🚀 About DeePTB

DeePTB is an innovative Python package that uses deep learning to accelerate ab initio electronic structure simulations. It offers versatile, accurate, and efficient simulations for a wide range of materials and phenomena. Trained on small systems, DeePTB can predict electronic structures of large systems, handle structural perturbations, and integrate with molecular dynamics for finite temperature simulations, providing comprehensive insights into atomic and electronic behavior. Furthermore, the Hamiltonians generated by DeePTB can be seamlessly integrated with DPNEGF, our official companion package for high-performance quantum transport simulations based on the Non-Equilibrium Green's Function (NEGF) formalism.

• Key Features DeePTB contains two main components:

  1. DeePTB-SK: deep learning based local environment dependent Slater-Koster TB.

     • Customizable Slater-Koster parameterization with neural network corrections for .
     • Flexible basis and exchange-correlation functional choices.
     • Handle systems with strong spin-orbit coupling (SOC) effects.

  2. DeePTB-E3: E3-equivariant neural networks for representing quantum operators.

     • Construct DFT Hamiltonians/density and overlap matrices under full LCAO basis.
     • Utilize (Strictly) Localized Equivariant Message-passing ((S)LEM) model for high data-efficiency and accuracy.
     • Employs SO(2) convolution for efficient handling of higher-order orbitals in LCAO basis.

For more details, see our papers:

• DeePTB-SK: Nat Commun 15, 6772 (2024)
• DeePTB-E3: ICLR 2025 Spotlight

For the underlying ideas in DeePTB-SK, see the author's research note (English / 中文). For the localized equivariant operators used in DeePTB-E3, see the SLEM note (English / 中文).

Peer-reviewed applications and observable-level validation

A learned Hamiltonian is an input to a physical calculation, not a guarantee that every downstream observable is accurate. These published studies give distinct examples of how electronic-structure models are connected to and validated against physical properties:

Application	What was computed and checked	Publication
Finite-temperature electronic structure	Environment-dependent DeePTB-SK Hamiltonians and tight-binding propagation for DOS and optical properties in a one-million-atom GaP structure; not a million-atom self-consistent DFT calculation	Nature Communications 15, 6772 (2024)
Localized equivariant quantum operators	SLEM prediction of atomic-orbital Hamiltonian, overlap and density matrices in a rotation-equivariant, strictly local representation	ICLR 2025 Spotlight
Nanoelectronic quantum transport	Learned Hamiltonians coupled with NEGF for over 10,000 break-junction configurations, and carbon-nanotube FET calculations benchmarked against experiments; lead contacts and device assumptions remain part of the transport model	npj Computational Materials 11, 375 (2025)
Two-dimensional material transport	DeePTB-derived Hamiltonians and NEGF transmission tests for graphene, h-BN and MoS2, with comparison to DFT-NEGF	IEEE MoSICom 2025
Finite-temperature optical spectra	DeePMD sampling with HSE-trained DeePTB electronic models for phonon-assisted spectra in Si, GaAs and GaP, including supercells up to 4,096 atoms	npj Computational Materials 12, 268 (2026)

These are different model, material and validation settings, not interchangeable performance claims. For device-level transport, see the separate DPNEGF solver; the DeePTB model does not by itself include electrode self-energies, gate electrostatics or full device transport.

📚 Documentation

• Online documentation

  For a comprehensive guide and usage tutorials, visit Documentation website.

• Contributing

  We welcome contributions to DeePTB. Please refer to our contributing guidelines for details.

🛠️ Installation

Installing DeePTB is straightforward with UV, a fast Python package manager.

• Requirements

  • Git
  • Python 3.10 to 3.13
  • UV, used by install.sh as the fast installer frontend
  • For GPU installs: an NVIDIA driver compatible with the selected CUDA runtime

• From Source (Recommended)

  1. Install UV (if not already installed):

     # On macOS and Linux
     curl -LsSf https://astral.sh/uv/install.sh | sh
     
     # Or using pip
     pip install uv
     
     # On Windows (PowerShell)
     powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

  2. Clone DeePTB:

     git clone https://github.com/deepmodeling/DeePTB.git
     cd DeePTB

  3. Install DeePTB with tested dependencies:

     Automatic CPU/GPU selection:

     ./install.sh

     CPU-only install:

     ./install.sh cpu

     GPU install:

     nvidia-smi  # check the driver-reported CUDA version
     ./install.sh gpu    # auto-detect CUDA backend
     
     # Or force a tested CUDA wheel path:
     ./install.sh cu128  # RTX 50 / CUDA 12.8 path tested with torch 2.10.0
     ./install.sh cu130  # requires a driver new enough for CUDA 13.0 runtime

     This single command will:

     • Automatically create a virtual environment (.venv)
     • Install a tested PyTorch / PyG / torch-scatter binary-wheel combination
     • Refuse unsupported Python or CUDA/backend combinations instead of falling back to source builds
     • Install all runtime and test dependencies

  4. Activate the standalone environment:

     install.sh installs DeePTB into .venv under the current DeePTB repository. Activate it before running dptb:

     source .venv/bin/activate  # On Unix/macOS
     .venv\Scripts\activate     # On Windows
     dptb --help

  5. Validate the installation: DeePTB is under active development, so new installations should run the test suite once before production use.

     python -m pytest ./dptb/tests/

     For a faster local check while iterating:

     python -m pytest ./dptb/tests/ -m "not slow"

  6. Install optional dependencies (if needed):

     # For 3D Fermi surface plotting
     ./install.sh auto --extra 3Dfermi
     
     # For TBtrans initialization
     ./install.sh auto --extra tbtrans_init
     
     # For pybinding support
     ./install.sh auto --extra pybinding

• Developer Install

  pyproject.toml declares the broader source-compatible range (Python >=3.10,<3.14, torch >=2.5.1,<=2.12.1). Developers who already manage their own Torch/PyG environment can still use:

  uv sync

  This path is for developers who intentionally manage their own environment. For a tested standalone DeePTB environment, prefer install.sh.

• Library Installation (Existing Environment)

  [!WARNING] This path is for downstream projects or existing environments that import DeePTB as a library. The environment must provide a compatible PyTorch and torch-scatter binary wheel.

  DeePTB keeps a broad Torch compatibility range so downstream projects can choose their own CPU/GPU backend. In that case, install PyTorch first, then install a torch-scatter binary wheel matching the current Torch version and backend. If you are working from a DeePTB source checkout, the helper can inspect the current environment and print the matching PyG wheel command:

  python - <<'PY'
  import torch
  print("torch:", torch.__version__)
  print("cuda runtime:", torch.version.cuda)
  print("cuda available:", torch.cuda.is_available())
  PY
  
  python docs/auto_install_torch_scatter.py --dry-run
  python docs/auto_install_torch_scatter.py

  Then install DeePTB from the current source checkout:

  pip install .

  Use pip install -e . instead if you want an editable developer install. Published package installs, such as pip install dptb, were not part of this compatibility test pass; prefer a source checkout until that path is tested.

  For standalone DeePTB use on a new machine, use the From Source installer above instead; it creates .venv and selects a tested Torch / PyG / torch-scatter combination.

  Example manual torch-scatter commands:

  # CPU, torch 2.12.1
  pip install --only-binary torch-scatter \
    torch-scatter==2.1.2 \
    -f https://data.pyg.org/whl/torch-2.12.1+cpu.html
  
  # CUDA 12.8 / torch 2.10.0
  pip install --only-binary torch-scatter \
    torch-scatter==2.1.2+pt210cu128 \
    -f https://data.pyg.org/whl/torch-2.10.0+cu128.html

• Julia Backend (Optional - for High-Performance Pardiso Solver)

  [!NOTE] Platform Support: Pardiso backend currently supports Linux only.

  • macOS: Not supported (Intel MKL limitations)
  • Windows: Use WSL2 (Windows Subsystem for Linux)

  If you want to use the Pardiso backend for accelerated band structure calculations:

  Automated Installation (Recommended):

  ./install_julia.sh

  Manual Installation:

  1. Install Julia:

     # Linux (macOS can install Julia, but the Pardiso backend is not supported)
     curl -fsSL https://install.julialang.org | sh

  2. Install required packages:

     julia install_julia_packages.jl

  Verify Installation:

  julia -e 'using Pardiso; println("Pardiso available: ", Pardiso.MKL_PARDISO_LOADED[])'

  Usage:

  dptb pdso band.json -i model.pth -stu structure.vasp -o ./output

  For more details, see:

  • Pardiso Backend README
  • Example Tutorial

Test code

To ensure the code is correctly installed, please run the unit tests first:

python -m pytest ./dptb/tests/

Be careful if not all tests pass!

🤝 How to Cite

The following references are required to be cited when using DeePTB. Specifically:

• For DeePTB-SK:

  Q. Gu, Z. Zhouyin, S. K. Pandey, P. Zhang, L. Zhang, and W. E, Deep Learning Tight-Binding Approach for Large-Scale Electronic Simulations at Finite Temperatures with Ab Initio Accuracy, Nat Commun 15, 6772 (2024).

• For DeePTB-E3:

  Z. Zhouyin, Z. Gan, S. K. Pandey, L. Zhang, and Q. Gu, Learning Local Equivariant Representations for Quantum Operators, In The 13th International Conference on Learning Representations (ICLR) 2025.