acceptodds
Under review as a conference paper at ICLR 2027

Few-Step Crystal Generation via Composition-Conditioned Moment Matching

Abstract

Deep generative models offer a scalable route to explore the vast composition–structure search space underlying materials discovery. Existing diffusion- and flow-based crystal generators, however, still rely on long iterative sampling trajectories, limiting their use in high-throughput settings where large candidate libraries must be generated efficiently. We propose CoMMat (Composition-conditioned Moment matching for Material generation), a few-step crystal generation framework that separates chemical composition modeling from geometric refinement. Given a composition, CoMMat obtains a composition-conditioned structural latent, predicts the lattice once, and refines only atomic coordinates while keeping species and lattice fixed, reducing iterative generation to a continuous geometric transport problem. To solve this problem in only a few network evaluations, we adapt inductive moment matching to periodic crystal geometry through a lattice-normalized Cartesian transport space coupled with periodic coordinate evaluation in wrapped fractional coordinates. Conditioning the structural prior on composition eliminates the need to encode target geometry, enabling a unified few-step generator for both crystal structure prediction and de novo generation. Across the MP-20, Perov-5, Carbon-24, and MPTS-52 benchmarks, CoMMat retains strong generation quality with only 2–8 neural function evaluations. Under matched few-step budgets, it achieves favorable CSP and de novo generation performance relative to recent diffusion- and flow-based baselines, with particularly strong results at the lowest sampling budgets. At its 32-NFE setting, CoMMat remains competitive with much higher-budget baselines while substantially reducing inference cost. These results demonstrate a practical route toward scalable few-step crystal generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal generation, Crystal structure prediction, Few-step generative modeling, Inductive moment matching, Composition-conditioned generation, Periodic generative models, Materials discovery.

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

This paper discusses an interesting problem: generating crystal structures in few steps through generative models. To achieve this goal, the paper proposes Composition-conditioned Moment matching for Material generation (CoMMAT). At a high level, CoMMAT trains a causal Transformer on a ground-truth posterior distribution by minimizing the KL divergence, and the sampled latent variables are fed into a lattice predictor that generates a lattice structure. Lastly, conditioned on this fixed lattice structure, it uses inductive moment matching (IMM) and CSPNet to generate the remaining coordinates.

== post rebuttal ==

Thank you for the thorough response to my concerns. However, I choose to maintain my score. The reason is that I feel my first concern (W1) remains, although the authors did answer my question on what the compute and performance tradeoff would look like for FLOPs. I understand where the authors’ motivation comes from; it is generally the case that we want more efficient generative models across various application domains, but I feel the means to achieve the authors’ claimed “extreme few-step generation” is through assuming a fixed lattice structure (W3) is a bit strong, as the authors have acknowledged.

Critical strengths

The paper proposes to use a single lattice predictor in place of performing multi-step generation on the lattice structure, which saves the number of forward passes in CoMMAT. The experiments conducted are also thorough, and ablations are performed on the design choices of the paper.

Critical weaknesses
  1. There seem to be a lack of motivation on why we need to perform few neural function evaluations (NFEs) on crystal generation task. It seems that the only sentence explaining the motivation in introduction is “For high-throughput screening, however, generation itself must be inexpensive.” without any relevant reference or explanations for such a motivation. This naturally raises the question of why we need CoMMAT in the first place, which is an important question to ask. The paper also aims to obtain good performance in 2-8 NFEs, but it didn’t say how much resources are being used for each NFE such as FLOPs or time consumption. Without this relevant information on marginal improvement of performance given costs, it becomes difficult to evaluate whether we can just apply existing methods using more NFEs while having higher-quality generation or whether we should use CoMMAT to save the cost.
  2. One central assumption in CoMMAT’s design is that a fixed lattice structure suffices for crystal generation task, and ablation studies in Table 3a does establish that a lattice predictor with a single forward pass even outperforms iterative approaches when using 32 NFEs. However, similar to (1), I do not know whether such benefits are meaningful since I do not have information about the tradeoff between performance and costs of the NFEs. If each NFE is cheap or affordable, I don’t see why we can’t run more NFEs with an iterative lattice predictor to improve performance, so more demonstration on the motivation would address this concern.
  3. I am uncertain whether the assumption that a fixed lattice structure suffices is a reasonable assumption. At first glance, I feel it is not a weak assumption, but it is also somewhat flexible since it assumes a fixed structure for a given latent variable. I am not an expert on crystal lattices, so I will leave this question to the other reviewers.

I will consider raising my rating if (1) and (2) are properly addressed.

Questions for authors
  1. Section 4.1 mentions that two paths are used together (one from p and the other from q) for the lattice predictor, I wonder how are they fed to the predictor?
  2. Section 4.2 mentions y, but I do not seem to be able to find how y is defined.
3 comments
AuthorsOpenReview ↗

W1. Motivation for few-step crystal generation and computational cost–quality trade-off

We thank the reviewer for this important observation. We agree that the motivation for few-step crystal generation should be supported by quantitative computational evidence rather than NFE alone. We address both the motivation and the trade-off between generation quality and computational cost.

