pymbd 0.9.3

Many-body dispersion library

Libmbd

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.

Installing

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

Libmbd uses CMake for compiling and installing, and requires a Fortran compiler, LAPACK, and optionally ScaLAPACK/MPI.

On Ubuntu:

apt-get install gfortran libblas-dev liblapack-dev [mpi-default-dev mpi-default-bin libscalapack-mpi-dev]

On macOS:

brew install gcc [open-mpi scalapack]

The compiling and installation can then proceed with

git clone https://github.com/jhrmnn/libmbd.git && cd libmbd
mkdir build && cd build
cmake .. [-DENABLE_SCALAPACK_MPI=ON]
make
make install

This installs the Libmbd shared library, C API header file, and high-level Fortran API module file.

Pymbd

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

Examples

from pymbd import mbd_energy_species, ang
from pymbd.fortran import MBDCalc

ene_py = mbd_energy_species(  # pure Python implementation
    [(0, 0, 0), (0, 0, 4*ang)], ['Ar', 'Ar'], [1, 1], 0.83
)
with MBDCalc() as calc:
    ene_f = calc.mbd_energy_species(  # Fortran implementation
        [(0, 0, 0), (0, 0, 4*ang)], ['Ar', 'Ar'], [1, 1], 0.83
    )
assert abs(ene_f-ene_py) < 1e-15

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
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()

Links

• Libmbd documentation: https://jhrmnn.github.io/libmbd
• Pymbd documentation: https://jhrmnn.github.io/libmbd/pymbd

Developing

For development, Libmbd doesn't have to be installed on the system, and Pymbd can be linked against local installation of Libmbd.

git clone https://github.com/jhrmnn/libmbd.git && cd libmbd
python3 -m venv venv && source venv/bin/activate
make
# development work...
make