G-metaD can be used to sample a training set of machine learning potentials via metadynamics, using an atom-centered symmetry function vector (G-space) as the collective variable.

D. Yoo, J. Jung, W. Jeong, and S. Han, Metadynamics sampling in atomic environment space for collecting training data for machine learning potentials, npj Computational Materials 7, 131 (2021), [https://doi.org/10.1038/s41524-021-00595-5]

vasp_wrap.py is a wrapper on the VASP quantum DFT code so it can work as a "server" code which LAMMPS drives as a "client" code to perform ab initio MD. LAMMPS performs the MD timestepping, sends VASP a current set of coordinates each timestep, VASP computes forces and energy and virial and returns that info to LAMMPS.

Messages are exchanged between MC and LAMMPS via a client/server library (CSlib), which is included in the LAMMPS distribution in lib/message. As explained below you can choose to exchange data between the two programs either via files or sockets (ZMQ).

To make waiting LAMMPS client not consume 100% CPU usage while waiting for MPI operations, you can use modified version of CSlib in this repository.

---

Requirement

• LAMMPS (29Oct2020)
• Eigen

---

Building

Build LAMMPS with its MESSAGE (OPENMP if needed) package installed: See the Build extras doc page and its MESSAGE package section for details. [doc]

cp -r cslib lammps/lib/message/  # copy modified cslib to lammps
cd lammps/lib/message
python Install.py -m -z          # build CSlib with MPI and ZMQ support
cp -r Eigen lammps/src/
cp pair_mtd.* symmetry_function.h lammps/src
cd lammps/src
make yes-message
make mpi

You can leave off the -z if you do not have ZMQ on your system.

---

Build the CSlib in a form usable by the vasp_wrapper.py script:

cd lammps/lib/message/cslib/src
make shlib            # build serial and parallel shared lib with ZMQ support
make shlib zmq=no     # build serial and parallel shared lib w/out ZMQ support

This will make a shared library versions of the CSlib, which Python requires. Python must be able to find both the cslib.py script and the libcsnompi.so library in your lammps/lib/message/cslib/src directory. If it is not able to do this, you will get an error when you run vasp_wrapper.py.

You can do this by augmenting two environment variables, either from the command line, or in your shell start-up script. Here is the sample syntax for the csh or tcsh shells:

export PYTHONPATH=$PYTHONPATH:/path/to/lammps/lib/message/cslib/src
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/lammps/lib/message/cslib/src

---

Running

Prepare to use VASP and the vasp_wrapper.py script

Insure you have the necessary VASP input files in this directory, suitable for the VASP calculation you want to perform:

• INCAR
• KPOINTS
• POSCAR_template
• POTCAR

Examples of all but the POTCAR file are provided. The POTCAR file is a proprietary VASP file, so use one from your VASP installation.

Note that the POSCAR_template file should be matched to the LAMMPS input script (# of atoms and atom types, box size, etc).

---

The parameters in pair_style are condition number of covariance matrix, the number of elements, symbol, the height (eV) and width (eV/Angstrom) of bias potential, update interval, the number of types, and the coefficient of bias.

pair_style mtd 1e-4 1 &
               Si 0.001 1.0 20 &
               1 &
               1.0

If the system consist of two elements such as GeTe, in.client is following:

pair_style mtd 1e-4 2 &
               Ge 0.001 1.0 20 &
               Te 0.001 1.0 20 &
               2 &
               1.0 1.0

---

To run in client/server mode:

NOTE: The vasp_wrap.py script must be run with Python version 2, not 3. This is because it used the CSlib python wrapper, which only supports version 2.

Both the client (LAMMPS) and server (vasp_wrap.py) must use the same messaging mode, namely file or zmq. This is an argument to the vasp_wrap.py code; it can be selected by setting the "mode" variable when you run LAMMPS. The default mode is file.

Here we assume LAMMPS was built to run in parallel, and the MESSAGE package was installed with socket (ZMQ) support. This means either of the messaging modes can be used and LAMMPS can be run in serial or parallel. The vasp_wrap.py code is always run in serial, but it launches VASP from Python via an mpirun command which can run VASP itself in parallel.

When you run, the server should print out thermodynamic info every timestep which corresponds to the forces and virial computed by VASP.

The examples below are commands you should use in two different terminal windows. The order of the two commands (client or server launch) does not matter. You can run them both in the same window if you append a "&" character to the first one to run it in the background.

---

File mode of messaging:

python vasp_wrap.py file POSCAR_template "mpirun -np 1 vasp.x" &
mpirun -np 1 lmp_mpi -v mode file -in in.client

ZMQ mode of messaging:

python vasp_wrap.py zmq POSCAR_template "mpirun -np 1 vasp.x" &
mpirun -np 1 lmp_mpi -v mode zmq -in in.client

---

You might have to set some environment variables for oversubscribing. For example, the commands below can enable both processes (vasp_wrap.py and LAMMPS) use all cores.

export PSM2_SHAREDCONTEXTS=YES
export PSM2_MAX_CONTEXTS_PER_JOB=8

---

Caution

1. Atom types in LAMMPS and VASP (POSCAR) may differ and cause unexpected problems. The code does not check for consistency.
2. The convergence of electronic step in VASP is not checked.
3. Some errors will not terminate your job. For example, when VASP calculation fails causing vasp_wrap.py to terminate, LAMMPS hangs waiting for server response.
4. Since VASP have to start and terminate on every step, it could be much slower than VASP MD for small systems due to the overhead.