Official Resources
- Source Repository: https://github.com/usnistgov/alignn
- Documentation: https://alignn.readthedocs.io/
- PyPI: https://pypi.org/project/alignn/
- License: Open source (MIT)
Overview
ALIGNN-FF (Atomistic Line Graph Neural Network Force Field) is a GNN potential that uses line graphs to capture bond angles and dihedrals. It covers 5-118 elements and achieves competitive accuracy for both property prediction and MD simulations.
Scientific domain: Line graph GNN potential with bond angle/dihedral awareness
Target user community: Researchers needing GNN potential with explicit angular information
Theoretical Methods
- Atomistic Line Graph Neural Network
- Bond angle and dihedral features
- Line graph message passing
- 5-118 elements coverage
- Property prediction + force field
Capabilities (CRITICAL)
- Line graph architecture
- 5-118 elements coverage
- Property prediction (formation energy, bandgap)
- Force field for MD
- JARVIS integration
Sources: GitHub repository, npj Comput. Mater. 7, 185 (2021)
Key Strengths
Line Graph:
- Explicit bond angles
- Dihedral information
- Better angular features
- Physically motivated
Broad Coverage:
- 5-118 elements
- Property prediction
- Force field
- JARVIS dataset integration
NIST:
- Government maintained
- Well documented
- JARVIS ecosystem
- Reproducible
Inputs & Outputs
- Input formats: Structures (ASE/cif)
- Output data types: Properties, energies, forces
Interfaces & Ecosystem
- JARVIS: Dataset and tools
- ASE: Calculator
- PyTorch: Backend
- Python: Core
Performance Characteristics
- Speed: Fast (GPU)
- Accuracy: Competitive
- System size: 1-10000+ atoms
- Automation: Full
Computational Cost
- Training: Hours on GPU
- MD: Fast
Limitations & Known Constraints
- Not universal pretrained: Needs training
- Line graph overhead: Slightly slower
- NIST focus: JARVIS benchmarks
Comparison with Other Codes
- vs CGCNN: ALIGNN has line graph, CGCNN does not
- vs M3GNet: ALIGNN is line graph, M3GNet is 3-body
- Unique strength: Line graph GNN with explicit bond angle/dihedral features covering 5-118 elements
Application Areas
Property Prediction:
- Formation energy
- Bandgap
- Elastic properties
- JARVIS benchmarks
MD:
- Materials simulation
- Structure relaxation
- Energy evaluation
Best Practices
- Use JARVIS training data
- Start with pretrained models
- Fine-tune for target systems
Community and Support
- Open source (MIT)
- NIST maintained
- PyPI installable
- ReadTheDocs documentation
Verification & Sources
Primary sources:
- GitHub: https://github.com/usnistgov/alignn
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- PyPI: AVAILABLE
- Specialized strength: Line graph GNN with explicit bond angle/dihedral features covering 5-118 elements