Torch Autodiff Utility

Compatibility:
Availability:
Status:

This library is a collection of utility functions that are used in PyTorch (re-)implementations of projects from the Grimme group. In particular, the tad-mctc library provides:

• autograd functions (Jacobian, Hessian)

• atomic data (radii, EN, example molecules, ...)

• batch utility (packing, masks, ...)

• conversion functions (numpy, atomic symbols/numbers, ...)

• coordination numbers (DFT-D3, DFT-D4, EEQ)

• io (reading/writing coordinate files)

• molecular properties (bond lengths/orders/angles, moment of inertia, ...)

• safeops (autograd-safe implementations of common functions)

• typing (base class for tensor-like behavior of arbitrary classes)

• units

The name is inspired by the Fortran pendant "modular computation tool chain library" (mctc-lib).

Citation

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

Installation

pip

tad-mctc can easily be installed with pip.

pip install tad-mctc

conda

tad-mctc is also available from conda.

conda install tad-mctc

From source

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

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

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
• opt_einsum
• psutil
• pytest (tests only)
• scipy (tests only)
• torch

Compatibility

:white_check_mark: tested in CI    :heavy_check_mark: supported, but not tested in CI (should still work)    :x: not supported

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.

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.

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

The following example shows how to calculate the coordination number used in the EEQ model for a single structure.

import torch
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],
    ]
)

# calculate EEQ coordination number
cn = mctc.ncoord.cn_eeq(numbers, positions)
torch.set_printoptions(precision=10)
print(cn)
# tensor([3.0519218445, 3.0177774429, 3.0132560730, 3.0197706223,
#         3.0779352188, 3.0095663071, 1.0991339684, 0.9968624115,
#         0.9943327904, 0.9947233200, 0.9945874214, 0.9945726395])

The next example shows the calculation of the coordination number used in DFT-D4 for a batch of structures.

import torch
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],
    ]),
))

# calculate coordination number
cn = mctc.ncoord.cn_d4(numbers, positions)
torch.set_printoptions(precision=10)
print(cn)
# tensor([[2.6886456013, 2.6886456013, 2.6314170361, 2.6314167976,
#          0.8594539165, 0.9231414795, 0.8605306745, 0.8605306745,
#          0.8594539165, 0.9231414795, 0.8568341732, 0.8568341732],
#         [2.6886456013, 0.8568335176, 2.6314167976, 0.8605306745,
#          0.8594532013, 0.9231414795, 0.0000000000, 0.0000000000,
#          0.0000000000, 0.0000000000, 0.0000000000, 0.0000000000]])

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.