Official Resources
- Source Repository: https://github.com/fenggo/I-ReaxFF
- License: Open source
Overview
I-ReaxFF (Intelligent-Reactive Force Field) is a differentiable ReaxFF framework based on TensorFlow. It enables gradient-based optimization and neural network augmentation of ReaxFF, combining reactive force fields with message passing neural networks.
Scientific domain: Differentiable ReaxFF with neural network augmentation
Target user community: Researchers combining ReaxFF with neural networks for reactive MD
Theoretical Methods
- Differentiable ReaxFF (TensorFlow)
- ReaxFF-MPNN hybrid potential
- Gradient-based optimization
- Neural network augmentation
- Higher-order derivatives
Capabilities (CRITICAL)
- Differentiable ReaxFF
- ReaxFF-MPNN hybrid
- Gradient-based fitting
- Neural network augmentation
- LAMMPS integration
Sources: GitHub repository, Phys. Chem. Chem. Phys. 23, 19457 (2021)
Key Strengths
Differentiable:
- Exact gradients
- Higher-order derivatives
- TensorFlow backend
- End-to-end optimization
Hybrid:
- ReaxFF + MPNN
- Best of both worlds
- Reactive chemistry
- ML accuracy boost
Integration:
- LAMMPS compatible
- ReaxFF-nn for LAMMPS
- Standard workflow
- Python interface
Inputs & Outputs
- Input formats: ReaxFF parameters, training data
- Output data types: Optimized ReaxFF parameters, hybrid models
Interfaces & Ecosystem
- TensorFlow: Backend
- LAMMPS: MD engine
- Python: Core
Performance Characteristics
- Speed: Standard ReaxFF + NN overhead
- Accuracy: Better than standard ReaxFF
- System size: Any (force field)
- Automation: Semi-automated
Computational Cost
- Fitting: Hours
- MD: Similar to ReaxFF
Limitations & Known Constraints
- TensorFlow: Required
- Complex setup: Multi-step
- Limited documentation: Academic code
- ReaxFF basis: Inherited limitations
Comparison with Other Codes
- vs JAX-ReaxFF: I-ReaxFF adds NN, JAX-ReaxFF is optimization only
- vs standard ReaxFF: I-ReaxFF is differentiable + NN
- vs pure MLIP: I-ReaxFF has ReaxFF physics
- Unique strength: Differentiable ReaxFF with ReaxFF-MPNN hybrid combining reactive FF with neural networks
Application Areas
Reactive MD:
- Combustion
- Catalysis
- Battery electrolytes
- Polymer degradation
Hybrid Potential:
- ReaxFF + ML accuracy
- Reactive + data-driven
- Transfer learning
Best Practices
- Start from existing ReaxFF parameters
- Use gradient-based optimization
- Validate with MD benchmarks
- Compare with standard ReaxFF
Community and Support
- Open source
- GitHub repository
- Academic development
Verification & Sources
Primary sources:
- GitHub: https://github.com/fenggo/I-ReaxFF
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- Specialized strength: Differentiable ReaxFF with ReaxFF-MPNN hybrid combining reactive FF with neural networks