acceptodds
Under review as a conference paper at ICLR 2027

DynaCrys: Crystal Generation with Dynamic Space-Group Diffusion

Abstract

The search for new crystalline materials spans an enormous compositional and structural space. Generating candidates in this space requires modeling the relationships between elemental composition, crystallographic symmetry, and crystal geometry. The space group specifies the crystal's symmetry and determines the admissible Wyckoff positions and geometric degrees of freedom; fixing it in advance prevents this global symmetry choice from being revised using information from evolving Wyckoff occupations and elemental composition. We introduce DynaCrys, the first framework for complete crystal generation in which the space group co-evolves with Wyckoff occupations and elements. We develop physics-informed space-group diffusion that explicitly incorporates crystallographic group–subgroup relations, and a legality-constrained stochastic decoder that jointly samples compatible space groups, Wyckoff occupations, and elements, followed by symmetry-constrained geometry generation. In large-scale evaluations on standard benchmarks, DynaCrys achieves state-of-the-art performance in symmetry-aware crystal generation while supporting fast sampling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
crystal generation, diffusion models, crystallographic symmetry, space groups, materials discovery

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_iq4kOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DynaCrys generates crystals by jointly updating the space group, Wyckoff occupations, and elements, followed by symmetry-constrained geometry generation. Its main contribution is to make the space-group choice revisable during symbolic generation, using group–subgroup relations and constrained sampling. Results on two benchmarks, supported by ablations and independent computational validation, make the approach convincing.

Critical strengths
  1. The paper offers a meaningful extension of symmetry-aware crystal generation. Allowing symmetry and composition to evolve together addresses a clear limitation of fixing the space group in advance, and the constrained decoder makes these choices compatible.

  2. The experimental support is substantial. Comparisons on two benchmarks, component ablations, and independent stability evaluations support the method's effectiveness, including its ability to generate stable, unique, and novel candidates efficiently.

Critical weaknesses
  1. Although DynaCrys improves structural quality and discovery rates, its compositional validity remains below some baselines on both benchmarks (Tables 4–5). The paper would benefit from explaining this gap and whether it can be reduced while preserving the discovery gains.

  2. The symbolic model restricts space-group choices to those represented in the training data (Appendix A.1, lines 655–658). While this still allows novel crystals within these classes, the applicability of dynamic space-group modeling beyond the observed classes has not been established.

Questions for authors
  1. What explains the lower compositional validity?

  2. Can the dynamic space-group model extend to symmetry classes absent from training, and what would be needed to evaluate this?

Reviewer_Xxe9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes DynaCrys, a crystal generation method. Compared with existing methods that are either un-aware of the symmetries or fix the space group before generating the remaining variables, DynaCrys co-evolves the space group with Wyckoff occupations and elements. The proposed DynaCrys method outperforms the symmetry-unaware baselines and the fixed space group baselines. It is also able to learn the distribution of the space groups within the training set. Extensive ablation reveals that each building block of the method contributes to its improved performance.

Critical strengths
  1. The proposed method outperforms fixed space-group baselines on MP20 and MPTS52 datasets. This demonstrate the advantage of co-evolving the space groups in the diffussion process.
  2. The authors conduct extensive ablation experiments that demonstrate the importance of each building components of DynaCrys.
  3. A smaller version of DynaCrys pushes the Pareto frontier over generation quality and efficiency.
Critical weaknesses
  1. Incomplete baseline comparison. In the materials generation experiments (Table 2, 3, 4, 5), the authors only compare with fixed space-group methods. How do the symmetry un-aware methods (e.g., FlowMM, MatterGen) perform compare to DynaCrys in these tables? Although the authors compare the space-group distribution with symmetry un-aware methods in Figure 3, this can only demonstrate the method learns the distribution of the space-groups better but not that DynaCrys is better generation method. I will raise my score if this question is resolved.
Questions for authors

See weaknesses.

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

The authors propose a crystal generative model which samples sequentially. First, the model does joint diffusion over discrete variables (space groups, Wyckoff positions, numbers of atoms, and atom types) where constraints are enforced at inference time via rejection sampling. Conditioned on these discrete variables, the model leverages prior work to diffuse continuous variables (crystal lattice and atom positions), constrained via per-step projections. They achieve strong empirical performance across several metrics and datasets, including with quantum mechanical simulations.

Critical strengths
  • Distinctly leverages crystallographic group-subgroup relationships for the forward noising of space groups
  • Strong empirical performance with evaluation from density functional theory simulations
Critical weaknesses
  • Lacking non-symmetry constrained baselines like MatterGen [1] which make it difficult to assess whether the empirical results are state-of-the-art
  • Writing was difficult to follow due to unconventional phrasing (e.g. "legality-constrained stochastic decoder" seems to just be rejection sampling, "symbolic framework" seems like odd phrasing for discrete variables)

[1] https://www.nature.com/articles/s41586-025-08628-5

Questions for authors
  • I didn't understand the point of the elementary symmetric polynomial in Eq 7.
Reviewer_ddviOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper presents a diffusion method for generating crystal materials, including elemental composition, atom coordinates, space group, Wyckoff occupation, and lattice parameters. The approach has two stages, 1. generating space group, Wyckoff occupances, and elements, and 2. lattice parameters and free parameters determining coordinates. The method is demonstrated on MP-20 and MPTS-52 and appears to generate structures that better match the data in terms of symmetries compared to symmetry unaware methods.

Critical strengths
  • The approach is a fairly complicated design, but design choices are reasonable and the method appears to provide an advantage.
  • Evaluation with VASP is appreciated.
  • Additional details available in a comprehensive supplement
Critical weaknesses
  • The writing is difficult to follow at times. The proposed method is fairly complex, and the presentation could improve by including additional pedagogical illustrations.
  • A direct comparison with e.g. WyckoffDiff wrt. generating protostructures would be informative, but is not included.
  • I am not completely convinced about the benefits of jointly generating space group with occupations/elements, rather than conditioning on it (sampling from prior/data distribution). Although the paper includes a ablation, I believe some details are missing.
Questions for authors

The following are all minor comments/questions that do not materially influence my assessment. The authors need not respond.

l.52: The introduction emphasizes the use of group-subgroup structure. Could you include a short explanation early in the paper on this?

l.73: Could you make Figure 1 more self contained? For example t and k are not defined, and it is unclear at first reading what other notation means. I did not get a complete understanding of the framework from the figure.

l.83: In your third contribution, it is not clear to me why one would need to sum over all legal configurations?

l.126: Could you briefly mention that Rg and tg are rotation and translations when introducing space groups?

l.133: Where is RG defined?

l.137: In Figure 2, you use a plane group used for illustration (rather than a space group)?

l.172: Did you consider using double braces to denote a multiset?

l.179-189: The description here is difficult to understand. Would you consider creating an illustration of the representation to help the reader?

l.192: It is not clear to me from the description here what is done in a continuously relaxed space? Diffusion over a graph implies discrete diffusion whereas the use of a codebook seems to imply a common vector space is used.

l.195: Rt and piG are not explicitly defined.

l.244: Does the sum range over all subsets of {1,...,M} of cardinality K?

l.253: I am not sure what exactly is meant by "without evaluating the global normalizer"? Isn't eK the normalizer, which is evaluated by an explicit sum?

l.323: Should it perhaps be mentioned that methods that condition on the space group can trivially realize any desired space group distribution, including the data distribution?

l.428: How is the fixed space group used to condition the remaining model components?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.