SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models
Daniel Levy, Siba Smarak Panigrahi, S\'ekou-Oumar Kaba, Qiang Zhu, Kin Long Kelvin Lee, Mikhail Galkin, Santiago Miret, Siamak Ravanbakhsh
TL;DR
SymmCD is a diffusion-based generative model that explicitly incorporates crystallographic symmetry to generate diverse, valid, and symmetry-preserving crystalline materials, advancing the design of novel crystals for various applications.
Contribution
We introduce SymmCD, a novel diffusion model that explicitly encodes symmetry in crystal generation, enabling the creation of realistic and diverse crystalline structures.
Findings
SymmCD generates diverse, symmetry-preserving crystals with realistic properties.
The model outperforms existing methods in generating valid crystal structures.
SymmCD generalizes across different crystallographic symmetry groups.
Abstract
Generating novel crystalline materials has the potential to lead to advancements in fields such as electronics, energy storage, and catalysis. The defining characteristic of crystals is their symmetry, which plays a central role in determining their physical properties. However, existing crystal generation methods either fail to generate materials that display the symmetries of real-world crystals, or simply replicate the symmetry information from examples in a database. To address this limitation, we propose SymmCD, a novel diffusion-based generative model that explicitly incorporates crystallographic symmetry into the generative process. We decompose crystals into two components and learn their joint distribution through diffusion: 1) the asymmetric unit, the smallest subset of the crystal which can generate the whole crystal through symmetry transformations, and; 2) the symmetry…
Peer Reviews
Decision·ICLR 2025 Poster
• The manuscript's structure and clarity are excellent overall. • The manuscript includes a well-written and comprehensive introduction, with a clear and well-developed motivation for the crystal generation problem as an application of diffusion models. • The method is well-formalized and understandable even to non-experts in crystal generation. • Experimental tasks and evaluation: The authors assess their method and the baselines on relevant additional tasks, such as S.U.N. structure predict
• Introduction: From my perspective, the problem of generating symmetric crystals is closely related to other structure generation tasks in general representation learning. For instance, in biological applications, such as neuron structure generation or vascular structure generation, it would be beneficial if the authors discussed the relation to other domains in structure generation and the types of methods that have been developed. For example, I see certain similarities to diffusion methods i
- The developed method for vectorizing crystalline structures, which explicitly accounts for both the spatial symmetry of the crystal and the point symmetry of the orbits, is to my knowledge the first of its kind, therefore unique, and holds a great promise for application in crystal structure prediction (CSP) for both inorganic and organic crystals. - The article is well-structured and clearly conveys information, allowing individuals unfamiliar with this field to understand the crystallographi
The major weaknesses of the paper lie in the discussion of the obtained empirical results. Addressing these will significantly enhance the presentation of the work accomplished: 1. Please indicate in the introduction that the initial focus is on non-molecular/inorganic crystals. 2. Please mention in the conclusion remarks that your method of structural representation seems to be well-suited for molecular crystals as well. For the latter, the presence of intrinsic point symmetry and its interac
1. Innovation in Representation: The paper introduces a physically motivated representation based on crystallographic symmetry, using a binary matrix to encode symmetry, which addresses data fragmentation and enables generalization across symmetry groups. 2. Computational Efficiency: By focusing on asymmetric units rather than full crystal structures, the model demonstrates significant improvements in memory usage and training speed, an aspect well-supported by experimental evidence. 3. Diversit
1.Comprehensive Evaluation of Generated Crystal Properties: While the model’s ability to generate symmetric and diverse crystals is demonstrated, additional quantitative evaluations of properties such as thermodynamic and mechanical stability would further solidify the model’s applicability to real-world scenarios. Metrics that reflect physical applicability, such as structural stability under various conditions, could significantly strengthen the evaluation section. 2.Efficiency on Larger Datas
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsCrystallization and Solubility Studies