These are examples that work the MDI package in LAMMPS which uses the MolSSI MDI library for coupling codes together and communicating between them with MDI messages. Within the MDI context, one code is the driver and another code is the engine. The 2 codes can be written in any language; C++ (LAMMPS) and Python are illustrated here. The 2 codes can each be stand-alone codes, in which case they can be run on different numbers of processors. The 2 codes can communicate either via TCP (sockets) or via MPI. For the TCP case, the driver and engine need to be launched separately, e.g. in 2 windows on your desktop machine. For the MPI case, a single mpirun command launches both codes. Alternatively the engine code can be a plugin library which the driver code loads, in which case the driver and engine run on the same processors. LAMMPS supports operating in all these MDI modes. It can be an engine operating either as a stand-alone code or as a plugin. It can also be a driver and couple to an engine that is either a stand-alone code or a plugin. Examples for all these use cases, using LAMMPS as a driver and as an engine are in this directory. The Run.sh file shows how run in all the modes. Type "sh Run.sh" to try them all out. Examples for using LAMMPS as a driver with the tight-binding code LATTE are in the LATTE sub-directory. See its README file for more info. Note that to use LAMMPS as a plugin engine for some the examples in this dir, you must build it as a shared library. Something like this with traditional make, which also builds the normal LAMMPS executable lmp_mpi: cd src make yes-mdi make mode=shlib mpi ------------------------------------------------- Examples 4 and 5 use Python scripts as MDI drivers. For this you will need Python 3 with Numpy and mpi4py installed. Make sure LAMMPS and Python/mpi4py are using same the same version of MPI. You will also need MDI installed in your Python. You cannot use the LAMMPS build of the MDI library for this, b/c LAMMPS builds MDI as a static library, while Python requires a shared library. Instead you can install MDI in your Python via conda: % conda install -c conda-forge pymdi=1.4.1 or via pip: % pip install pymdi==1.4.1 It is likely fine to leave off the version number, to get the latest MDI version. But to be safe, 1.4.1 is the version LAMMPS is currently using. ------------------------------------------------- 5 example use-cases are explained below. In the first 3 examples, results running with MDI should be identical to running without MDI (alone log files). Example #4 has no non-MDI run. Example #5 results should match the non-MDI run of example #1. ------------------------------------------------- * Example #1 = run ab initio MD (AIMD) Two instances of LAMMPS operate as a driver and engine. As an engine, LAMMPS is a surrogate for a quantum code. You can compare the thermo output in log.aimd.alone.1 to the thermo output in any of the log.aimd.driver* files. It should be identical. Note that the "alone" and "driver" input scripts have options for running in NVE vs NPT Comment in/out the appropriate line to make change. Nothing needs to be changed in the "engine" scripts. ------------------------------------------------- ------------------------------------------------- * Example #2 = run LAMMPS, compute QM forces on snapshots from a long run Two instances of LAMMPS operate as a driver and engine. As an engine, LAMMPS is a surrogate for a quantum code You can compare the thermo output in log.snapshot.alone.1 to the thermo output in any of the log.snapshot.driver* files. It should be identical. You can compare the dumped forces in dump.snapshot.alone.1 to the forces in any of the dump.snapshot.* files. They should be identical, although at step 0 the forces are "zero" and may be epsilon different. ------------------------------------------------- ------------------------------------------------- * Example #3 = run LAMMPS, compute QM forces on series of independent systems Two instances of LAMMPS operate as a driver and engine. As an engine, LAMMPS is a surrogate for a quantum code You can compare the thermo output in log.series.alone.1 to the thermo output in any of the log.series.driver* files. It should be identical. You can compare the dumped forces in dump.series.alone.1 to the forces in any of the dump.series.* files. They should be identical, ------------------------------------------------- ------------------------------------------------- * Example #4 = Python driver runs a sequence of unrelated LAMMPS calculations Each calculation can be a single-point evaluation, MD run, or minimization The sequence_driver.py code allows for optional switches in addition to -mdi (required) and the -plugin and -plugin_args switches which are used to link to an engine as a plugin library. The example run commands below just use the default values of the optional switches. The switches are also explained at the top of the file; the info is copied here: # -n 3 # number of calculations to perform, default = 3 # -mode eval/run/min # style of calculations: single snapshot evals, dynamics, minimization # default = run # -size Nx Ny Nz # cubic lattice, default = 2 2 2 # -rho 0.75 0.1 # reduced density and random variation thereof, default = 0.75 0.1 # -delta 0.1 # randomly perturb atoms initially by this distance, default 0.1 # -nsteps 100 # number of timesteps in dynamics runs, default = 100 # -temp 1.0 # initial temperature in dynamics runs, default = 1.0 # -tol 0.001 # tolerance for minimizations, default = 0.001 # -seed 12345 # random number seed > 0, default = 12345 You can compare the thermo output in any of the log.sequence.engine.* files. It should be identical. ------------------------------------------------- ------------------------------------------------- * Example #5 = run AIMD with Python driver code and 2 LAMMPS instances as engines First LAMMPS instance performs the MD timestepping. Second LAMMPS instance is surrogate QM to compute forces. The aimd_driver.py code allows for an optional switch in addition to -mdi (required) and the -plugin and -plugin_args swiches which are used to link to the 2 engines as a plugin libraries. The example run commands below use the default values of the optional switch. The switch is also explained the top of the file; the info is copied here: # -nsteps 10 # number of timesteps in dynamics runs, default = 10 This calculation is the same as Example #1 with a LAMMPS driver and a LAMMPS engine. Now there is a Python driver and two LAMMPS engines. You can compare the thermo output in log.aimd.alone.1 output to the thermo output is any of the log.sequence.engine.* files. It should be identical for the Total Energy printed out by the Python driver script. E.g. Step 10: MM energy 1.69875, QM energy -6.31535, Total energy -4.6166