Implementation of the DFT-D3 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
- J. Chem. Phys., 2024, 161, 062501. (DOI)
For details on the D3 dispersion model see
For alternative implementations also check out
- simple-dftd3:
- Simple reimplementation of the DFT-D3 dispersion model in Fortran with Python bindings
- torch-dftd:
- PyTorch implementation of DFT-D2 and DFT-D3
- dispax:
- Implementation of the DFT-D3 dispersion model in Jax M.D.
The project can easily be installed with pip.
pip install tad-dftd3
tad-dftd3 is also available from conda.
conda install tad-dftd3
This project is hosted on GitHub at dftd3/tad-dftd3. Obtain the source by cloning the repository with
git clone https://github.com/dftd3/tad-dftd3 cd tad-dftd3
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.yml mamba activate torch
Install this project with pip in the environment
pip install .
The following dependencies are required
PyTorch<2.3.0 is compiled against the NumPy 1.x C-API and requires numpy<2.
Since this cannot be expressed in the package metadata (dependency markers cannot refer to the PyTorch version), pin numpy<2 yourself if you use PyTorch<2.3.0.
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
Some tests compare tad-dftd3's output against reference values computed at
test time by calling the s-dftd3
Fortran implementation through its Python bindings, rather than against
numbers stored in this repository. The dftd3 package is therefore a
required test dependency, not an optional one -- it is pulled in by the
[dev] extra, and there is no offline fallback if it is missing.
All examples can also be found in the examples directory.
The following example shows how to calculate the DFT-D3 dispersion energy for a single structure.
import torch
import tad_dftd3 as d3
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())
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],
]
)
param = {
"a1": torch.tensor(0.49484001),
"s8": torch.tensor(0.78981345),
"a2": torch.tensor(5.73083694),
}
energy = d3.dftd3(numbers, positions, param)
torch.set_printoptions(precision=10)
print(energy)
# tensor([-0.0004075971, -0.0003940886, -0.0003817684, -0.0003949536,
# -0.0003577212, -0.0004110279, -0.0005385976, -0.0001808242,
# -0.0001563670, -0.0001503394, -0.0001577045, -0.0001764488])The next example shows the calculation of dispersion energies for a batch of structures, while retaining access to all intermediates used for calculating the dispersion energy.
import torch
import tad_dftd3 as d3
import tad_mctc as mctc
sample1 = dict(
numbers=mctc.convert.symbol_to_number("Pb H H H H Bi H H H".split()),
positions=torch.tensor(
[
[-0.00000020988889, -4.98043478877778, +0.00000000000000],
[+3.06964045311111, -6.06324400177778, +0.00000000000000],
[-1.53482054188889, -6.06324400177778, -2.65838526500000],
[-1.53482054188889, -6.06324400177778, +2.65838526500000],
[-0.00000020988889, -1.72196703577778, +0.00000000000000],
[-0.00000020988889, +4.77334244722222, +0.00000000000000],
[+1.35700257511111, +6.70626379422222, -2.35039772300000],
[-2.71400388988889, +6.70626379422222, +0.00000000000000],
[+1.35700257511111, +6.70626379422222, +2.35039772300000],
]
),
)
sample2 = dict(
numbers=mctc.convert.symbol_to_number(
"C C C C C C I H H H H H S H C H H H".split(" ")
),
positions=torch.tensor(
[
[-1.42754169820131, -1.50508961850828, -1.93430551124333],
[+1.19860572924150, -1.66299114873979, -2.03189643761298],
[+2.65876001301880, +0.37736955363609, -1.23426391650599],
[+1.50963368042358, +2.57230374419743, -0.34128058818180],
[-1.12092277855371, +2.71045691257517, -0.25246348639234],
[-2.60071517756218, +0.67879949508239, -1.04550707592673],
[-2.86169588073340, +5.99660765711210, +1.08394899986031],
[+2.09930989272956, -3.36144811062374, -2.72237695164263],
[+2.64405246349916, +4.15317840474646, +0.27856972788526],
[+4.69864865613751, +0.26922271535391, -1.30274048619151],
[-4.63786461351839, +0.79856258572808, -0.96906659938432],
[-2.57447518692275, -3.08132039046931, -2.54875517521577],
[-5.88211879210329, 11.88491819358157, +2.31866455902233],
[-8.18022701418703, 10.95619984550779, +1.83940856333092],
[-5.08172874482867, 12.66714386256482, -0.92419491629867],
[-3.18311711399702, 13.44626574330220, -0.86977613647871],
[-5.07177399637298, 10.99164969235585, -2.10739192258756],
[-6.35955320518616, 14.08073002965080, -1.68204314084441],
]
),
)
numbers = mctc.batch.pack(
(
sample1["numbers"],
sample2["numbers"],
)
)
positions = mctc.batch.pack(
(
sample1["positions"],
sample2["positions"],
)
)
ref = d3.reference.Reference()
rcov = mctc.data.COV_D3()[numbers]
rvdw = mctc.data.VDW_PAIRWISE()[numbers.unsqueeze(-1), numbers.unsqueeze(-2)]
r4r2 = d3.data.R4R2()[numbers]
param = {
"a1": torch.tensor(0.49484001),
"s8": torch.tensor(0.78981345),
"a2": torch.tensor(5.73083694),
}
cn = mctc.ncoord.cn_d3(
numbers, positions, counting_function=mctc.ncoord.exp_count, rcov=rcov
)
weights = d3.model.weight_references(numbers, cn, ref, d3.model.gaussian_weight)
c6 = d3.model.atomic_c6(numbers, weights, ref)
energy = d3.disp.dispersion(
numbers,
positions,
param,
c6,
rvdw,
r4r2,
d3.disp.rational_damping,
)
torch.set_printoptions(precision=10)
print(torch.sum(energy, dim=-1))
# tensor([-0.0014092578, -0.0057840119])Since the dispersion energy is differentiable with respect to the atomic positions, the D3 contribution to the gradient (and hence the forces) is obtained from a simple backward pass.
import torch
import tad_dftd3 as d3
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())
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],
],
requires_grad=True,
)
param = {
"a1": torch.tensor(0.49484001),
"s8": torch.tensor(0.78981345),
"a2": torch.tensor(5.73083694),
}
energy = d3.dftd3(numbers, positions, param)
(grad,) = torch.autograd.grad(energy.sum(), positions)
forces = -gradThe full example, including a comparison against numerical gradients, is available in forces.py. Second derivatives, i.e., the D3 contribution to the Hessian, are obtained by applying reverse-mode automatic differentiation twice (see hessian.py).
The current implementation only works for molecular structures. Periodic boundary conditions are not implemented, i.e., no stress tensor or lattice gradient is available.
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).
This is a volunteer open source projects and contributions are always welcome. Please, take a moment to read the contributing guidelines.
Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file 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.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project by you, as defined in the Apache-2.0 license, shall be licensed as above, without any additional terms or conditions.