Additional Features¶
Δ-NNP using LJ-like force field¶
Before training the neural network (step 4), the parameters of the classical force field are optimized.
sannp --classical
Executing the above outputs the file sannp.class
, which contains the parameters of the classical force field.
Then, set withLJlike 1
in sannp.prop
and perform step 4 and onward.
Δ-NNP using ReaxFF¶
Set withReaxFF 1
in sannp.prop
and perform training of the neural network (step 4).
The parameter definition file ffield.reax
is required when training or executing the molecular dynamics calculation (step 5). You can use one included in NanoLabo Tool. The syntax of ffield.reax
file is compatible with LAMMPS.
Hint
Existing parameters are used for ReaxFF and no optimization is performed within this product
Structure Generation by Metropolis Method¶
It is executed after the neural network training (step 4) is completed.
Prepare the settings for the Metropolis method in the file sannp.metro
. No file is required with the default settings.
sannp --metro
When executing the above, a Monte Carlo calculation is performed and a structure is generated. The generated structure is output to the dft_geom
folder in the form of a Quantum ESPRESSO input file.
Hint
The Monte Carlo calculation process is saved in xyz format in the movie folder.
Dragging and dropping the xyz file onto the Advance/NanoLabo screen make it visible as a movie.
Following that, run
sannp --dft
to generate the shell script for the calculations in dft_geom
, dft_run.sh
(or the batch file dft_run.bat
on Windows).
In-situ Tests¶
This function calculates and displays the residual RMS for test data at the same time when training a neural network (Step 4). While learning proceeds, that makes it possible to check the performance against the test data at any time. In addition, if the RMS of the test data is large even though one of the training data is small, it can be assumed that overlearning has occurred.
First, prepare test data that is separate from the training data. Either run sannp --dft
and select “test” to generate it, or split the existing training data sannp.train with sannp --split
. Once the test data sannp.test
is ready, set insituTest 1
in sannp.prop
and run the training.
Conversion of ASE trajectory file¶
You can convert Atomic Simulation Environment(ASE) `trajectory (.traj) file <https://wiki.fysik.dtu.dk/ase/ase/io/trajectory.html>`_to training data and use it for the training of the neural network.
Go to the install folder and use the Python script traj_to_train.py
in python
folder as
python traj_to_train.py filename.traj
to output the training data sannp.train
. The procedure following the neural network training (step 4) is the same as normal.
Following options are available.
- -output_file <file_name>¶
Specify the output file. Default is
sannp.train
.
- -mode {a,w,x}¶
Specify the behavior of file output.
a
: append,w
: overwrite andx
: create new file (error if file already exists). Default isx
.
- -start <int>¶
First index of the data to be converted. Default is 0.
- -step <int>¶
The step of the data to be converted. Default is 1.