Torch Autodiff for DFT-D4

Compatibility:
Availability:
Status:

Implementation of the DFT-D4 dispersion model in PyTorch. This module allows to process a single structure or a batch of structures for the calculation of atom-resolved dispersion energies.

If you use this software, please cite the following publication:

• M. Friede, C. Hölzer, S. Ehlert, S. Grimme, J. Chem. Phys., 2024, 161, 062501. DOI: 10.1063/5.0216715

For details on the D4 dispersion model, see:

• E. Caldeweyher, C. Bannwarth and S. Grimme, J. Chem. Phys., 2017, 147, 034112. DOI: 10.1063/1.4993215
• E. Caldeweyher, S. Ehlert, A. Hansen, H. Neugebauer, S. Spicher, C. Bannwarth and S. Grimme, J. Chem. Phys., 2019, 150, 154122. DOI: 10.1063/1.5090222
• E. Caldeweyher, J.-M. Mewes, S. Ehlert and S. Grimme, Phys. Chem. Chem. Phys., 2020, 22, 8499-8512. DOI: 10.1039/D0CP00502A

For alternative implementations, also check out:

• dftd4: Implementation of the DFT-D4 dispersion model in Fortran with Python bindings.
• cpp-d4: Implementation of the DFT-D4 dispersion model in C++.

Installation

pip

tad-dftd4 can easily be installed with pip.

pip install tad-dftd4

conda

tad-dftd4 is also available from conda.

conda install tad-dftd4

From source

This project is hosted on GitHub at dftd4/tad-dftd4. Obtain the source by cloning the repository with

git clone https://github.com/dftd4/tad-dftd4
cd tad-dftd4

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 .

The following dependencies are required

• numpy
• tad-mctc
• tad-multicharge
• torch
• pytest (tests only)

Compatibility

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.

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

Development

For development, additionally install the following tools in your environment.

mamba install black covdefaults mypy pre-commit pylint pytest pytest-cov pytest-xdist tox
pip install pytest-random-order

With pip, add the option -e for installing in development mode, and add [dev] for the development dependencies

pip install -e .[dev]

The pre-commit hooks are initialized by running the following command in the root of the repository.

pre-commit install

For testing all Python environments, simply run tox.

tox

Note that this randomizes the order of tests but skips "large" tests. To modify this behavior, tox has to skip the optional posargs.

tox -- test

Examples

All examples can also be found in the examples directory.

• single.py
• batch.py
• forces.py
• d4s.py

The following example shows how to calculate the DFT-D4 dispersion energy for a single structure.

import torch
import tad_dftd4 as d4
import tad_mctc as mctc

numbers = mctc.convert.symbol_to_number(symbols="C C C C N C S H H H H H".split())

# coordinates in Bohr
positions = torch.tensor(
    [
        [-2.56745685564671, -0.02509985979910, 0.00000000000000],
        [-1.39177582455797, +2.27696188880014, 0.00000000000000],
        [+1.27784995624894, +2.45107479759386, 0.00000000000000],
        [+2.62801937615793, +0.25927727028120, 0.00000000000000],
        [+1.41097033661123, -1.99890996077412, 0.00000000000000],
        [-1.17186102298849, -2.34220576284180, 0.00000000000000],
        [-2.39505990368378, -5.22635838332362, 0.00000000000000],
        [+2.41961980455457, -3.62158019253045, 0.00000000000000],
        [-2.51744374846065, +3.98181713686746, 0.00000000000000],
        [+2.24269048384775, +4.24389473203647, 0.00000000000000],
        [+4.66488984573956, +0.17907568006409, 0.00000000000000],
        [-4.60044244782237, -0.17794734637413, 0.00000000000000],
    ]
)

# total charge of the system
charge = torch.tensor(0.0)

# TPSSh-D4-ATM parameters
param = {
    "s6": positions.new_tensor(1.0),
    "s8": positions.new_tensor(1.85897750),
    "s9": positions.new_tensor(1.0),
    "a1": positions.new_tensor(0.44286966),
    "a2": positions.new_tensor(4.60230534),
}

# parameters can also be obtained using the functional name:
# param = d4.get_params("tpssh")

energy = d4.dftd4(numbers, positions, charge, param)
torch.set_printoptions(precision=10)
print(energy)
# tensor([-0.0020841344, -0.0018971195, -0.0018107513, -0.0018305695,
#         -0.0021737693, -0.0019484236, -0.0022788253, -0.0004080658,
#         -0.0004261866, -0.0004199839, -0.0004280768, -0.0005108935])

The next example shows the calculation of dispersion energies for a batch of structures.

