Libmbd implements the many-body dispersion (MBD) method in several programming languages and frameworks:
- The Fortran implementation is the reference, most advanced implementation, with support for analytical gradients and distributed parallelism, and additional functionality beyond the MBD method itself. It provides a low-level and a high-level Fortran API, as well as a C API. Furthermore, Python bindings to the C API are provided.
- The Python/Numpy implementation is intended for prototyping, and as a high-level language reference.
- The Python/Tensorflow implementation is an experiment that should enable rapid prototyping of machine learning applications with MBD.
The Python-based implementations as well as Python bindings to the Libmbd C API are accessible from the Python package called Pymbd.
Libmbd is included in FHI-aims, Quantum Espresso, DFTB+, and ESL Bundle.
TL;DR Install prebuilt Libmbd binaries via Conda-forge and Pymbd with Pip.
conda install -c conda-forge libmbd pip install pymbd
One can also install the ScaLAPACK/MPI version.
conda install -c conda-forge 'libmbd=*=mpi_*' mpi4py pip install pymbd[mpi]
Verify installation with
$ python -m pymbd Expected energy: -0.0002462647623815428 Calculated energy: -0.0002462647623817456
Libmbd uses CMake for compiling and installing, and requires a Fortran compiler, LAPACK, and optionally ScaLAPACK/MPI.
apt-get install gfortran libblas-dev liblapack-dev [mpi-default-dev mpi-default-bin libscalapack-mpi-dev]
brew install gcc [open-mpi scalapack]
The compiling and installation can then proceed with
cmake -B build [-DENABLE_SCALAPACK_MPI=ON] make -C build install [ctest --test-dir build]
This installs the Libmbd shared library, C API header file, high-level Fortran API module file, and Cmake package files, and optionally runs tests.
Pymbd can be installed and updated using Pip, but requires installed Libmbd as a dependency (see above).
pip install pymbd
To support Libmbd built with ScaLAPACK/MPI, the
mpi extras is required, which installs
mpi4py as an extra dependency. In this case one has to make sure that
mpi4py is linked against the same MPI library as Libmbd (for instance by compiling both manually, or installing both via Conda-forge).
pip install pymbd[mpi]
If Libmbd is installed in a non-standard location, you can point Pymbd to it with
env LIBMBD_PREFIX=<path to Libmbd install prefix> pip install pymbd
If you don’t need the Fortran bindings in Pymbd, you can install it without the C extension, in which case
pymbd.fortran becomes unimportable:
env LIBMBD_PREFIX= pip install pymbd
from pymbd import mbd_energy_species from pymbd.fortran import MBDGeom # pure Python implementation energy = mbd_energy_species([(0, 0, 0), (0, 0, 7.5)], ['Ar', 'Ar'], [1, 1], 0.83) # Fortran implementation energy = MBDGeom([(0, 0, 0), (0, 0, 7.5)]).mbd_energy_species( ['Ar', 'Ar'], [1, 1], 0.83 )
use mbd, only: mbd_input_t, mbd_calc_t type(mbd_input_t) :: inp type(mbd_calc_t) :: calc real(8) :: energy, gradients(3, 2) integer :: code character(200) :: origin, msg inp%atom_types = ['Ar', 'Ar'] inp%coords = reshape([0d0, 0d0, 0d0, 0d0, 0d0, 7.5d0], [3, 2]) inp%xc = 'pbe' call calc%init(inp) call calc%get_exception(code, origin, msg) if (code > 0) then print *, msg stop 1 end if call calc%update_vdw_params_from_ratios([0.98d0, 0.98d0]) call calc%evaluate_vdw_method(energy) call calc%get_gradients(gradients) call calc%destroy()
- Libmbd documentation: https://libmbd.github.io
- Pymbd documentation: https://libmbd.github.io/pymbd
For development, a top-level
Makefile is included, which configures and compiles Libmbd, compiles the Pymbd C extension, and runs both Libmbd and Pymbd tests.
git clone https://github.com/libmbd/libmbd.git && cd libmbd python3 -m venv venv && source venv/bin/activate make # development work... make