Official Resources
- Homepage: https://github.com/mir-group/flare
- Documentation: https://flare.readthedocs.io/
- Source Repository: https://github.com/mir-group/flare
- License: MIT
Overview
FLARE (Fast Learning of Atomistic Rare Events) is an open-source Python package for creating fast and accurate interatomic potentials using Gaussian process regression with Bayesian active learning. It enables on-the-fly training during molecular dynamics simulations.
Scientific domain: Bayesian ML potentials, active learning, on-the-fly training
Target user community: Researchers needing uncertainty-aware ML potentials
Theoretical Methods
- Gaussian process regression
- Sparse Gaussian processes
- Bayesian active learning
- Mapped Gaussian processes
- On-the-fly learning
Capabilities (CRITICAL)
- On-the-fly potential training
- Bayesian uncertainty quantification
- Active learning
- Sparse GP for efficiency
- LAMMPS integration
- ASE calculator
Key Strengths
Uncertainty Quantification:
- Bayesian framework
- Prediction uncertainties
- Active learning
- Automatic data selection
On-the-fly Learning:
- Train during MD
- Automatic DFT calls
- Efficient data collection
Inputs & Outputs
-
Input formats:
- ASE Atoms
- DFT calculator interface
-
Output data types:
- Energies with uncertainties
- Forces with uncertainties
- Model files
Interfaces & Ecosystem
- ASE: Calculator
- LAMMPS: Integration
- VASP/QE: DFT backends
- flare_pp: C++ acceleration
Advanced Features
- On-the-fly: Train during MD
- Sparse GP: Scalable inference
- Mapped GP: Fast evaluation
- Active learning: Automatic sampling
- Uncertainty: Bayesian errors
Performance Characteristics
- Mapped GP very fast
- Uncertainty adds overhead
- Good scaling
- C++ acceleration available
Computational Cost
- Training: Automatic during MD
- Inference: Fast with mapping
- DFT calls: As needed
- Overall: Efficient active learning
Best Practices
- Start with small systems
- Validate uncertainty estimates
- Use mapped GP for production
- Monitor DFT call frequency
Limitations & Known Constraints
- GP scaling with data
- Requires DFT backend
- Complex setup
- Active development
Application Areas
- Materials discovery
- Phase transitions
- Rare events
- Defect dynamics
- Surface reactions
Comparison with Other Codes
- vs NequIP/MACE: FLARE Bayesian with uncertainty, others deterministic NN
- vs DeepMD-kit: FLARE active learning, DeepMD fixed training
- vs N2P2: FLARE GP-based, N2P2 neural network
- Unique strength: On-the-fly training, Bayesian uncertainty, active learning
Community and Support
- Active development (Harvard)
- GitHub issues
- Good documentation
- Growing community
Verification & Sources
Primary sources:
- GitHub: https://github.com/mir-group/flare
- J. Vandermause et al., npj Comput. Mater. 6, 20 (2020)
Secondary sources:
- FLARE tutorials
- flare_pp C++ documentation
- Active learning publications
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: OPEN (GitHub, MIT)
- Published in npj Computational Materials
- Academic citations: >300
- Active development: Harvard group