DeePTB

Overview

DeePTB provides a unified Python package for building electronic-structure models with deep learning, rather than a single pretrained predictor. Its documented scope includes tight-binding models, Kohn–Sham Hamiltonians and density matrices. In a materials-modelling workflow, it supplies learned electronic representations for electronic-structure calculations and downstream analysis. The README describes applying models trained on small systems to larger structures, handling structural perturbations and coupling with molecular dynamics for finite-temperature simulations.

The package has two principal modelling components. DeePTB-SK uses local atomic environments to modify Slater–Koster tight-binding parameterizations through neural-network corrections, with configurable basis and exchange-correlation functional choices and support for spin–orbit coupling. DeePTB-E3 uses E3-equivariant networks to represent quantum operators, including DFT Hamiltonian, density and overlap matrices in a full linear combination of atomic orbitals (LCAO) basis. Its documented architecture includes localized equivariant message passing and SO(2) convolution for higher-order orbitals.

The supplied band-calculation example takes a model checkpoint, a structure.vasp file and a band.json configuration, and writes results to a designated output directory. Generated Hamiltonians can also feed DPNEGF, a separate companion package for NEGF quantum transport. The excerpts do not specify complete training-data schemas or establish accuracy for a particular material. Operational constraints include compatible PyTorch and torch-scatter binaries; the optional Julia/Pardiso backend is Linux-only, with WSL2 suggested for Windows. The README recommends installation tests before production use and states that published-package installation was outside its described compatibility test pass.

Key Features

• DeePTB-SK provides local-environment-dependent Slater–Koster tight binding with neural-network corrections and configurable basis and exchange-correlation functional choices.
• DeePTB-SK supports systems with strong spin–orbit coupling effects.
• DeePTB-E3 represents DFT Hamiltonian, density and overlap matrices in a full LCAO basis using E3-equivariant neural networks.
• DeePTB-E3 incorporates localized equivariant message passing and SO(2) convolution for higher-order LCAO orbitals.
• Generated Hamiltonians can be integrated with the separate DPNEGF companion package for NEGF quantum-transport calculations.
• An optional Julia/Pardiso backend supports band-structure calculations using a configuration, model checkpoint and atomic structure.

Use Cases

• Suggested evaluation: compare DeePTB-SK electronic-structure predictions against reference calculations for representative perturbed structures, including spin–orbit coupling where relevant.
• Suggested evaluation: assess DeePTB-E3 Hamiltonian, density-matrix or overlap-matrix representations for a material with an appropriate LCAO reference basis.
• Suggested evaluation: test transfer from small-system training examples to larger structures or molecular-dynamics configurations relevant to finite-temperature studies.
• Suggested evaluation: pass generated Hamiltonians to DPNEGF and assess their suitability for a selected quantum-transport problem.

How to Use

1. Read the official documentation and choose DeePTB-SK or DeePTB-E3 according to the required tight-binding or quantum-operator representation. Confirm the relevant basis choices and tutorial inputs before preparing data.
2. Consult the installation section in the repository. Check the documented Python, Git and UV requirements, and GPU-driver requirements if applicable. Prefer its recommended source-installation path for standalone use.
3. Follow the documented environment setup and activation instructions. Run the supplied unit-test procedure before production use; investigate failures rather than assuming the installation is usable.
4. Use the documentation to prepare the chosen workflow. For the illustrated Pardiso band calculation, identify the model checkpoint, band.json configuration, structure.vasp input and output directory; check the backend's Linux restriction.
5. Evaluate results against reference electronic-structure calculations for your target material. If transport is required, consult DPNEGF separately. Follow the README's component-specific citation guidance when reporting work.