Motivation. Recent crystal generators, including FlowMM, CrysBFN, SymmBFN, Crystalite, and ZAtom-1, increasingly emphasize sampling efficiency alongside generation quality. For example, FlowMM reports approximately 3× fewer integration steps, while CrysBFN reduces network evaluations by 200× relative to diffusion-based methods. This is particularly relevant to high-throughput materials discovery, where many generated candidates are discarded during screening for structural validity, thermodynamic stability, novelty, and desired properties, making large candidate pools necessary. Efficient generation reduces the computational cost of exploring these pools under limited resources. Motivated by this need, we investigate the extreme few-step regime (as few as 2 coordinate evaluations) to determine whether competitive crystal generation quality can be maintained at substantially lower computational cost.

We will strengthen the Introduction with this motivation and explicit compute–quality comparisons.

Computational cost per NFE. We agree that NFE alone is insufficient for comparing computational efficiency, as different architectures incur different costs per evaluation. We therefore measured FLOPs using the same profiling procedure across models and generation time for 10,000 de novo structures on a single NVIDIA A30 GPU (FP32).

ModelGFLOPs/NFEs/NFE*NFETotal GFLOPs/structureTime/10k
CoMMat-2560.381.02/8/321.6/3.8/12.96/12/37 s
CoMMat-5121.442.32/8/323.7/12.4/47.09/23/77 s
Crystalite1.961.27991570957 s
CrysBFN2.034.2100020304160 s
FlowMM2.307.8100023007801 s
DiffCSP2.593.2200051806459 s
CrystalFlow3.565.9100356592 s

For CoMMat, s/NFE represents incremental decoder time over 10,000 samples; baseline values are native-budget runtimes divided by NFE. CoMMat's DNG totals include fixed generation overhead (~0.8 GFLOPs/structure). Thus, timing definitions are not strictly identical.

Both CoMMat variants require fewer GFLOPs per evaluation than the evaluated baselines. CoMMat-512 requires 1.44 GFLOPs/NFE, compared with 2.30 for FlowMM and 3.56 for CrystalFlow, representing approximately 37% and 60% lower computational cost per evaluation, respectively.

Quality–cost trade-off. To directly address whether existing methods can achieve better quality using additional evaluations, we compare MP-20 CSP match rates against computational cost. Unlike the DNG totals above, these CSP costs exclude composition-generation overhead.

ModelNFEGFLOPs/structure ↓MR (%) ↑
CoMMat-25620.841.9
CoMMat-51222.942.1
FlowMM24.640.3
CrystalFlow27.129.2
CoMMat-25641.545.1
CoMMat-51245.849.3
FlowMM49.252.1
CrystalFlow414.250.5
CoMMat-25683.046.1
CoMMat-512811.550.8
FlowMM818.456.0
CrystalFlow828.555.8

Low-compute performance. At 2 NFE, CoMMat-512 achieves a higher match rate (42.1%) than FlowMM (40.3%) and CrystalFlow (29.2%), while requiring 37% and 59% fewer GFLOPs, respectively. CoMMat-256 further reduces computational cost to only 0.8 GFLOPs/structure while maintaining a comparable 41.9% match rate. At 4 NFE, CoMMat-512 achieves 49.3% MR at 5.8 GFLOPs, approaching CrystalFlow's 50.5% at 14.2 GFLOPs. Notably, CoMMat-512 at 8 NFE achieves 50.8% MR at 11.5 GFLOPs, slightly exceeding CrystalFlow at 4 NFE (50.5%) while requiring approximately 19% less compute.

Marginal quality–cost trade-off. Increasing CoMMat-512 from 2 to 4 NFE improves match rate by 7.2 percentage points while doubling computational cost, whereas increasing from 4 to 8 NFE yields only 1.5 additional points. CoMMat-256 similarly improves from 41.9% to 46.1% across 2–8 NFE while requiring only 0.8–3.0 GFLOPs/structure. These results demonstrate diminishing returns from additional evaluations. Although FlowMM and CrystalFlow achieve higher absolute accuracy at larger budgets, CoMMat provides competitive structural quality at substantially lower computational cost, particularly in the extreme few-step regime.

Therefore, CoMMat offers a favorable quality–cost trade-off under constrained computational budgets, rather than universally outperforming higher-budget methods.

AuthorsOpenReview ↗

W2. Why use one-shot lattice prediction instead of iterative lattice refinement?

Our motivation is to decouple global lattice geometry from atomic-coordinate refinement. CoMMat first predicts the lattice from the composition-conditioned structural latent, L^=hω(z,N)\hat{\mathbf L}=h_\omega(\mathbf z,N), and subsequently refines atomic coordinates within this fixed periodic cell. This provides a consistent lattice geometry throughout coordinate refinement and allows the iterative decoder to focus on atomic arrangements.

We agree that additional lattice evaluations can be affordable and may improve structural quality. To investigate this, we evaluated one-shot prediction, decoupled iterative lattice generation (IMM/Flow), and joint lattice–coordinate refinement across multiple sampling budgets.

Lattice ablation on MP-20 (CSP match rate %, three sampling seeds):

Lattice configuration2 NFE8 NFE32 NFE
One-shot CoMMat42.150.851.65
Decoupled Lattice-IMM (+4 lattice NFE)41.144.745.3
Decoupled Lattice-Flow (+8 lattice NFE)41.544.444.9
One-shot, without decoupled type training28.931.832.6
Joint lattice–coordinate refinement, without decoupled type training17.517.918.0

The upper three rows compare lattice objectives in the final CoMMat training setting. The lower two rows compare joint refinement against its matched type-training control. The joint variant additionally uses fractional-coordinate transport.

