Symmetry Functions

This library calculates symmetry functions and their derivatives for machine learning potentials. For a short introduction see:

J. Behler, J. Chem. Phys. 134, 074106 (2011) (https://doi.org/10.1063/1.3553717)

The implementation is based on the concept of symmetry groups, as introduced in

A. Singraber, J, Behler, C. Dellago, J. Chem. Theory Comput. 15, 1827 (2019) (https://pubs.acs.org/doi/full/10.1021/acs.jctc.8b00770)

Supported Symmetry Functions

• Element Embracing Symmetry Functions.

  M. Eckhoff, M. Reiher, J. Chem. Theory Comput. 19, 3509 (2023) (https://pubs.acs.org/doi/full/10.1021/acs.jctc.3c00279)

Requirements

Main Library

• Git.
• OpenMP.
• cmake 3.22 (or later).
• GCC/G++ 11.4, C++ 17 standard (other compilers may work but are not tested; there are/were known compatibility problems between Torch and GCC < 9, GCC >= 12. You can try setting -DWARNINGS_AS_ERRORS=OFF to avoid compiler errors caused by warnings from this issue).

Python Bindings

• Pybind11 (installed automatically if not available).

Unit Tests

• GTest (installed automatically if not available).
• Boost/filesystem (1.74.0 or later).

Torch Bindings (available in Python)

• Torch/PyTorch (IMPORTANT: Check CXX11 ABI (see CMake flags below); see https://download.pytorch.org/whl/torch/ for different wheels).
• MKL (see scripts/download_mkl.sh for installation tips).

Installation

Method 1: Manual Build and Install

git submodule init
git submodule update
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../inst -DCMAKE_PREFIX_PATH="$(python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)')" ..
make -j10
make install

After calling make install, your installation directory will include the header files (include), and a directory lib containing the shared library for the Python bindings (SymFuncPy.cpython-<version-specification>.so), the static library with the main implementation (libSymmetryFunctions.a), and the Torch bindings libSymFuncTorch.so. Export the path to the lib directory in your PYTHON_PATH to access the Python bindings.

Other optional CMake flags:

• -DSYM_PYTHON_BINDINGS=ON/OFF, default ON.
• -DSYM_ENABLE_TESTS=ON/OFF, default ON.
• -DSYMMETRY_FUNCTIONS_MARCH=<march-flag>, default native.
• -DPYTHON_EXECUTABLE=<path to Python interpreter>.
• -DSYM_TORCH_BINDINGS=ON/OFF, default ON.
• -D_GLIBCXX_USE_CXX11_ABI=0 to use pre CXX11 ABI (REQUIRED by some torch versions.).
• -DWARNINGS_AS_ERRORS=ON/OFF, default ON. Compile with the -Werror.

Required CMake flags for Torch bindings:

• Add -DCMAKE_PREFIX_PATH="$(python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)')" to the cmake call.

Method 2: Install Python Bindings via pip

git submodule init
git submodule update
python3 -m pip install .

You can pass arguments to cmake by exporting the variable CMAKE_ARGS, e.g., CMAKE_ARGS=-DSYM_ENABLE_TESTS=OFF to disable compiling the tests. Note that installing with pip will also change the directory of the Torch bindings shared library to your python-environment site-package directory.

Tests

After compiling (installation not required), the test executable SymmetryFunctions_tests will be in the directory build/bin. You can run the tests by executing ./SymmetryFunctions_tests.

Python Bindings

See src/Tests/python_bindings_test.py for an example.

Torch Bindings

See src/Tests/torch_bindings_test.py for an example.