Official Resources
- Source Repository: https://github.com/deepmodeling/dpgen
- Documentation: https://docs.deepmodeling.com/projects/dpgen/
- License: Open source (LGPL-3.0)
Overview
DP-GEN (Deep Potential GENerator) is an active learning workflow for generating deep learning interatomic potentials. It automates the cycle of training, exploration, and labeling to build high-quality DeePMD potentials with minimal human intervention.
Scientific domain: Active learning workflow for DeePMD potential generation
Target user community: Researchers building DeePMD potentials via active learning
Theoretical Methods
- Active learning loop (train-explore-label)
- Model disagreement-based selection
- Multiple DeePMD model training
- DFT labeling of selected structures
- Convergence monitoring
Capabilities (CRITICAL)
- Automated active learning
- Multiple DFT code support (VASP, QE, ABACUS, Gaussian)
- Model disagreement selection
- Convergence monitoring
- HPC job management
Sources: GitHub repository, Nat. Commun. 11, 5713 (2020)
Key Strengths
Automated:
- Train-explore-label loop
- Model disagreement selection
- Convergence monitoring
- Minimal human intervention
Multi-Code:
- VASP, QE, ABACUS for DFT
- Gaussian, CP2K for molecules
- DeePMD-kit for training
- LAMMPS for exploration
Production:
- Well-tested at scale
- HPC job management
- Slurm/PBS support
- Checkpoint/restart
Inputs & Outputs
- Input formats: Initial structures, DFT parameters
- Output data types: Trained DeePMD models, training data
Interfaces & Ecosystem
- DeePMD-kit: Training
- LAMMPS: Exploration
- VASP/QE/ABACUS: DFT labeling
- Python: Core
Performance Characteristics
- Speed: Limited by DFT calls
- Accuracy: Near-DFT (converged)
- System size: Any
- Automation: Full
Computational Cost
- DFT calls: Expensive (but minimized)
- Training: Hours per iteration
- Total: Days to weeks
Limitations & Known Constraints
- DFT cost: Still need DFT calculations
- DeePMD only: Only for DeePMD models
- Complex setup: Multi-step workflow
- LGPL license: Copyleft
Comparison with Other Codes
- vs FLARE: DP-GEN is batch, FLARE is on-the-fly
- vs Metatrain: DP-GEN is active learning, Metatrain is training
- vs manual fitting: DP-GEN is automated
- Unique strength: Automated active learning workflow for DeePMD with multi-DFT-code support
Application Areas
Potential Development:
- Automated MLIP construction
- High-throughput potential generation
- Multi-component systems
- Phase diagram exploration
Production:
- Converged DeePMD potentials
- Published potential generation
- Multi-fidelity training
Best Practices
- Start with small training set
- Use model disagreement for selection
- Monitor convergence metrics
- Use HPC for DFT calculations
Community and Support
- Open source (LGPL-3.0)
- DeepModeling community
- Comprehensive documentation
- Published in Nature Communications
Verification & Sources
Primary sources:
- GitHub: https://github.com/deepmodeling/dpgen
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- Specialized strength: Automated active learning workflow for DeePMD with multi-DFT-code support