acceptodds
Under review as a conference paper at ICLR 2027

Optimal Crystal Flow for Fast and Reliable Crystal Generation

Abstract

crystal generation aims to discover novel materials by exploring the vast combinatorial space of atomic species and lattice geometries. While recent models based on diffusion and flow matching have shown strong generative capabilities, they often require a large number of sampling steps. More importantly, they struggle to generate reliable structures in out-of-distribution regimes, especially for systems larger than those seen during training. To address these limitations, we introduce (OCFlow), a flow matching framework that reduces unnecessarily long and curved coordinate paths by resolving ambiguity within each crystal. Experiments on the MP-20 and MPTS-52 benchmarks show that OCFlow is superior to or competitive with state-of-the-art baselines in sample validity and stability, with particularly strong performance on larger out-of-distribution systems while remaining effective with as few as ten network function evaluations. Together, these results show that per-structure optimal couplings provide both efficient sampling and robust generalization in crystal generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
De novo Crystal Generation, Flow matching, Few step generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nnGkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose an all-atom joint generative model, OCFlow, for inorganic materials de novo generation (DNG) which leverages torus-aware optimal transport coupling to yield stronger out-of-distribution inference and more efficient sampling measured in terms of NFE. They demonstrate the efficacy of their proposed architecture on unit cells of varying size, showcasing strong size-transferability performance and claim robustness at low NFE.

The main contributions are:

  • A torus-aware optimal transport coupling.
  • Quantitative evidence that OCFlow balances the speed vs. quality tradeoff in DNG).
  • An MLIP-based benchmark for evaluating generative models for DNG is proposed and used to stress-test OCFlow.

OCFlow is validated using stability/metastability, uniqueness, and novelty (SUN/MSUN) metric as well as structural and compositional validity rates with respect to competing models. Formation energy of generated structures is also reported demonstrating a lower-shifted density of EfE_f for OCFlow. OCFlow is competitive with respect to state-of-the-art models.

Overall the paper is well-written and the results support the central claim that optimal transport coupling during training results in more robust OOD sampling. My main concerns regarding this work are the proposed novelty of the optimal transport coupling for periodic fractional coordinates, the purported contribution of a new benchmark, as well as a lacking comparison to existing models. An additional concern, though not critical, is that while validity remains high for low-NFE sampling, the performance of the model in terms of SUN and MSUN drops dramatically.

Critical strengths

The most convincing demonstration of OCFlow is its performance on size-transferable inference for inorganic crystals. it is clear based on the results in Table 2 and Figure 2 that OOD performance is superior to other architectures. It is also clear that OCFlow trained on MP-20 and MPTS-52 is more robust to OOD systems than other models while remaining strong in distribution. I think that size-transferable generation is a consequential and yet relatively underexplored avenue for inorganic DNG and the authors demonstrate that, at least according to validity and MSUN, OCFlow is well-suited to this task.

Critical weaknesses

Broadly speaking I think that some of the central claims are overstated. Firstly, I believe that the proposed optimal transport coupling is not completely novel. While the authors present torus-aware OC for fractional coordinate flow matching as their first novel contribution, it is my understanding that the Open Materials Generation (OMatG) model attempted something very similar to this already. A high-level comparison of the two approaches and how they differ would make it more clear why the optimal coupling proposed by OCFlow is advantageous and distinct. Similarly, I think that the claimed "few-step benchmark" is somewhat of an overstatement. The authors rather seem to apply well-known benchmarks and evaluation metrics to existing models at various NFE but I am not convinced that this is enough to constitute the claims of a new benchmark.

A second key weakness that I believe should be addressed during the rebuttal period is more comparisons with existing models. There are a few noticeable missing competitors like Crystalite, MiAD, and OMatG. Crystalite, for example, is the current SOTA on LeMat-GenBench and, thus, a comparison with them seems quite relevant. How does OCFlow compare to these models when considering standard and low-NFE inference? On this note, the best LeMat-GenBench SUN rate on MP-20 models has bee reported over 1 % whereas in Figure 3 it does not appear that OCFlow is at this level. Clarification on if OCFlow structures were pre relaxed before SUN computation would help shed light on this.

