Official Resources
- Source Repository: https://github.com/materialsvirtuallab/matgl
- Documentation: https://matgl.readthedocs.io/
- PyPI: https://pypi.org/project/matgl/
- License: Open source (MIT)
Overview
matgl (Materials Graph Library) is a graph deep learning library for materials properties. It implements M3GNet, CHGNet, TensorNet, and MEGNet models with pretrained weights, providing a unified interface for property prediction and potential energy surface modeling.
Scientific domain: Graph deep learning for materials, unified MLIP framework
Target user community: Researchers needing unified access to multiple GNN potentials
Theoretical Methods
- M3GNet (3-body graph network)
- CHGNet (charge-informed GNN)
- TensorNet (tensor-based equivariant)
- MEGNet (materials graph network)
- CGCNN (crystal graph CNN)
Capabilities (CRITICAL)
- Multiple GNN models in one package
- Pretrained universal potentials
- Property prediction (formation energy, bandgap, etc.)
- Potential energy surface modeling
- ASE calculator interface
- PyTorch backend
Sources: GitHub repository
Key Strengths
Unified Framework:
- M3GNet, CHGNet, TensorNet, MEGNet
- Same API across models
- Easy model comparison
- Pretrained weights
Property Prediction:
- Formation energy
- Bandgap
- Elastic properties
- Bulk modulus
Integration:
- ASE calculator
- pymatgen structures
- PyTorch backend
Inputs & Outputs
- Input formats: Structures (pymatgen/ASE)
- Output data types: Properties, energies, forces, stresses
Interfaces & Ecosystem
- pymatgen: Structure handling
- ASE: Calculator
- PyTorch: Backend
- DGL: Graph library
Performance Characteristics
- Speed: Fast (GPU)
- Accuracy: Model-dependent
- System size: Any
- Automation: Full
Computational Cost
- Inference: Milliseconds
- Training: Hours on GPU
Limitations & Known Constraints
- DGL dependency: Deep Graph Library required
- Model-specific limitations: Inherited from each model
- Memory: Large graphs need GPU
Comparison with Other Codes
- vs individual packages: matgl unifies multiple models
- vs MACE: matgl is multi-model, MACE is single architecture
- Unique strength: Unified framework for M3GNet, CHGNet, TensorNet, MEGNet with pretrained weights
Application Areas
Materials Property Prediction:
- Formation energy prediction
- Bandgap estimation
- Elastic property calculation
- Screening studies
MLIP:
- MD with M3GNet/CHGNet
- Structure relaxation
- Energy evaluation
Best Practices
- Use pretrained models for quick results
- Compare multiple architectures
- Validate critical predictions with DFT
Community and Support
- Open source (MIT)
- PyPI installable
- Materials Virtual Lab maintained
- ReadTheDocs documentation
Verification & Sources
Primary sources:
- GitHub: https://github.com/materialsvirtuallab/matgl
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- PyPI: AVAILABLE
- Specialized strength: Unified framework for M3GNet, CHGNet, TensorNet, MEGNet with pretrained weights