Reference implementation in TensorFlow 2 of the geometric message passing neural network (GemNet). You can find its PyTorch implementation in another repository. GemNet is a model for predicting the overall energy and the forces acting on the atoms of a molecule. It was proposed in the paper:
GemNet: Universal Directional Graph Neural Networks for Molecules
by Johannes Klicpera, Florian Becker, Stephan Günnemann
Published at NeurIPS 2021.
Adjust config.yaml (or config_seml.yaml) to your needs.
This repository contains notebooks for training the model (train.ipynb
) and for generating predictions on a molecule loaded from ASE (predict.ipynb
). It also contains a script for training the model on a cluster with Sacred and SEML (train_seml.py
).
You can either use the precomputed scaling_factors (in scaling_factors.json) or compute them yourself by running fit_scaling.py. Scaling factors are used to ensure a consistent scale of activations at initialization. They are the same for all GemNet variants.
Please contact [email protected] if you have any questions.
Please cite our paper if you use the model or this code in your own work:
@inproceedings{klicpera_gemnet_2021,
title = {GemNet: Universal Directional Graph Neural Networks for Molecules},
author = {Klicpera, Johannes and Becker, Florian and G{\"u}nnemann, Stephan},
booktitle={Conference on Neural Information Processing Systems (NeurIPS)},
year = {2021}
}