A modification of DPGEN code, we add the dimer method as a method of model deviation, which could make our MLFF model predict the saddle points with high precision and few training data.

python>=3.8.0

dpgen

You just need to put our run.py and arginfo.py into your work directory, then prepare the param.json file and machine.json file required by dpgen. You also need a file named input.lammps, no mattter what content is, we didn't use it but the code needs it. To use the dimer method, you need to set the model_devi_engine to dimer in the machine.json file, and set the command to your command to run the python code dimer.py. For instance:

CUDA_VISIBLE_DEVICES=1 srun python /home/yinbc/test/dpgen/dpgen/Li-TiO2//dimer.py True  24  0.05  24  2

You can also use other model devi methods of dpgen, such as enhanced sampling. And you can also try to mix them.

auto.py in the example folder: A script which could automatically generate input files for multiple examples. To use this, you need to put your initial vasp MD caculation folders together, and use dpdata to generate the initial data, for instance:

import dpdata
import numpy as np
vaspmd_dir = '.'
data=dpdata.LabeledSystem('OUTCAR')
data.to_deepmd_raw('deepmd')
data.to_deepmd_npy('deepmd')

Besides the calculation folders, put run.py, arginfo.py, dimer.py, Templates of INCAR, param.json, machine.json, input.lammps together also. Then you can run python auto.py. Then you may adjust some of the parameters for each specific example.

test_process.py:This code could generate muliple search trajectory for the given initial configuration, and give the predicted energy. The energy are in the ener.txt, and we would get each trajectory folder.

test_saddle.py: This code could generate muliple dimer search to get saddle points and corresponding barriers for the given initial configuration. The barriers are in the barrier_sad.txt.

In our code, you just need one POTCAR, not one POTCAR for each element.