Atomistic Force Fields based on GNNFF
GNNFF [1] is a graph neural network framework to directly predict atomic forces from automatically extracted features of the local atomic environment that are translationally-invariant, but rotationally-covariant to the coordinate of the atoms. This package is an atomistic force fields that constructed based on GNNFF.
Requirements:
- python 3
- ASE
- numpy
- PyTorch (>=1.7)
- schnetpack
Note: We recommend using a GPU for training the neural network.
Install this project by git
.
git clone https://github.com/ken2403/gnnff.git
cd gnnff
pip install -r requirements.txt
pip install .
You're ready to go!
The example scripts provided by gnnff are inserted into your PATH during installation. These scripts are in gnnff/src/scripts
.
Prepare the dataset in ase database format.
Or prepare a atoms .xyz
file and convert it to ase database using the script provided by schnetpack.
spk_parse.py /path_to_xyz /save_path_to_db
Prepare the .json
file that defines the hyperparameters for training.
The example file can be found in gnnff/src/scripts/argument
If the argument .json
file is ready, you can train the model like this way.
gnnff_run from_json /path_to_train.json
Learned model are saved as best_model
in model direcctory.
The same script can also be used to evaluate the learned model.
Prepare a .json
file for evaluation.
gnnff_run from_json /path_to_eval.json
Write a result file evaluation.txt
into the model directory.
You can also input one POSCAR
structure file into the learned model to calculate total energy and interatomic forces.
This is useful for phonon calculations with phonopy.
gnnff_run from_poscar /path_to_POSCAR /path_to_learned_model --cutoff cutoff_radious [--cuda]
If GPU is availabel, set --cuda
.
The calculation results are written to gnnff_run.xml
file and saved in the same directory as the POSCAR
.
-
[1] C. W. Park, M. Kornbluth, J. Vandermause, C. Wolverton, B. Kozinsky, J. P. Mailoa, Accurate and scalable graph neural network force field and molecular dynamics with direct force architecture. npj Computational Materials. 7, 1–9 (2021). link
-
[2] T. Xie, J. C. Grossman, Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties. Phys. Rev. Lett. 120, 145301 (2018). link
-
[3] implementaion of Crystal Graph Convolutional Neural Networks