Computational cost. The extra lattice evaluations require only 0.003–0.007 GFLOPs per structure, compared with 1.44 GFLOPs for one CoMMat coordinate evaluation. At 8 coordinate NFE, measured sampling times are 23.0/23.3/23.1 seconds per 10,000 structures for one-shot/IMM/Flow, respectively. Thus, lattice refinement is inexpensive; CoMMat's primary computational savings arise from reducing expensive coordinate evaluations, not from eliminating lattice evaluations.

Performance. Neither decoupled iterative alternative consistently improves accuracy over one-shot prediction. Joint lattice–coordinate refinement performs substantially worse than its matched one-shot control (by 11.4–14.6 percentage points across 2–32 NFE). Although the joint variant also differs in coordinate representation, an independent representation ablation changes MR by at most 0.6 points. These findings suggest that simultaneous lattice–coordinate refinement is not automatically beneficial in our evaluated formulation.

Therefore, our results support one-shot lattice prediction as a simpler and competitive architectural choice, rather than demonstrating that iterative lattice refinement is universally unnecessary.

W3. Is the fixed-lattice assumption reasonable?

The lattice is fixed within each coordinate-refinement trajectory, not across all generated crystals. In CoMMat, L^=hω(z,N)\hat{\mathbf L}=h_\omega(\mathbf z,N). During de novo generation, sampling z∼pψ(z∣A)\mathbf z\sim p_\psi(\mathbf z\mid\mathbf A) permits different lattices for the same composition. In CSP, using the conditional-prior mean z=μp(A)\mathbf z=\boldsymbol\mu_p(\mathbf A) produces one predicted lattice per composition.

This factorization simplifies coordinate generation by maintaining consistent periodic geometry, but introduces an important limitation: coordinate refinement cannot correct errors in the initially predicted lattice.

We quantified this limitation through a ground-truth-lattice oracle experiment. On MP-20, the predicted lattice has a median cell-volume error of 3.6%, with 84% of predictions within 10% of the reference volume. Under the corrected validity-gated scorer, replacing the predicted lattice with the ground-truth lattice increases CSP match rate from 42.1% to 54.8% at 2 NFE, 50.8% to 59.9% at 8 NFE, and 51.65% to 60.5% at 32 NFE. These results demonstrate that lattice prediction is a substantial remaining bottleneck that additional coordinate refinement alone cannot resolve.

The limitation is especially relevant for polymorphic compositions, where multiple distinct unit cells may be valid. Although a sampled structural latent permits lattice diversity in de novo generation, the deterministic prior-mean lattice used in CSP cannot explicitly represent multiple cells for the same composition.

Thus, one-shot lattice prediction is a deliberate architectural tradeoff, not an assumption that a fixed lattice is universally sufficient. Our experiments support its effectiveness in the evaluated few-step setting, while identifying more accurate, polymorph-aware lattice prediction and improved joint lattice–coordinate modeling as promising directions for further improvement.

AuthorsOpenReview ↗

Questions for Authors

Q1. How are the two latent paths fed to the lattice predictor?

The two paths are processed separately through the same shared lattice predictor. During training, �0��0� and �1��1� independently produce predicted lattices and coordinate outputs. Each path contributes its own structural loss, combined with λp=1\lambda_{p} = 1 . The latent vectors are neither concatenated nor averaged. During inference, only the composition-conditioned prior path is used. We have clarified this in Section 4.1.

Q2. How is y defined in Section 4.2?

It is the six-dimensional lattice regression target:

yL=(aN1/3,bN1/3,cN1/3,α,β,γ),y_L=\left(\frac{a}{N^{1/3}},\frac{b}{N^{1/3}}, \frac{c}{N^{1/3}},\alpha,\beta,\gamma\right),

where a,b,c represent lattice lengths, α,β,γ\alpha, \beta, \gamma represent lattice angles, and N is the atom count. The six components are subsequently standardized using training-set statistics. We have added this definition directly to Section 4.2.

We thank the reviewer again for the constructive suggestions. We hope the additional computational analysis and lattice-refinement experiments directly address concerns (1) and (2), and we would appreciate reconsideration of the rating.

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

The paper presents CoMMat, a fast generative model for de novo crystal generation and crystal structure prediction. The model generates the crystal lattice in one shot, then uses inductive moment matching to refine the atom coordinates, using comparatively fewer iterations to other published models.

CoMMat is evaluated on the MP-2, MPTS-52, Carbon-24 and Perov-5 datasets, and compared with a set of diffusion and flow-matching baselines in both speed and quality of generation.

Critical strengths

Time comparisons are a Pandora’s box - all models support some sort of a trade off between the inference speed and accuracy. The authors report both NFE and clock speed on common hardware, this is reasonable.

From the practitioner’s point of view, once published, CoMMat will offer a way to quickly sample structures with reasonable quality.

