MDDatasetBuilder is a script to construct reference datasets for the training of neural network potentials from given LAMMPS trajectories.
Complex reaction processes in combustion unraveled by neural network-based molecular dynamics simulation, Nature Communications, 11, 5713 (2020), DOI: 10.1038/s41467-020-19497-z
Author: Jinzhe Zeng
Email: [email protected]
MDDatasetBuilder can be installed with pip:
pip install mddatasetbuilder
The installation process should be very quick, taking only a few minutes on a “normal” desktop computer.
datasetbuilder -d dump.ch4 -b bonds.reaxc.ch4_new -a C H O -n ch4 -i 25
dump.ch4 is the name of the dump file.
bonds.reaxc.ch4_new is the name of the bond file, which is optional.
C H O is the element in the trajectory.
ch4 is the name of the dataset.
25 means the time step interval and the default value is 1.
Then you can generate Gaussian input files for each structure in the dataset and calculate the potential energy & atomic forces (assume the Gaussian 16 has already been installed.):
qmcalc -d dataset_ch4_GJf/000 qmcalc -d dataset_ch4_GJf/001
Next, prepare a DeePMD dataset and use DeePMD-kit to train a NN model.
preparedeepmd -p dataset_ch4_GJf -a C H O cd train && dp train train.json
The runtime of the software depends on the amount of data. It is more suited to running on a server rather than desktop computer.
The MDDatasetBuilder package has been integrated with DP-GEN software:
dpgen init_reaction reaction.json machine.json
where an example of
reaction.json can be found here, and
machine.json should include the following keys:
reaxff_command is the LAMMPS command,
build_command is the MDDatasetbuilder command, and
fp_command is the Gaussian 16 command.