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:

Note: We recommend using a GPU for training the neural network.

Installation

Install this project by git.

Clone the repository

git clone https://github.com/ken2403/gnnff.git
cd gnnff

Install requirements

pip install -r requirements.txt

Install this package

pip install .

You're ready to go!

Getting started

The example scripts provided by gnnff are inserted into your PATH during installation. These scripts are in gnnff/src/scripts.

Preparing the dataset

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

Preparing argument file

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.

Evaluating learned model

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.

Using POSCAR file as an input

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.

References

[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

hitergelei/gnnff