Installation

pip

dxtb can easily be installed with pip.

pip install dxtb[libcint]

Installing the libcint interface is recommended, as it is faster than the pure PyTorch implementation on the CPU. The pure PyTorch integral driver provides the overlap, dipole and quadrupole integrals (and thereby GFN2-xTB) with derivatives of any order, on CPU and CUDA. However, the interface is currently only available on Linux.

conda

dxtb is also available on conda from the conda-forge channel.

mamba install dxtb

The optional libcint interface (not on conda) can be installed via pip install tad-libcint.

For Windows, dxtb is not available via conda, because PyTorch itself is not registered in the conda-forge channel.

From source

This project is hosted on GitHub at grimme-lab/dxtb. Obtain the source by cloning the repository with

git clone https://github.com/grimme-lab/dxtb
cd dxtb

We recommend using a conda environment to install the package. You can setup the environment manager using a mambaforge installer. Install the required dependencies from the conda-forge channel.

mamba env create -n torch -f environment.yaml
mamba activate torch

Install this project with pip in the environment

pip install .

Without pip

If you want to install the package without pip, start by cloning the repository.

DEST=/opt/software
git clone https://github.com/grimme-lab/dxtb $DEST/dxtb

Next, add <path to dxtb>/dxtb/src to your $PYTHONPATH environment variable. For the command line interface, add <path to dxtb>/dxtb/bin to your $PATH environment variable.

export PYTHONPATH=$PYTHONPATH:$DEST/dxtb/src
export PATH=$PATH:$DEST/dxtb/bin

Dependencies

The following dependencies are required

• numpy

• opt_einsum

• psutil

• scipy

• tad-mctc

• tad-multicharge

• tad-dftd3

• tad-dftd4

• tad-libcint

• torch

For tests, we also require

• pytest

• pyscf

• tox