Critical weaknesses

I Evaluation

…is not robust Evaluations use SUN with the threshold E_hull ≤ 0 eV/atom. This is a valid metric, but it is very sensitive to measurement noise:

  1. The small absolute numbers of SUN structures in some settings make some comparisons in Table 1 not statistically significant. Unless the authors show that the difference between each best and second-best is statistically significant, I suggest they uniformly highlight the entries whose SUN is statistically insignificantly different from the best. // On top of this, there is also variance caused by the training process, which is hard to measure.
  2. The tight threshold makes SUN sensitive to the fine details of the evaluation protocol (MLIP, reference hull structures relaxation and energy computation) - see Questions. This is somewhat mitigated by subjecting all the models to the same protocol - but we still don’t know whether the reported SUN difference is driven by the genuine SUN difference, or by how well does each model match CHGNet quirks. To illustrate the influence of the protocol details, see the reported SUN and model ranking dfiference between the submission and FlowMM paper:
ModelSUN, submitted (CHGNet)SUN, FlowMM (DFT)
FlowMM5.782.34
CDVAE7.051.43
DiffCSP9.053.34

A potential way to address this will be to use a publicly available DFT dataset not included in the CHGNet's training dataset to evaluate the validity of the proposed protocol. Potential options: GNoMe, WyFormer & Co (they also computed CHGNet relaxations).

…does not separate the algorithmic contribution from the training regime

Tables 1 & 2 contain two classes of results: few-step results and native budget. Few-step results compare models at equal budget - but the budget is suboptimal for the baselines by construction . An informative comparison would build the models for this budget, not truncated inference.

Native budget pits CoMMat at 32 NFE against 100+ NFE baselines. It performs similarly for de novo, and worse for CSP - a promising practical result, but we don’t know how much did the authors’ architecture innovations contribute to it. If CoMMat is allowed to match the baselines’ budget, will it outperform them? If the strongest baselines are trained and tuned from ground-up for 32 NFE inference, how much quality will they retain?

II Baseline choice

Raison d’être of this work is inference speed, and yet it chooses to compare itself only to the models whose published checkpoints have an order of magnitude slower inference. There are models which report similar speed: DiffCrysGen (~33s/10k structures), SymmBFN (~70s/10k structures).

Questions for authors

I Evaluation protocol

A. Clearly mark truncated sampling trajectories in Tables 1 & 2

Do tables 1 & 2 few-step regime contain two rather different models?

  1. Continuous-time models, where design organically allows choosing solver steps at inference time (e. g. FlowMM)
  2. Models which were trained to have N steps, but then the inference was truncated (e. g. DiffCSP)

B. Report MSUN (fraction of metastable, unique, and novel). It has larger sample size, physically meaningful (thermodynamically metastable structures very well exist in experiments), less sensitive to the details of convex hull and energy computation. Line 1186 states that metastability was computed.

C. Define the SUN evaluation procedure

  1. Which dataset was used as the novelty reference?
    1. How were structures in it relaxed? DFT/CHGNet/other?
  2. Which dataset was used as the convex hull reference?
    1. How were structures in it relaxed? DFT/CHGNet/other?
    2. How were the reference energy values computed? DFT/CHGNet/other?

II Which applications benefit from the generative model inference speed?

Could the authors please show which specific published downstream materials design pipelines are constrained by the generative model inference speed and will benefit from CoMMat?

Most pipelines follow the generative model with a relaxation and energy computation: DFT is orders of magnitude more expensive, MLIP-based relaxation is likely much slower than CoMMat. This relaxation step is hard to avoid for principled reasons. CoMMat’s reported stability is ≤11.75% (Table 11) - if we take this at a face value, it implies that 88.25% structures CoMMat produces are unstable and need to be filtered out before being used.

III Reconciling good de novo performance with bad CSP performance

CoMMat has lower MR (Table 2) compared to the top baselines - and yet similar SUN. A possible explanation is that CoMMat is capable of finding the right energy basin - but is bad at precise coordinate relaxation. This can be checked by computing CSP MR after MLIP relaxation. This is an important evaluation for CoMMat specifically. It offers generation speed - but to take advantage of it, users must use the structures it produces raw, without subjecting them to slower relaxation. Quantifying the speed-quality trade off will help them choose.

P. S. This doesn’t matter for the accept/reject decision, just a formatting bit. The submissions header says ICLR 2026.

18 comments
AuthorsOpenReview ↗

SUN Significance and MSUN (W-I.1, Q-I.B)

We thank the reviewer for emphasizing the importance of statistical significance in SUN comparisons. To address this concern quantitatively, we conducted additional evaluations using 10,000 generated structures per run, 95% Clopper–Pearson confidence intervals, and two-sided proportion tests.

W-I.1. Statistical significance of SUN comparisons.

CoMMat's few-step advantages are substantially larger than the small numerical differences observed among closely ranked native-budget methods.

At the evaluated reference settings, Crystalite achieves 12.0% SUN (95% CI: 11.6–12.4%) at 299 NFE, compared with 11.0% (10.7–11.4%) for CoMMat(D) at only 32 NFE. These results demonstrate that CoMMat achieves comparable discovery quality using substantially fewer network evaluations, while the confidence intervals provide a more precise characterization of the observed differences. Similarly, CoMMat(D) exceeds ADiT (9.7%), whereas CDVAE (7.1%) and FlowMM (6.8%) are statistically indistinguishable under the reported sampling-level test.

More importantly, CoMMat's advantage is pronounced in the intended few-step regime. On MP-20, CoMMat(D) achieves 10.5%, 10.9%, and 11.3% SUN at 2, 4, and 8 NFE, compared with 3.0%, 6.3%, and 7.3% for the strongest completed competitors of CoMMat. The differences are substantial under the reported sampling-level significance tests (p<10−26p < 10^{-26}). Actual NFE counts are explicitly reported for each model, including methods requiring multiple network evaluations per solver step.