import torch
import tad_dftd4 as d4
import tad_mctc as mctc

# S22 system 4: formamide dimer
numbers = mctc.batch.pack((
    mctc.convert.symbol_to_number("C C N N H H H H H H O O".split()),
    mctc.convert.symbol_to_number("C O N H H H".split()),
))

# coordinates in Bohr
positions = mctc.batch.pack((
    torch.tensor([
        [-3.81469488143921, +0.09993441402912, 0.00000000000000],
        [+3.81469488143921, -0.09993441402912, 0.00000000000000],
        [-2.66030049324036, -2.15898251533508, 0.00000000000000],
        [+2.66030049324036, +2.15898251533508, 0.00000000000000],
        [-0.73178529739380, -2.28237795829773, 0.00000000000000],
        [-5.89039325714111, -0.02589114569128, 0.00000000000000],
        [-3.71254944801331, -3.73605775833130, 0.00000000000000],
        [+3.71254944801331, +3.73605775833130, 0.00000000000000],
        [+0.73178529739380, +2.28237795829773, 0.00000000000000],
        [+5.89039325714111, +0.02589114569128, 0.00000000000000],
        [-2.74426102638245, +2.16115570068359, 0.00000000000000],
        [+2.74426102638245, -2.16115570068359, 0.00000000000000],
    ]),
    torch.tensor([
        [-0.55569743203406, +1.09030425468557, 0.00000000000000],
        [+0.51473634678469, +3.15152550263611, 0.00000000000000],
        [+0.59869690244446, -1.16861263789477, 0.00000000000000],
        [-0.45355203669134, -2.74568780438064, 0.00000000000000],
        [+2.52721209544999, -1.29200800956867, 0.00000000000000],
        [-2.63139587595376, +0.96447869452240, 0.00000000000000],
    ]),
))

# total charge of both system
charge = torch.tensor([0.0, 0.0])

# TPSSh-D4-ATM parameters
param = {
    "s6": positions.new_tensor(1.0),
    "s8": positions.new_tensor(1.85897750),
    "s9": positions.new_tensor(1.0),
    "a1": positions.new_tensor(0.44286966),
    "a2": positions.new_tensor(4.60230534),
}

# calculate dispersion energy in Hartree
energy = torch.sum(d4.dftd4(numbers, positions, charge, param), -1)
torch.set_printoptions(precision=10)
print(energy)
# tensor([-0.0088341432, -0.0027013607])
print(energy[0] - 2*energy[1])
# tensor(-0.0034314217)

The last example shows how to use the D4SModel.

import torch
import tad_dftd4 as d4
import tad_mctc as mctc

numbers = mctc.convert.symbol_to_number(symbols="C C C C N C S H H H H H".split())

# coordinates in Bohr
positions = torch.tensor(
    [
        [-2.56745685564671, -0.02509985979910, 0.00000000000000],
        [-1.39177582455797, +2.27696188880014, 0.00000000000000],
        [+1.27784995624894, +2.45107479759386, 0.00000000000000],
        [+2.62801937615793, +0.25927727028120, 0.00000000000000],
        [+1.41097033661123, -1.99890996077412, 0.00000000000000],
        [-1.17186102298849, -2.34220576284180, 0.00000000000000],
        [-2.39505990368378, -5.22635838332362, 0.00000000000000],
        [+2.41961980455457, -3.62158019253045, 0.00000000000000],
        [-2.51744374846065, +3.98181713686746, 0.00000000000000],
        [+2.24269048384775, +4.24389473203647, 0.00000000000000],
        [+4.66488984573956, +0.17907568006409, 0.00000000000000],
        [-4.60044244782237, -0.17794734637413, 0.00000000000000],
    ]
)

# total charge of the system
charge = torch.tensor(0.0)

# Create the D4S model
model = d4.model.D4SModel(numbers, **dd)

param = d4.get_params("tpssh")
energy = d4.dftd4(numbers, positions, charge, param, model=model)

torch.set_printoptions(precision=10)
print(energy)
# tensor([-0.0020843975, -0.0019013016, -0.0018165035, -0.0018363572,
#         -0.0021877293, -0.0019495023, -0.0022923108, -0.0004326892,
#         -0.0004439871, -0.0004362087, -0.0004454589, -0.0005344027])

Limitations

The current implementation only works molecular structures. Periodic boundary conditions are not implemented.

The code is fully vectorized for maximum efficiency. Therefore, all quantities are stored as full tensors, which makes calculations rather memory intensive. Especially, the ATM term can become limiting as it requires a 3D tensor of dimension (n_atoms, n_atoms, n_atoms).

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.