SymmCD

Official Resources

• Homepage: https://openreview.net/forum?id=xnssGv9rpW
• Source Repository: https://github.com/sibasmarak/symmcd
• Documentation: https://github.com/sibasmarak/symmcd
• License: As specified in repository (open-source)

Overview

SymmCD (Symmetric Crystal Diffusion) is a novel diffusion-based generative model that explicitly incorporates crystallographic symmetry into the crystal generation process. Presented at ICLR 2025, SymmCD addresses a key limitation of existing crystal generation methods: the failure to generate materials displaying the symmetries of real-world crystals, or simply replicating symmetry information from database examples.

SymmCD decomposes crystals into two components and learns their joint distribution through diffusion: (1) the asymmetric unit, the smallest subset of the crystal that can generate the whole crystal through symmetry transformations, and (2) the symmetry transformations needed to be applied to each atom in the asymmetric unit. The method uses a novel and interpretable representation for these transformations based on Wyckoff positions, enabling generalization across different crystallographic symmetry groups. SymmCD demonstrates competitive performance on a subset of the Materials Project, producing diverse and valid crystals with realistic symmetries and predicted properties.

Scientific domain: Generative models for crystals, crystallographic symmetry, materials discovery
Target user community: ML researchers and materials scientists in generative modeling

Theoretical Methods

• Diffusion-based generative modeling
• Crystallographic symmetry (space groups, Wyckoff positions)
• Asymmetric unit decomposition
• Joint distribution learning via diffusion
• Interpretable symmetry transformation representation
• SE(3) equivariance for crystal generation
• Generalization across crystallographic symmetry groups

Capabilities (CRITICAL)

• Generation of crystals with realistic crystallographic symmetries
• Explicit incorporation of space group symmetry
• Asymmetric unit + symmetry transformation decomposition
• Wyckoff position-based interpretable representation
• Generalization across different space groups
• Diverse and valid crystal generation
• Competitive performance on Materials Project subset
• Predicted properties for generated crystals

Inputs & Outputs

Input formats:

• Training data from Materials Project (crystal structures with symmetry)
• Space group and Wyckoff position specifications
• Standard crystal representations (CIF, etc.)

Output data types:

• Generated crystal structures with explicit symmetry
• Asymmetric unit specifications
• Space group assignments
• Predicted material properties
• Validity and diversity metrics

Interfaces & Ecosystem

• Programming language: Python (PyTorch-based)
• Source: GitHub repository (sibasmarak/symmcd)
• Dataset: Materials Project subset
• Related models: CDVAE, CrystalDiff, FlowMM
• Platforms: Linux, GPU required

Limitations & Known Constraints

• Research code, not production-ready
• Limited to crystallographic space groups in training data
• Requires GPU for training and generation
• Diffusion process can be computationally expensive
• Quality depends on training data coverage of space groups

Performance Characteristics

• Diffusion-based generation: iterative denoising process
• Competitive with state-of-the-art crystal generation methods
• Symmetry-aware generation reduces invalid structures
• Training requires GPU and Materials Project data
• Generation produces structures closer to valid crystals

Comparison with Other Codes

• vs CDVAE: CDVAE doesn't explicitly model symmetry; SymmCD generates with correct space groups
• vs FlowLLM: FlowLLM uses LLM + flow matching; SymmCD uses diffusion with symmetry decomposition
• vs CrystalDiff: CrystalDiff uses diffusion without explicit symmetry; SymmCD incorporates Wyckoff positions
• vs USPEX: USPEX uses evolutionary search; SymmCD uses generative diffusion

Best Practices

• Ensure training data covers diverse space groups
• Use Wyckoff position representation for interpretable generation
• Validate generated structures with symmetry analysis tools
• Evaluate using diversity, validity, and novelty metrics
• Compare symmetry distribution with real crystal databases

Verification & Sources

Primary sources:

1. ICLR 2025 paper: https://openreview.net/forum?id=xnssGv9rpW
2. GitHub: https://github.com/sibasmarak/symmcd
3. D. Levy et al., "SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models," ICLR 2025
4. arXiv: https://arxiv.org/abs/2502.03638

Confidence: VERIFIED - Published at ICLR 2025 with GitHub repository available