Overview
DiffCSP is a crystal structure prediction method using joint equivariant diffusion on lattice and atomic coordinates. It treats CSP as a conditional generation task given chemical composition.
Theoretical Basis
- Equivariant diffusion models
- Joint lattice and coordinate diffusion
- Periodic E(3) equivariance
- Conditional generation
- Score-based denoising
Key Capabilities
- Crystal structure prediction from composition
- Joint lattice and atom diffusion
- Equivariant architecture
- Conditional generation
- State-of-the-art performance
Sources: NeurIPS 2023, arXiv:2309.04475
Key Strengths
Equivariance:
- Periodic E(3) equivariance
- Proper symmetry handling
- Rotation/translation invariance
Joint Diffusion:
- Lattice and coordinates together
- Consistent generation
- Physical constraints
Performance:
- State-of-the-art results
- Multiple benchmarks
- Competitive accuracy
Inputs & Outputs
- Input formats: Chemical composition, atom types
- Output data types: Predicted crystal structures
Interfaces & Ecosystem
- Framework: PyTorch
- Datasets: MP-20, Perov-5, Carbon-24, MPTS-52
- Evaluation: Match rate, RMSD
Workflow and Usage
- Train on crystal structure dataset
- Input chemical composition
- Run diffusion sampling
- Generate candidate structures
- Rank and validate
Performance Characteristics
- GPU-accelerated
- Fast sampling
- Good match rates
Computational Cost
- Training: GPU hours
- Sampling: moderate
- Validation: DFT-dependent
Best Practices
- Use appropriate training data
- Generate multiple samples
- Validate top candidates
- Check structural validity
Limitations & Known Constraints
- Training data dependent
- Complex compositions challenging
- Requires validation
Application Areas
- Crystal structure prediction
- Materials discovery
- Composition-to-structure
- High-throughput screening
Comparison with Other Codes
- vs CDVAE: DiffCSP pure diffusion, CDVAE VAE+diffusion
- vs EquiCSP: Similar approach, different implementation
- Unique strength: Joint equivariant diffusion, NeurIPS 2023
Community and Support
- Open-source (GitHub)
- Academic development
- Active research area
Verification & Sources
Primary sources:
- GitHub: https://github.com/jiaor17/DiffCSP
- Paper: NeurIPS 2023
- arXiv: https://arxiv.org/abs/2309.04475
Secondary sources:
- DiffCSP++ (ICLR 2024): https://github.com/jiaor17/DiffCSP-PP
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Website: ACTIVE (GitHub)
- Source: OPEN
- Development: ACTIVE
- Applications: Crystal structure prediction, generative modeling