These results establish that CoMMat's central few-step conclusion is supported by substantial measured differences rather than marginal SUN rankings. We will report uncertainty alongside the corresponding results and distinguish close comparisons from the clearly separated few-step results.

Q-I.B. Metastable SUN (MSUN).

We additionally evaluate MSUN at Ehull≤0.10E_{\mathrm{hull}} \leq 0.10 eV/atom, with the same uniqueness and novelty criteria (% of 10,000 samples):

Model2 steps4 steps8 stepsReference budget (NFE)
CoMMat(D)49.050.150.851.2 (32)
CoMMat(R)35.536.638.140.9 (32)
Crystalite0.02.037.155.2 (299)
DiffCSP---50.2 (2000)
ADiT0.228.239.846.6 (200)
OMatG11.017.520.543.6 (2127)
FlowMM13.626.033.441.9 (1000)
CrystalFlow4.016.427.740.4 (100)
CrysBFN---40.3 (1000)
ZAtom-11.31.11.222.5 (128)

Note: CrysBFN and DiffCSP uses a fixed sampling schedule; therefore, we report only its native-budget MSUN to avoid evaluating truncated sampling trajectories.

At the reference budgets, Crystalite reaches the highest MSUN (55.2%), followed by CoMMat(D) (51.2%) and DiffCSP (50.2%); CoMMat(D) and DiffCSP are not significantly different (p=0.08). At 2, 4, and 8 steps, CoMMat(D) already reaches 49.0%, 50.1%, and 50.8%, respectively, compared with at most 39.8% for any other model. CoMMat thus maintains its few-step lead under both the strict SUN and the more permissive MSUN criterion.

AuthorsOpenReview ↗

CHGNet Reliability and SUN Evaluation Protocol (W-I.2, Q-I.C)

We agree that SUN depends on the interatomic potential, energy convention, and reference hull. However, CHGNet-based screening is an established practice in crystal-generation research, including SymmCD (ICLR 2025) and SymmBFN (2026), which the reviewer explicitly recommends. Our evaluation applies a common scoring protocol to all baselines rather than combining results reported under different procedures.

W-I.2. Does CHGNet determine the exact SUN results?

To examine this concern directly, we evaluated our CHGNet screening procedure against DFT-labelled structures from WyFormer (9,910 and 2,000 structures), WBM (10,000), and a separate recall-only GNoME set (10,000).

After excluding detected structure-level overlaps with CHGNet's training data, strict CHGNet stability classification at Ehull≤0E_{\mathrm{hull}} \leq 0 achieves DFT-stability recall of 0.725–0.754, with precision of 0.31–0.45 at a standardized 7% stable prevalence. At Ehull≤0.10E_{\mathrm{hull}} \leq 0.10 eV/atom, recall increases to 0.936–0.965, with precision of 0.67–0.73 at a standardized 47% prevalence.

These results quantify the sensitivity of the strict stability threshold and demonstrate substantially improved agreement at the metastability threshold.

Alternative-potential robustness. We additionally re-scored the generated structures of all evaluated models using MACE-MP-0 and MatterSim single-point energies on CHGNet-relaxed geometries, with and without MP2020 corrections.

SUN (%) under different evaluation protocols:

ModelCHGNetMACEMACE + MP2020MatterSimMatterSim + MP2020
Crystalite11.46.49.94.67.4
CoMMat(D)10.86.08.73.75.9
ADiT9.75.17.93.45.8
DiffCSP9.14.57.32.94.6
CrysBFN8.04.67.53.65.7
CDVAE7.11.94.91.02.8
FlowMM6.83.55.52.13.5
CrystalFlow6.42.84.71.63.0
OMatG5.01.83.70.92.2
CoMMat(R)3.92.43.31.22.0
ZAtom-13.03.74.40.91.6

CoMMat(D) is evaluated at 8 NFE and CoMMat(R) at 4 NFE; the other models use the stated reference operating settings.

CoMMat(D) ranks second in all five scoring configurations. The SUN rankings exhibit Kendall correlations of 0.64–0.86 relative to CHGNet, indicating meaningful ranking consistency despite differences in absolute values.

Across 22 independently evaluated WyFormer generator sets, the correlation between CHGNet- and DFT-based stable fractions increases from Spearman 0.62 at the strict threshold to 0.92 at 0.10 eV/atom after removing detected training-set overlaps. This supports MSUN as a useful complementary screening metric, while not establishing DFT-level stability for CoMMat's own samples.

Q-I.C. Exact SUN evaluation procedure.

We clarify the complete protocol:

  • Novelty reference: MP-20 training split containing 27,136 DFT-relaxed Materials Project structures.
  • Hull reference: Materials Project phase diagram dated February 7, 2023, distributed with Matbench Discovery; reference energies use MP2020-corrected DFT values.
  • Generated structures: CHGNet 0.3.0 relaxation, maximum 1,500 steps, force threshold 0.05 eV/Å; CHGNet energies are evaluated against the reference hull without an additional MP2020 correction.
  • Stability: Ehull≤0E_{\mathrm{hull}} \leq 0 eV/atom; MSUN uses Ehull≤0.10E_{\mathrm{hull}} \leq 0.10 eV/atom.
  • Uniqueness and novelty: StructureMatcher on relaxed structures (ltol=0.2, stol=0.3, angle_tol=5°), with all generated samples included in the denominator.

The mixed CHGNet/DFT energy convention and the training-split novelty reference are explicitly identified.

