DeePTB

Official Resources

• Homepage: https://github.com/deepmodeling/DeePTB
• Source Repository: https://github.com/deepmodeling/DeePTB
• License: GNU Lesser General Public License v3.0

Overview

DeePTB (Deep Learning for Tight-Binding) is a powerful software package that creates highly accurate tight-binding (TB) Hamiltonians using deep learning techniques. While traditional TB models rely on analytical Slater-Koster rules or manual parameterization, DeePTB trains deep neural networks to predict the Hamiltonian matrix elements directly from the atomic structure. This allows it to achieve ab initio accuracy (matching DFT bands and energies) while retaining the low computational cost of tight-binding, enabling simulation of large-scale systems with quantum accuracy.

Scientific domain: Machine Learning, Tight-Binding, Electronic Structure Target user community: Materials scientists, Physicists studying large/disordered systems

Theoretical Methods

• Deep Neural Networks (DNN): Mapping structure to Hamiltonian.
• Refined Slater-Koster (DeePTB-SK): Using NN to correct/modulate SK parameters.
• Equivariant Neural Networks (DeePTB-E3): Directly predicting Hamiltonian blocks using E3-equivariant representations (no scaling laws assumed).
• Spectral Properties: Calculation of bands, DOS, and Berry curvature.

Capabilities (CRITICAL)

• Hamiltonian Prediction: Fast generation of $H(k)$ for any configuration.
• Band Structure: Highly accurate reproduction of DFT bands.
• Electronic Properties: Total energy, forces, atomic charges.
• Transferability: Models trained on small supercells transfer to large systems/defects.
• Efficiency: Orders of magnitude faster than DFT.

Key Strengths

Ab Initio Accuracy:

• Unlike standard TB, DeePTB captures subtle environmental effects and hybridization shifts.
• Can describe complex bonding (halides, oxides) difficult for simple SK models.

Symmetry Preserving:

• DeePTB-E3 guarantees that the predicted Hamiltonian transforms correctly under rotation and translation, ensuring physical validity without data augmentation.

Efficient Workflow:

• Part of the DeepModeling ecosystem (DeePMD), allowing seamless integration with MD workflows.

Inputs & Outputs

• Inputs:
  • Atomic config (POSCAR/XYZ).
  • Training Data: DFT Hamiltonians/Eigenvalues (from VASP/ABACUS/OpenMX).
• Outputs:
  • Predicted TB Hamiltonian (sparse format).
  • Band structures.
  • Electronic Free Energy and Forces.

Interfaces & Ecosystem

• DeepModeling: Compatible with DeePMD-kit workflows.
• DFT Codes: Interfaces to read Hamiltonians from VASP, ABACUS, and OpenMX.
• Wannier90: Can interface with Wannierized Hamiltonians.

Advanced Features

• Force Training: Can train on forces to enable stable Molecular Dynamics.
• Self-Consistency: Can be coupled with a self-consistent density loop (though primarily used in non-SCF mode for bands).

Performance Characteristics

• Speed: Hamiltonian generation is fast (inference time); Diagonalization cost dominates for large systems.
• Scalability: Linear scaling generation; Eigensolver depends on method.
• Training: Requires GPU for efficient training of E3 networks.

Computational Cost

• Inference: Very Low (milliseconds for small cells).
• Training: Moderate (requires sufficient DFT data).

Limitations & Known Constraints

• Data Dependence: The model is only as good as the DFT data it saw; extrapolation to unknown phases can be physical (due to E3) but inaccurate.
• Hamiltonian Size: Requires defining a basis set size (e.g., orbitals per atom) consistent with reference.

Comparison with Other Codes

• vs DFTB+: DFTB+ uses a fixed physical model (2-center integrals); DeePTB uses a flexible Neural Network model (multi-center effects). DeePTB is more accurate but requires training.
• vs Wannier90: Wannier90 is a post-processing tool for one structure; DeePTB learns the function $H(R)$ to predict new structures.
• vs SchNetPack: SchNetPack predicts scalars (Energy) or vectors (Forces); DeePTB predicts Matrices (Hamiltonian).
• Unique strength: The State-of-the-Art implementation of Equivariant Neural Networks for electronic Hamiltonians.

Application Areas

• Disordered Systems: Alloys, amorphous semiconductors.
• Defects: Electronic levels of vacancies/dopants in large supercells.
• Twisted Bilayers: Moiré physics requiring huge unit cells (thousands of atoms).
• Finite Temperature: Band structure renormalization due to thermal vibrations.

Best Practices

• Data Diversity: Include distorted structures in training to learn bond-length dependence.
• Check Symmetry: Use DeePTB-E3 for crystals to ensure band degeneracies are respected.
• Basis Consistency: Ensure all DFT training data uses the exact same orbital projection/basis definition.

Community and Support

• GitHub: Active issue tracking.
• DeepModeling Community: Large user base in China and global.

Verification & Sources

Primary sources:

1. Repository: https://github.com/deepmodeling/DeePTB
2. Publication: "DeePTB: A deep learning package for tight-binding hamiltonians", Phys. Rev. B (etc.).

Verification status: ✅ VERIFIED

• Source code: OPEN (LGPLv3)
• Maturity: Research grade, rapidly adopting.