Questions for authors
  • There seem to be many different proxies for thermodynamic stability used in this work. Tables 1 and 2 list stability in terms of formation energy but do these rankings hold when considering energy above the convex hull?
  • How does the ablation in Table 5 change if one considers 100 or 1000 NFE? It is a strong finding that optimal transport coupling improves performance appreciably for 10 NFE but is this consistent at greater NFE as well?
Reviewer_FoGvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OCFlow is a flow-matching generator for periodic crystals, targeting inference time and out of distribution. For each (prior, data) pair, the species-blind prior sites are matched to the data sites by a Hungarian assignment under the periodic minimum-image cost, and the mean displacement is subtracted. The paper reports near-perfect structural validity, lower formation energy and relaxation RMSD, and graceful degradation. It attributes these gains to shorter, less ambiguous coordinate paths.

Critical strengths
  • The problems addressed (throughputs and extrapolation to larger cells) are important and open.
  • The findings & method are useful for crystal designers.
Critical weaknesses
  • The scientific claim is about the coupling. However, the evidence says a strong recipe plus a cheap coupling that helps a little.
  • The main technical idea has prior works:
    • Equivariant Flow Matching (Klein, Krämer and Noé, NeurIPS 2023) aligns each prior-data pair of interchangeable particles with the Hungarian algorithm (plus Kabsch) to shorten paths, with an accompanying theory of the equivariant OT coupling.
    • OMatG (Höllmer et al., ICML 2025) has an optional data-dependent coupling that permutes prior atoms to minimise fractional-coordinate distance on the torus.
Questions for authors
  • Is "without OC" in Table 5 an independently trained model with the same recipe? Did it use the same sampling constants (c, r, species temperature schedule), or were they retuned for it? If the uncoupled model was sampled with constants tuned for the coupled one, the reported gap is not a clean comparison.
  • Are the standard deviations in Tables 1 and 2 over sampling seeds of one checkpoint (as Appendix B.1 suggests) rather than over independent training runs?
  • How does your coupling differ technically from OMatG's minimum-permutational-distance coupling and from Klein et al.'s per-pair Hungarian alignment?
Reviewer_h7YCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OCFlow identifies intra-sample ambiguity—arising from atom permutations, periodic images, and global translations—as an important source of unnecessarily long coordinate paths in crystal flow matching. It introduces a periodicity-aware local optimal coupling that uses minimum-image distances and Hungarian matching to determine atom correspondences, followed by a translation correction that shortens the resulting transport paths. Experiments demonstrate that this formulation improves few-step generation, relaxation quality, and generalization to crystals containing more atoms than those observed during training, while adding only modest training overhead and no extra inference-time assignment cost.

Critical strengths
  1. Principled geometric formulation. The periodic Hungarian matching and translation correction provide a clear solution to intra-sample atom-correspondence ambiguity. 2, Strong few-step and OOD performance. OCFlow preserves high validity and good relaxation behavior with as few as 10 NFEs, particularly for systems larger than those seen during training.
  2. Careful statistical and computational evaluation. The paper uses three sampling seeds, large sample sets, extensive NFE sweeps, retrained baselines, and explicit runtime measurements.
Critical weaknesses
  1. Low absolute discovery yield. The best strict S.U.N. rate is only about 0.25% on MP-20, while no strict S.U.N. candidates are obtained in the OOD setting.
  2. No DFT validation. Stability and relaxation conclusions are based entirely on MLIPs.
  3. The primary metrics are not sufficient evidence of material stability. Structural validity, small relaxation displacement, and negative formation energy do not necessarily imply thermodynamic stability or synthesizability.
  4. The central ablation is limited. The comparison with and without optimal coupling is mainly reported at 10 NFEs on MP-20, while sampling hyperparameters also vary across datasets and NFE budgets.
Questions for authors

Same as weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.