These additional analyses provide quantitative evidence of both the strengths and limitations of surrogate stability screening. Although absolute strict SUN is potential-sensitive, CoMMat remains among the leading models across the evaluated scoring protocols, while its few-step CSP and runtime results provide independent support for the proposed method.

AuthorsOpenReview ↗

Algorithmic Contribution and Training Regime (W-I.3, Q-I.A)

We appreciate the reviewer’s concern regarding the influence of baseline training regimes on few-step performance. CoMMat’s central methodological contribution is enabling effective coordinate generation with extremely few network evaluations. To distinguish this contribution from training-regime effects, we provide budget-specific baseline comparisons, training-compute analyses, and a controlled moment matching-versus-flow-matching ablation.

Q-I.A. Truncated trajectories versus budget-specific training.

We distinguish flexible-step continuous-time generators (FlowMM, CrystalFlow, OMatG, Crystalite, ADiT, and ZAtom-1) from fixed-schedule generators (DiffCSP, CDVAE, SyMat, and CrysBFN), whose released checkpoints were evaluated using shortened trajectories. We clarify these distinctions in the revised tables and do not interpret truncated inference as the optimal few-step performance achievable after dedicated retraining.

W-I.3(a). How do baselines perform when trained specifically for smaller inference budgets?

To address this question, we retrained DiffCSP for 32 diffusion steps and evaluated CrysBFN's released budget-specific checkpoints. We also examined the training cost required to support different sampling budgets.

Model / training configurationEvaluation NFEMP-20 CSP MR (%)Training compute
CoMMat-512, single joint CSP/DNG checkpoint1 / 2 / 4 / 8 / 3228.7 / 42.1 / 49.3 / 50.8 / 51.65670 PFLOPs
CoMMat-256, single joint CSP/DNG checkpoint1 / 2 / 4 / 8 / 3227.3 / 41.9 / 45.1 / 46.1 / 46.3210 PFLOPs
DiffCSP, retrained for 32 steps3240.4220 PFLOPs
DiffCSP, released 1,000-step model2,00052.1220 PFLOPs
CrysBFN, released 10-step CSP checkpoint1054.5≥570 PFLOPs
CrysBFN, released 1,000-step DNG checkpoint1,000— (DNG task)≥810 PFLOPs

Training cost and sampling cost are distinct. A DiffCSP training run requires approximately 220 PFLOPs for each of its diffusion-step count, but different step-specific models require separate training runs. Their inference costs differ substantially:

DiffCSP configurationSampling NFESampling GFLOPs/structure
32-step model, predictor only (corrector off)3264
32-step model with predictor-corrector64129
1,000-step model with predictor-corrector2,0004,021

Importantly, CoMMat’s coordinate IMM formulation learns across noise levels, allowing a single checkpoint jointly trained for CSP and DNG to operate at 1, 2, 4, 8, 32, or higher NFE without budget-specific retraining or optimization.

The additional comparisons clarify the competitive landscape. Retrained DiffCSP achieves 40.4% MR at 32 NFE, versus 51.65% for CoMMat-512. However, CrysBFN’s dedicated 10-step checkpoint achieves 54.5% MR, while FlowMM reaches 50.3% and 56.5% at 4 and 8 NFE, respectively. Thus, our contribution targets the ultra-low-NFE regime, not universal superiority across budgets.

These targeted experiments address the reviewer’s concern without requiring the substantial additional effort of retraining every competing architecture across all budgets, datasets, and tasks.

W-I.3(b). Does CoMMat outperform baselines when given larger inference budgets?

CoMMat's MP-20 CSP performance largely saturates by 32 NFE: 51.65% MR at 32, 52.4% at 128, and 52.8% at 1,000 NFE. We do not claim superior native-budget CSP accuracy over the strongest baselines. Instead, CoMMat is designed to reach useful structural quality with minimal iterative refinement.

An additional oracle experiment using the ground-truth lattice increases MR from 51.65% to 60.5% at 32 NFE, indicating that lattice prediction contributes substantially to the remaining reconstruction gap.

W-I.3(c). How much does the proposed algorithm contribute independently of the training architecture?

We directly isolate the coordinate-learning objective by replacing IMM with conventional flow matching while retaining the same network architecture, training data, and remaining training settings.

MP-20 CSP MR (%)1 NFE2 NFE4 NFE8 NFE32 NFE
CoMMat (IMM)28.742.149.350.851.65
Coordinate flow matching4.525.940.844.746.6

At 1 and 2 NFE, coordinate moment matching improves CSP MR by 24.2 and 16.2 percentage points, respectively. These results identify coordinate moment matching's particular effectiveness in the extreme few-step regime. Additional component-level ablations are presented in our response to Reviewer 98At.

The controlled moment matching-versus-flow-matching ablation demonstrates the proposed objective’s substantial gains at one and two evaluations, while a single jointly trained CSP/DNG checkpoint supports multiple inference budgets.

Reviewer_98AtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CoMMat, a crystal generator built to work with very few sampling steps. CoMMat uses a causal Transformer to model the composition autoregressively and to produce a composition-conditioned latent (trained against a DimeNet++ posterior with a KL term). An MLP predicts the lattice once from that latent. Only the atomic coordinates are refined iteratively, using a CSPNet decoder trained with Inductive Moment Matching (IMM). IMM runs in a lattice-normalized Cartesian space, and the network sees wrapped fractional coordinates. Species and lattice stay fixed during refinement, so one model handles both structure prediction from a given composition (CSP) and unconditional de novo generation (DNG). On four datasets, the authors compare CoMMat with 10 baselins in the few-step setting with (2/4/8) network evaluations (NFE). They report robust quality at 2–8 NFE and competitive results at 32 NFE, about 8× faster than the fastest baseline.

