 6,275.

Therefore, after expanding from the original 1,000 to 10,000 structures per method, CrystalJTF continues to improve the joint yield of stable or metastable, unique, and novel structures. Our objective is to increase useful new-crystal discoveries, rather than solely maximize the marginal stability fraction across all generated samples.

2) Novelty Is Evaluated After Relaxation, and Simple Post-Processing Does Not Reproduce the Gains

We appreciate the reviewer’s concern that geometric perturbations might be classified as novel structures. Our novelty assessment is performed after CHGNet relaxation, rather than directly on unrelaxed guided outputs. Candidates counted toward SUN/MSUN must remain novel after relaxation while also satisfying uniqueness and the corresponding thermodynamic threshold.

We additionally evaluated a control that applies post-hoc pre-relaxation with the same MACE potential to unguided candidates, followed by the same CHGNet relaxation and evaluation procedure. Each group below contains 10,000 structures:

MethodSUNMSUN
Baseline10.84%50.36%
Baseline + MACE post-hoc pre-relaxation10.84%50.52%
CrystalJTF11.42%62.75%

MACE post-hoc pre-relaxation does not increase SUN and improves MSUN by only 0.16 percentage points. In contrast, CrystalJTF achieves 11.42% SUN and 62.75% MSUN. This control shows that the observed discovery gains are not reproduced by the tested post-generation energy-optimization procedure, supporting the value of incorporating physical feedback during generation.

3) DFT Validation

We agree that DFT validation of representative candidates would provide stronger independent evidence of their physical stability and structural character. Given our currently available computational resources, we have not been able to complete sufficiently converged DFT structural relaxations and the corresponding stability validation for this response.

We therefore prioritized a tenfold expansion of the original evaluation, joint stability–novelty statistics, and a post-processing control using the same potential to strengthen the experimental evidence for the main findings. The present results demonstrate discovery gains under the stated relaxation and MLIP evaluation protocol, rather than DFT-confirmed new phases. We thank the reviewer for this suggestion and recognize the importance of first-principles validation for further assessing individual candidates.

Thank you for the detailed response. I am keeping my rating at 2: Weak rejection.

Addressed: The three paired 10000 samples runs show that the SUN and MSUN gains on DiffCSP MP-20 are reliable. The wall-time cost (3.5x) is now reported. The tuned uniform-guidance sweep and the larger MatterGen evaluation answer my questions on guidance strength and sample size, and the post-hoc relaxation control shows that the gain is not a relaxation effect. The argument against comparing with MatterGen's finetuned conditional model is reasonable.

Remaining: First, the new baseline supports my novelty concern. A uniform cap on the Point Q-JTF direction recovers about half of the SUN gain and most of the MSUN gain. CrystalJTF beats it by 0.65 SUN and 0.44 MSUN pts in one test run and the ranking reverses on the development set. The added value of Relative-BB is therefore not established. Second, the matched-compute filtering baseline was not run, so the benefit at 3.5x cost is unclear, and the cap fraction, clipping range and guidance window of CrystalJTF itself are still not swept. Third, on MatterGen Alex-MP-20, the MSUN gain is driven by novelty while the unstable fraction rises and SUN changes by about 0.6 pts, and there are no symmetry statistics or DFT checks.

Reviewer_3hvdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work introduces a inference/generation time only plug-in to steer diffusion models into generating (meta)stable, unique, and novel crystals. Specifically, the authors machine learning interatomic potential models, e.g. MACE and CHGNet to act as a physical force field "corrector" to bring samples closer to stable crystals during sampling. The authors claim that there are two challenges in doing this. First, the coordinates are still noisy during denoising steps, those machine learning models are trained on clean physical crystals, so feeding noisy coordinates is problematic. Second, different atoms have different local environments, so that an uniform step size might pushes stiff bonds too hard or loose atoms too little; also it might collapse atoms directly into their periodic mirror images. The authors solve these problems by JFT.

First, they use the noisy sample x_t to predict the clean state x_0 directly and then using the clean state to get the physical forces, and then apply it back through Jacobian. They do relative scaling for different atoms for the second issue. Experiments show that their approach can improve SUN when applied to 3 different backbone models on different datasets.

