Overview
CDVAE (Crystal Diffusion Variational Autoencoder) is an SE(3)-invariant autoencoder for generating periodic crystal structures. It combines variational autoencoders with diffusion models to generate stable materials.
Theoretical Basis
- Variational autoencoder (VAE)
- Score-based diffusion model
- SE(3) equivariance
- Periodic structure handling
- Energy-guided generation
Key Capabilities
- Periodic material generation
- Stable structure generation
- Property optimization
- Reconstruction and generation
- SE(3) invariance
Sources: ICLR 2022, arXiv:2110.06197
Key Strengths
Architecture:
- SE(3)-invariant
- Diffusion-based decoder
- VAE latent space
Generation:
- Stable structures
- Diverse compositions
- Property-guided
Validation:
- Energy evaluation
- Stability checks
- Benchmark datasets
Inputs & Outputs
- Input formats: Crystal structures (training), latent vectors (generation)
- Output data types: Generated crystal structures
Interfaces & Ecosystem
- Framework: PyTorch, PyTorch Geometric
- Datasets: MP-20, Carbon-24, Perov-5
- Evaluation: DFT validation
Workflow and Usage
- Train on crystal structure dataset
- Encode structures to latent space
- Sample from latent space
- Decode with diffusion process
- Evaluate generated structures
Performance Characteristics
- GPU-accelerated training
- Fast generation after training
- Good reconstruction accuracy
Computational Cost
- Training: GPU hours
- Generation: fast
- Evaluation: DFT-dependent
Best Practices
- Use appropriate training data
- Validate with DFT
- Check stability metrics
- Use property guidance
Limitations & Known Constraints
- Training data dependent
- May generate unstable structures
- Limited to training distribution
Application Areas
- Materials discovery
- Crystal structure generation
- Property-guided design
- Stable material prediction
Comparison with Other Codes
- vs DiffCSP: CDVAE VAE-based, DiffCSP pure diffusion
- vs MatterGen: Different architectures
- Unique strength: VAE+diffusion, SE(3) invariance, ICLR 2022 landmark
Community and Support
- Open-source (GitHub)
- MIT/Meta AI development
- Widely cited
Verification & Sources
Primary sources:
- GitHub: https://github.com/txie-93/cdvae
- Paper: ICLR 2022
- arXiv: https://arxiv.org/abs/2110.06197
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Website: ACTIVE (GitHub)
- Source: OPEN
- Development: LANDMARK (2022)
- Applications: Crystal structure generation, materials discovery