dxtb 0.4.0

Fully Differentiable Approach to Extended Tight Binding

Fully Differentiable Extended Tight-Binding

- Combining semi-empirical quantum chemistry with machine learning in PyTorch -

The xTB methods (GFNn-xTB) are a series of semi-empirical quantum chemical methods that provide a good balance between accuracy and computational cost.

With dxtb, we provide a re-implementation of the xTB methods in PyTorch, which allows for automatic differentiation and seamless integration into machine learning frameworks.

NOTE: If you encounter any bugs or have questions on how to use dxtb, feel free to open an issue.

Installation

pip

dxtb can easily be installed with pip.

pip install dxtb[libcint]

Installing the libcint interface is highly recommended, as it is significantly faster than the pure PyTorch implementation and provides access to higher-order multipole integrals and their derivatives (required for GFN2-xTB). However, the interface is currently only available on Linux.

conda

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

mamba install dxtb

Don't forget to install the libcint interface (not on conda) via pip install tad-libcint. The libcint interface is required for GFN2-xTB.

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

Other

For more options, see the installation guide in the documentation.

Example

The following example demonstrates how to compute the energy and forces using GFN1-xTB.

import torch
import dxtb

dd = {"dtype": torch.double, "device": torch.device("cpu")}

# LiH
numbers = torch.tensor([3, 1], device=dd["device"])
positions = torch.tensor([[0.0, 0.0, 0.0], [0.0, 0.0, 1.5]], **dd)

# instantiate a calculator
calc = dxtb.calculators.GFN1Calculator(numbers, **dd)

# compute the energy
pos = positions.clone().requires_grad_(True)
energy = calc.get_energy(pos)

# obtain gradient (dE/dR) via autograd
(g,) = torch.autograd.grad(energy, pos)

# Alternatively, forces can directly be requested from the calculator.
# (Don't forget to manually reset the calculator when the inputs are identical.)
calc.reset()
pos = positions.clone().requires_grad_(True)
forces = calc.get_forces(pos)

assert torch.equal(forces, -g)

All quantities are in atomic units.

For more examples and details, check out the documentation.

Compatibility

PyTorch \ Python	3.8	3.9	3.10	3.11	3.12	3.13	3.14
1.11.0	:white_check_mark:	:white_check_mark:	:x:	:x:	:x:	:x:	:x:
1.12.1	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:	:x:	:x:
1.13.1	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:	:x:
2.0.1	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:	:x:
2.1.2	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:	:x:
2.2.2	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:
2.3.1	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:
2.4.1	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:
2.5.1	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:	:x:
2.6.0	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:
2.7.1	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:
2.8.0	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:x:
2.9.1	:x:	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:
2.10.0	:x:	:x:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:	:white_check_mark:

Note that only the latest bug fix version is listed, but all preceding bug fix minor versions are supported. For example, although only version 2.2.2 is listed, version 2.2.0 and 2.2.1 are also supported.

Restriction for macOS and Windows:

On macOS and Windows, PyTorch<2.0.0 does only support Python<3.11.

The libcint interface is not available for macOS and Windows. Correspondingly, the integral evaluation can be considerably slower. Moreover, higher-order multipole integrals (dipole, quadrupole, ...) are not implemented. While macOS support may be considered in the future, native Windows support is not possible, because the underlying libcint library does not work under Windows.

Citation

If you use dxtb in your research, please cite the following paper:

• M. Friede, C. Hölzer, S. Ehlert, S. Grimme, dxtb -- An Efficient and Fully Differentiable Framework for Extended Tight-Binding, J. Chem. Phys., 2024, 161, 062501. (DOI)

The Supporting Information can be found here.

For details on the xTB methods, see

• C. Bannwarth, E. Caldeweyher, S. Ehlert, A. Hansen, P. Pracht, J. Seibert, S. Spicher, S. Grimme, WIREs Comput. Mol. Sci., 2020, 11, e01493. (DOI)
• C. Bannwarth, S. Ehlert, S. Grimme, J. Chem. Theory Comput., 2019, 15, 1652-1671. (DOI)
• S. Grimme, C. Bannwarth, P. Shushkov, J. Chem. Theory Comput., 2017, 13, 1989-2009. (DOI)

Contributing

This is a volunteer open source projects and contributions are always welcome. Please, take a moment to read the contributing guidelines.

License

This project is licensed under the Apache License, Version 2.0 (the "License"); you may not use this project's files except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.