Critical strengths
  • Overall, the idea of applying machine learning interatomic potential models to diffusion models as guidance is novel to the best of my knowledge
Critical weaknesses
  • I have a few questions that I won't frame as critical weaknesses
Questions for authors
  • What is the computational head of this approach, can you provide a wallclock comparison? Since every denoising step invokes a separate inference to a machine learning interatomic potential model, the computational cost seems very high
  • In Table 5, the guidance schedule is set to the final 200 of 1000 steps. Partially, this saves computational cost, but also, in the beginning, the predicted clean states are probably very bad, what's your take on this?
  • In appendix A3 you list some conditions that skips CrystalJTF, can you provide some statistics to show how often this happens?
  • Why does physical force descent cause a decrease in raw stability for MatterGen on Alex-MP-20 while driving up novelty by over 21 percentage?
3 comments
AuthorsOpenReview ↗

Part 1: Computational Efficiency and Guidance Schedule

We sincerely thank the reviewer for taking the time to carefully read our paper and for the positive assessment of the novelty and overall technical soundness of our approach. We particularly appreciate your accurate understanding of the core design motivations and your specific questions concerning computational cost, guidance timing, the frequency of skipped updates, and the relationship between stability and novelty. These constructive suggestions help us explain the practical behavior of CrystalJTF more clearly and provide a more complete account of its experimental results.

To address these questions, we have supplemented our evaluation with wall-clock measurements and detailed guidance-execution statistics. We have also expanded the MatterGen Alex-MP-20 evaluation from 1,000 structures per method in the original manuscript to 10,000 structures per method, enabling a more detailed analysis of the joint changes in stability and novelty. We respond to each question below.

1. Computational Overhead and Wall-Clock Comparison

Thank you for highlighting the importance of computational cost. We first clarify that, under the configuration in Table 5, MLIP guidance is applied only during the final 200 of the 1,000 denoising steps, rather than at every step throughout the entire sampling trajectory. MLIP evaluations are performed in batches, and the potential remains loaded across steps.

We have added a wall-clock comparison for DiffCSP under the same hardware configuration. We generated 100 structures per method with a batch size of 100, executing the complete 1,000-step sampling process in both cases.

MethodFull sampling timeAmortized time per structure
Baseline44.12 s0.441 s
CrystalJTF154.48 s1.545 s

Under this configuration, the complete sampling time of CrystalJTF is 3.50 times that of the baseline, corresponding to approximately 1.10 additional seconds per structure on an amortized basis. The additional computation includes MLIP force evaluation, vector–Jacobian products through the denoiser, and step-size and geometry checks.

The reported values are CUDA-synchronized measurements of the complete sampling loop after warm-up. They exclude model initialization, warm-up, output serialization, and subsequent relaxation and evaluation. The per-structure values represent amortized batch-processing costs, rather than the latency of generating a single structure independently. We will include this wall-clock comparison and its measurement protocol in the revised manuscript.

2. Why Is Guidance Restricted to the Final 200 Steps?

We agree with the reviewer's observation: this choice considers the reliability of the predicted clean structures as well as computational cost.

CrystalJTF evaluates physical forces on the predicted clean structure, avoiding direct evaluation of highly noisy coordinates by the MLIP. At the early, high-noise stages, the clean-state estimate can still have substantial uncertainty, and its local atomic environments may differ from the physical structures represented in the MLIP's training data.

Our late-stage guidance schedule therefore allows the base generative model to first establish a more plausible structure, before introducing physical feedback for refinement at lower noise levels. In this way, the schedule combines the structural prior of the generative model with physical corrections applied when the clean-state estimate is more informative.

The final-200-step schedule thus reflects a balance between the reliability of physical feedback and computational efficiency. We will explain this motivation more explicitly in the revision, emphasizing that late-stage guidance serves a methodological purpose beyond reducing the number of MLIP evaluations.

AuthorsOpenReview ↗

Part 2: Guidance Execution and Stability–Novelty Analysis

3. How Frequently Is Guidance Skipped Under the Conditions in Appendix A3?

Thank you for this helpful suggestion. To quantify the practical impact of these conditions, we analyzed the complete generation logs and counted the outcome of each guidance opportunity.

