Official Resources
- Source Repository: https://github.com/google-deepmind/deepmind-research/tree/master/density_functional_approximation_dm21
- Paper: Science 374, 1385 (2021)
- License: Open source (Apache-2.0)
Overview
DM21 (DeepMind 21) is a neural network exchange-correlation functional that solves the fractional electron problem in DFT. It enforces piecewise linearity and size-consistency constraints, improving predictions for charge transfer, dissociation, and band gaps.
Scientific domain: Neural network XC functional with fractional electron constraints
Target user community: Researchers needing improved DFT for charge transfer and dissociation
Theoretical Methods
- Neural network XC functional
- Fractional charge constraint (piecewise linearity)
- Fractional spin constraint
- Size-consistency
- PySCF integration
Capabilities (CRITICAL)
- Piecewise linear XC
- Fractional charge/spin constraints
- PySCF integration
- DM21, DM21m, DM21mu variants
- Improved charge transfer and dissociation
Sources: GitHub repository, Science 374, 1385 (2021)
Key Strengths
Constraints:
- Piecewise linearity
- Fractional spin constraint
- Size-consistency
- Delocalization error reduction
Accuracy:
- Better charge transfer
- Correct dissociation limits
- Improved band gaps
- Reduced self-interaction error
Integration:
- PySCF interface
- Standard DFT workflow
- Easy to use
- Multiple variants
Inputs & Outputs
- Input formats: Molecular structures (PySCF)
- Output data types: XC energy, potential, total energy
Interfaces & Ecosystem
- PySCF: DFT engine
- TensorFlow: Backend
- Python: Core
Performance Characteristics
- Speed: Similar to hybrid DFT
- Accuracy: Better than standard hybrids for charge transfer
- System size: Molecular
- Automation: Full
Computational Cost
- SCF: Similar to hybrid DFT
- No training needed: Pretrained
Limitations & Known Constraints
- Molecular only: No periodic systems
- PySCF only: Limited code support
- Gamma point only: For periodic
- Convergence issues: Some systems
Comparison with Other Codes
- vs Skala: DM21 is constraint-based, Skala is message-passing
- vs DeePKS: DM21 is fixed, DeePKS is trainable
- vs standard hybrids: DM21 has fractional constraints
- Unique strength: Neural XC functional with fractional charge/spin constraints solving delocalization error
Application Areas
Improved DFT:
- Charge transfer systems
- Dissociation curves
- Band gap prediction
- Reaction barriers
Benchmarking:
- DFT functional comparison
- Delocalization error studies
- Fractional electron systems
Best Practices
- Use DM21m for better convergence
- Compare with standard functionals
- Validate on target chemistry
- Use PySCF for SCF
Community and Support
- Open source (Apache-2.0)
- DeepMind maintained
- Published in Science
Verification & Sources
Primary sources:
- GitHub: https://github.com/google-deepmind/deepmind-research/tree/master/density_functional_approximation_dm21
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- Specialized strength: Neural XC functional with fractional charge/spin constraints solving delocalization error