Overview
FlowMM is a generative model for materials using Riemannian flow matching. It generates crystal structures by learning flows on the Riemannian manifold of periodic structures, achieving state-of-the-art performance on CSP tasks.
Theoretical Basis
- Riemannian flow matching
- Continuous normalizing flows
- Periodic structure manifold
- Equivariant architecture
- Conditional flow matching
Key Capabilities
- Crystal structure prediction
- De novo material generation
- Riemannian geometry handling
- State-of-the-art performance
- FlowLLM extension (LLM base)
Sources: arXiv:2406.04713, Facebook Research
Key Strengths
Methodology:
- Riemannian flow matching
- Proper geometry handling
- Continuous flows
Performance:
- State-of-the-art results
- Multiple benchmarks
- Efficient sampling
Extensions:
- FlowLLM (LLM integration)
- CrystalLLM base
- Flexible architecture
Inputs & Outputs
- Input formats: Chemical composition (CSP), nothing (de novo)
- Output data types: Generated crystal structures
Interfaces & Ecosystem
- Framework: PyTorch, PyTorch Lightning
- Datasets: Standard CSP benchmarks
- Extensions: FlowLLM, CrystalLLM
Workflow and Usage
- Train on crystal dataset
- Define generation task (CSP or de novo)
- Sample from flow model
- Generate structures
- Validate candidates
Performance Characteristics
- GPU-accelerated
- Efficient sampling
- Good generation quality
Computational Cost
- Training: GPU hours
- Sampling: moderate
- Validation: DFT
Best Practices
- Use appropriate training data
- Generate multiple samples
- Validate with DFT
- Consider FlowLLM for enhanced performance
Limitations & Known Constraints
- Training data dependent
- Complex implementation
- Requires GPU
Application Areas
- Crystal structure prediction
- De novo material generation
- Materials discovery
- Generative materials science
Comparison with Other Codes
- vs DiffCSP: FlowMM flow-based, DiffCSP diffusion
- vs CDVAE: Different generative approach
- Unique strength: Riemannian flow matching, Facebook Research
Community and Support
- Open-source (GitHub)
- Facebook Research
- Active development
Verification & Sources
Primary sources:
- GitHub: https://github.com/facebookresearch/flowmm
- arXiv: https://arxiv.org/abs/2406.04713
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Website: ACTIVE (GitHub)
- Source: OPEN
- Development: ACTIVE (Meta)
- Applications: Crystal structure prediction, generative modeling