For one DiffCSP MP-20 experiment, we generated 10,000 structures with a 200-step guidance window per structure, giving a total of 2,000,000 structure–guidance-step opportunities. The execution statistics are as follows:

OutcomeCountPercentage of all guidance opportunities
Guidance update applied1,805,87490.2937%
History initialization or rebuilding; no update at this step188,6189.4309%
Weak gradient4,3000.2150%
Invalid forces at the predicted clean structure9980.0499%
Current structure fails the geometry check2040.0102%
Proposed update fails the geometry check60.0003%
Total2,000,000100%

Overall, 90.2937% of guidance opportunities result in an applied update, while 9.7063% do not. The latter are predominantly associated with initializing or rebuilding the history required by Relative-BB. For example, a change in predicted atom types requires the history to be reset. All other reasons together account for only 0.2754% of opportunities, including 0.0105% attributable to geometry-check rejection.

Thus, most skipped updates reflect the history management required for step-size estimation, rather than frequent geometric or numerical failures.

The denominator here includes only opportunities within the final 200-step guidance window, not the preceding 800 steps. Importantly, skipping guidance only means that the additional physical correction is omitted at that step. The underlying diffusion sampler continues normally, and the sample is not discarded. We will add these statistics and their definitions to Appendix A3.

4. Why Does Stability Decrease While Novelty Increases on MatterGen Alex-MP-20?

Thank you for drawing attention to this result. Two complementary aspects help explain the observation: the distinction between local force-based correction and final thermodynamic evaluation, and the distribution of generated samples across stability–novelty categories.

First, force guidance uses the local gradient of a surrogate potential to modify the sampling trajectory. Final stability, in contrast, is assessed using the relaxed structures and their energies relative to competing phases. The local correction and the final hull-based evaluation therefore assess different aspects of the generation process, making the joint stability–novelty statistics particularly informative.

Second, we expanded the evaluation tenfold, from 1,000 structures per method in the original manuscript to 10,000 structures per method, and classified the samples jointly by stability and novelty. The key categories below report per-sample counts before deduplication:

CategoryBaselineCrystalJTFChange
Stable and novel1,0841,142+58
Stable and non-novel1,169565−604
Metastable-only and novel3,9565,138+1,182
Metastable-only and non-novel2,5621,301−1,261

In the expanded evaluation, the overall stable fraction decreases from 22.53% to 17.07%. However, the reduction of 546 stable samples comprises 604 fewer stable non-novel samples and 58 more stable novel samples. Thus, the lower overall stable fraction does not imply fewer stable novel discoveries.

The increase in novelty is also not confined to high-energy structures: the number of metastable-only novel samples increases by 1,182. After deduplication under the original evaluation protocol, SUN improves from 10.84% to 11.42%, and MSUN from 50.36% to 62.75%.

We additionally evaluated a control that applies post-hoc pre-relaxation with the same MACE potential to unguided candidates, followed by the same CHGNet relaxation and evaluation procedure. Each group contains 10,000 structures:

MethodSUNMSUN
Baseline10.84%50.36%
Baseline + MACE post-hoc pre-relaxation10.84%50.52%
CrystalJTF11.42%62.75%

This control shows that the tested post-generation energy-optimization procedure does not reproduce the joint discovery gains achieved by guidance during generation. Moreover, novelty is assessed after CHGNet relaxation, rather than directly on the unrelaxed guided outputs.

Taken together, these results reveal a stability–novelty trade-off: the marginal stable fraction decreases, while the joint yield of stable or metastable, unique, and novel structures increases. We will present the expanded evaluation and the joint statistics in the revision to make this distinction explicit.

We again thank the reviewer for these constructive comments, which have helped us strengthen the experimental evidence and clarify the practical operation and design rationale of CrystalJTF.

Thank you for the detailed experiments. In the 10k MatterGen table, summing the four categories shows that total valid structures (stable + metastable) actually drop from 87.7% to 81.5% (so unstable generations go up from 12.3% to 18.5%), while ground-state SUN only moves slightly (10.84%to11.42%). Almost all the gain is in metastable structures (MSUN +12.4%). Please explicitly note this increase in the unstable fraction and be careful not to overclaim ground-state stability improvements.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.