Critical strengths
  1. In the 2/4/8 few-step setting, the proposed method CoMMat outperforms 10 baselines in the DNG tasks. With 32 steps quota, CoMMat is at least 8x faster than the baseline methods.
  2. The proposed design can handle both DNG and CSP tasks.
Critical weaknesses
  1. The proposed method underperforms FlowMM in most of the few-step CSP cases.
  2. The authors do not provide a clear ablation study showing which component in CoMMat actually make it work in the few-step setting.
Questions for authors

See weaknesses.

2 comments

We thank the reviewer for recognizing CoMMat's efficient de novo generation (DNG) and its unified support for crystal structure prediction (CSP) and DNG. We address the CSP comparison (W1) and clarify why few-step performance arises from CoMMat's integrated training and generation design, rather than a single detachable component (W2). All baseline comparisons in the submitted manuscript use a common evaluation protocol and scoring pipeline.

W1. CSP performance relative to FlowMM

CoMMat does not underperform FlowMM in most matched few-step CSP comparisons. In Tables 2 and 12, CoMMat achieves a higher match rate in 10 of the 12 comparisons at 2, 4, and 8 NFE across the four datasets:

DatasetModel2 NFE4 NFE8 NFE
MP-20CoMMat42.1049.3050.80
FlowMM40.9450.3056.50
MPTS-52CoMMat6.508.499.12
FlowMM0.122.194.45
Perov-5CoMMat32.8031.8033.10
FlowMM24.4525.6030.90
Carbon-24CoMMat23.9025.0025.50
FlowMM22.1224.1920.54

Entries are CSP match rates (%); bold indicates the higher value in each comparison.

Although FlowMM achieves higher MP-20 match rates at 4 and 8 NFE, CoMMat outperforms it in 10 of 12 matched few-step comparisons across four datasets, demonstrating that FlowMM's advantage is not consistent across benchmarks. At their respective native budgets, FlowMM achieves 58.82% MR at 1,000 NFE, compared with CoMMat's 51.65% at only 32 NFE. These results highlight the trade-off between prediction accuracy and sampling efficiency. CoMMat achieves strong CSP performance with substantially fewer network evaluations, consistent with its primary objective of effective few-step crystal generation, rather than maximizing accuracy at unrestricted sampling budgets. Moreover, CoMMat achieves these few-step CSP match rates using a single model trained jointly for both CSP and DNG, whereas FlowMM employs separate task-specific models. This further highlights CoMMat's efficiency and unified generation capability.

AuthorsOpenReview ↗

W2. No ablation shows which component makes few-step generation work

We thank the reviewer for raising this question. Evidently, CoMMat's few-step capability arises equally from its individual architectural components. The primary mechanism enabling few-step generation is Atomic Coordinates Moment Matching. The remaining components—autoregressive composition modeling, composition-conditioned structural representations, and one-shot lattice prediction—support the complete generation pipeline, rather than independently reducing the number of coordinate evaluations.

Why coordinate moment matching enables few-step generation. Conventional diffusion- and flow-based crystal generators typically require many iterative coordinate updates to achieve their best performance. In contrast, CoMMat uses coordinate moment matching to learn finite-time mappings between noise levels, enabling larger transitions along the generation trajectory with very few evaluations.

Why the other design choices matter. Autoregressive composition modeling and one-shot lattice prediction do not independently enable few-step coordinate generation, as similar components are used in models that still require iterative coordinate refinement. In CoMMat, these components provide effective structural conditioning, reduce unnecessary computation, and support both CSP and DNG within a single model. They can be modified while retaining coordinate moment matching, but the ablations show that the selected combination provides better overall performance.

Controlled design ablations. Although replacing IMM changes the central generative formulation, we performed this comparison and retrained additional variants to clarify the contribution of each design choice. The MP-20 CSP match rates (%) are:

Variant1 NFE2 NFE4 NFE8 NFE32 NFE
Full CoMMat28.742.1049.3050.8051.65
Coordinate flow matching instead of IMM4.525.940.844.746.6
Standard-normal prior5.98.08.48.48.5
MLP composition prior19.527.529.129.329.6
Without decoupled training22.029.031.231.832.6
Lattice-IMM (4 additional lattice evaluations)27.441.144.044.745.3
Lattice-flow (8 additional lattice evaluations)29.341.543.644.444.9
Fractional-coordinate transport27.941.544.845.646.1

The ablation results support three main findings:

(1) IMM enables few-step generation. At 2 NFE, CoMMat achieves 42.1% MR, compared with 25.9% using conventional coordinate flow matching, a 16.2 percentage-point improvement. CoMMat also performs better at 4 NFE (49.3% vs. 40.8%), 8 NFE (50.8% vs. 44.7%), and 32 NFE (51.65% vs. 46.6%). The largest gap occurs at 2 NFE, confirming IMM's particular effectiveness under extremely limited coordinate evaluations.

(2) Other components improve generation quality rather than independently enabling few-step inference. Removing the composition-conditioned prior, replacing the Transformer prior with an MLP, or removing decoupled training reduces CSP accuracy. These components provide effective structural conditioning, while IMM enables coordinate generation with very few evaluations.

