SOMD is an ab-initio molecular dynamics (AIMD) package designed for the SIESTA DFT code. The SOMD code provides some common functionalities to perform standard Born-Oppenheimer molecular dynamics (BOMD) simulations, and contains a simple wrapper to the Neuroevolution Potential (NEP) package. The SOMD code may be used to automatically build NEPs by the mean of the active-learning methodology.

The SOMD code is designed to be maintained by one person, thus many important functionalities may be absent. Besides, the code should be considered EXPERIMENTAL since it has not been extensively tested. So if you would like to perform production runs with SOMD, please take your own risk.

SOMD only runs on GNU/Linux distros. The installation requires a working g++ compiler (with C++11 supports), a Python3 interpreter and four additional Python3 libraries (cython, h5py, mdtraj and toml). You could install SOMD by the following steps.

1. Install the required dependencies with:

   conda config --add channels conda-forge
   conda install cython h5py mdtraj toml -c conda-forge

   or

   pip install cython h5py mdtraj toml

2. Clone this repo:

   git clone https://www.github.com/initqp/somd
   cd somd
   git submodule update --init

   Note: if you would like to proceed your installation with the tarball downloaded from GitHub, you should manually download the NEP_CPU package and put it in the somd/somd/potentials/src directory. Besides, the version number of the installed package may be wrong.
3. Install SOMD:

   python setup.py install

   or

   pip install .

4. Start a python REPL and enter the following lines:

   >>> import somd
   >>> print(somd.__version__)

   If the installation is successful, a version string should be printed. Likewise, you could enter the following command under your shell:

   somd -v

   If the installation is successful, a version string should be printed as well.
5. Compile the SIESTA code. SOMD could work with the 4.1.5 or the git master version of SIESTA. When compiling, you are suggested to link your binary against the ELPA library (and using ELPA as the diagonalization algorithm). This is because of one of the memory leakage bugs in SIESTA (read this page for details). The usage of the ELPA library could be found in the SIESTA documentation.
6. If you would like to use DFTD3, DFTD4 and PLUMED with SOMD, you should also install the corresponding packages:

   conda install dftd3-python dftd4-python py-plumed -c conda-forge

   or

   pip install dftd3 dftd4 plumed

   Specifically, the above commands do not install the PLUMED kernel library for you. You should compile it separately and export the PLUMED_KERNEL environment variable before actually perform your PLUMED aided MD runs.

First, install the pytest package with:

conda install pytest -c conda-forge

or

pip install pytest

Then, enter the somd/tests directory and invoke this command (you need to change the SIESTA_COMMAND variable to the actual path of your siesta binary):

SIESTA_COMMAND='/path/to/siesta' py.test

SOMD has a naive command line interface, which reads the TOML format configure file. A typical input file looks like this (which defines a NVT run of a water molecule):

[system]
        structure = "H2O.POSCAR"
[[group]]
        atom_list = "all"
        initial_temperature = 300.0
[[potential]]
        type = "SIESTA"
        siesta_options = """
        xc.functional          GGA
        xc.authors             PBE
        PAO.BasisSize          DZP
        Mesh.Cutoff            300 Ry
        """
        siesta_command = "mpirun -np 4 /path/to/siesta"
[[trajectory]]
        format = "H5"
        file_name = "traj.h5"
        interval = 10
[[logger]]
        format = "CSV"
        file_name = "data.csv"
        interval = 10
[integrator]
        type = "BAOAB"
        timestep = 0.0005
        temperatures = 300.0
        relaxation_times = 0.1
[run]
        n_steps = 500

Based on this file (e.g., it is called input.toml), you could run your simulation via the following command:

somd -i input.toml

You may also invoke SOMD as a library and implement your own simulation protocols. For example, the above configure file equals to the following python script:

import somd

siesta_command = 'mpirun -np 4 /path/to/siesta'
siesta_options = r"""
xc.functional          GGA
xc.authors             PBE
PAO.BasisSize          DZP
Mesh.Cutoff            300 Ry
"""

system = somd.core.systems.create_system_from_poscar('H2O.POSCAR')
g = {'atom_list': list(range(0, system.n_atoms)), 'has_translations': False}
system.groups.create_from_dict(g)
system.groups[0].add_velocities_from_temperature(300)
potential = somd.potentials.create_siesta_potential(system,
                                                    range(0, system.n_atoms),
                                                    siesta_options,
                                                    siesta_command)
system.potentials.append(potential)

integrator = somd.core.integrators.baoab_integrator(0.0005,
                                                    temperatures=[300],
                                                    relaxation_times=[0.1],
                                                    thermo_groups=[0])
trajectory = somd.apps.trajectories.H5WRITER('traj.h5', write_forces=False,
                                             interval=10)
logger = somd.apps.loggers.DEFAULTCSVLOGGER('data.csv', interval=10)
simulation = somd.apps.simulations.SIMULATION(system=system,
                                              integrator=integrator,
                                              trajectories=[trajectory],
                                              loggers=[logger])

simulation.run(500)

Based on this script (e.g., it is called input.py), you could run your simulation via the following command:

python input.py

A problem-oriented documentation could be found here.

Tutorials of SOMD could be found here. Going through these tutorials is considered as an efficient way to get familiar with SOMD.

• Q: There are millions of MD packages out there, why do you need another one?

  A: Because I write it for fine. If you are seeking for unity, you may want to use some mainstream packages like i-Pi, ASE, Tinker or even CP2K. Each of them provides excellent functionalities.

• Q: Will SOMD support other ab-initio packages, like VASP?

  A: Maybe. But I can not afford a VASP license, it is way luxurious for me.

• Q: Are you riding on the wave of machine learning potentials?

  A: I am. And I'm tired of pretending I'm not. 🤡

• Q: How to cite the code?

  A: You do not have to. But if your publisher forces you to do so, you may cite the GitHub repo directly.