(3) The architectural combination is empirically supported. Iterative lattice prediction provides no improvement over one-shot prediction despite additional evaluations, while changing the coordinate representation offers no consistent accuracy gain. These results support IMM-based coordinate generation combined with composition-conditioned modeling and one-shot lattice prediction.

Overall, IMM is the principal mechanism enabling few-step inference, while the supporting components improve structural quality and enable unified CSP/DNG generation. The ablations demonstrate that these choices serve complementary roles and justify CoMMat's overall design.

Taken together, our responses to W1 and W2 clarify CoMMat's contributions and the sources of its few-step efficiency. CoMMat outperforms FlowMM in 10 of 12 matched few-step CSP comparisons across four benchmarks, while achieving competitive generation quality at substantially lower inference cost. These ablations establish coordinate IMM as the principal mechanism enabling few-step generation, with autoregressive composition modeling, composition-conditioned structural conditioning, and one-shot lattice prediction supporting the effectiveness of the unified CSP/DNG framework. While we do not claim superiority at every sampling budget, these findings demonstrate the effectiveness of CoMMat's deliberate architectural design.

We believe the additional evidence addresses both concerns and respectfully request that the reviewer reconsider the initial assessment.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their thoughtful feedback. Your comments motivated substantial additional experiments that strengthen CoMMat's methodological justification, evaluation reliability, computational efficiency, and practical relevance. We particularly appreciate Reviewer fXXh's acknowledgment of our additional evidence and decision to increase the score. We hope the other reviewers will also consider the new results and additional analyses presented in our rebuttal, and we would greatly appreciate any further questions or discussion. In particular, Reviewer LchZ indicated in their initial review that they would consider increasing their score if Concerns 1 and 2 were resolved. Our rebuttal now addresses both concerns, and we would greatly appreciate the reviewer’s reassessment. If our responses have satisfactorily addressed your concerns, we would be grateful if you could reconsider your scores and support the acceptance of our work.

We briefly summarize how we addressed the principal concerns.

Reviewer 98At — CoMMat outperforms FlowMM in 10 of 12 matched few-step CSP comparisons across four benchmarks. Seven controlled ablations identify coordinate IMM as the principal few-step mechanism: replacing IMM with conventional flow matching reduces MP-20 CSP match rate from 42.1% to 25.9% at 2 NFE. Other components contribute to generation quality and unified CSP/DNG.

Reviewer fXXh — We conducted SUN significance tests, MSUN evaluations, external DFT-labeled validation, and independent MACE/MatterSim rescoring. We also investigated budget-specific training, fast baselines, and post-relaxation CSP. Reviewer fXXh acknowledged that the SUN evaluation concerns were addressed, recognized effective few-step IMM generation, and increased their score.

The reviewer identified pipeline-level utility as the key remaining question: why does few-step generation matter when downstream relaxation dominates cost?

Our original motivation extends beyond a single discovery pipeline. Conventional iterative crystal generation incurs substantial computational costs, limiting research that requires repeated sampling and evaluation. For instance, latent-space disentanglement, controlled latent perturbations, and conditional generation analysis require exploring large structural ensembles. By enabling rapid generate–evaluate cycles, CoMMat's few-step generation opens new opportunities for systematic latent-space exploration, model understanding, and materials discovery that would otherwise be computationally expensive.

Motivated by the reviewer’s suggestion, we additionally investigated complete discovery-pipeline efficiency:

  • A generate–screen–relax pipeline includes generation, inexpensive screening, and model-specific relaxation.
  • Screening achieves AUC = 0.81 and 7.7× SUN enrichment in the top 5% of candidates.
  • At 10% and 5% relaxation selection, CoMMat achieves 1.9× and 3.0× higher SUN/GPU-hour than Crystalite, respectively.
  • A separate fresh-sample end-to-end experiment confirms a 3.4× throughput improvement at 5% selection.
  • Stricter novelty and independent stability evaluations further assess discovery relevance.

These results show complementary benefits: faster iterative crystal exploration and higher discovery-pipeline throughput through selective screening and relaxation. The latter provides direct quantitative evidence that few-step inference can deliver meaningful computational gains beyond the generator.

Reviewer LchZ — We report FLOPs, latency, and quality–cost comparisons. At 2 NFE, CoMMat-512 achieves 42.1% CSP MR at 2.9 GFLOPs/structure, while CoMMat-256 achieves 41.9% at only 0.8 GFLOPs. Lattice ablations support one-shot prediction, and oracle experiments identify opportunities for further accuracy improvements. Both implementation questions were addressed. The computational analysis and new pipeline results directly address the motivation and practical value of few-step generation.

Overall, the rebuttal explains both how CoMMat enables effective few-step generation and why it matters for scientific experimentation and discovery workflows. The new evidence reinforces our central contribution: competitive crystal generation with extremely few evaluations using a unified, coordinate-IMM-based CSP/DNG model.

We hope the additional experiments and analyses presented during the rebuttal have satisfactorily addressed the reviewers’ concerns and further strengthened the paper. We respectfully ask the Area Chair to consider this substantial new evidence, along with any updated reviewer assessments, when making the final recommendation. We sincerely appreciate the reviewers’ and Area Chair’s time, constructive feedback, and thoughtful consideration of our work.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.