GATGNN-VOLTAGE

This software package implements our work of GATGNN-VOLTAGE for the problem voltage prediction. With GATGNN-VOLTAGE, one can predict the material's voltage from:

1. a materials' formation-energy prediction alone or
2. the reaction of a low and high potential energy materials.

Please read our paper for the detailed implementation of GATGNN-VOLTAGE:

Accurate Prediction of Voltage of Battery Electrode Materials using Attention based Graph Neural Networks

Machine Learning and Evolution Laboratory
Department of Computer Science and Engineering
University of South Carolina

Table of Contents

• Installation
• Data
• Usage
• How to cite
• References

Installation

1. Inside of your Python environemnt, install the basic dependencies required for GATGNN-VOLTAGE by running code below:

pip install -r requirements.txt

2. Follow the instructions listed on Pytorch-Geometric's documentations to install pytorch-geometric for using Graph Neural Network.

Data

To obtain the dataset, run the get_data.py file.

python get_data.py

Usage

Reaction based voltage

For the reaction based voltage, run the voltage-reaction.py file. The 3 running options (evaluation, training, cross-validation or CV) can be set by using the --mode flag

evaluation

• Details

for evaluating the performance of the trained reaction-model. Running this mode predicts voltage of electrodes from the testing-set and saves those results to RESULTS/voltage--prediction.csv.

• Usage example:

python voltage-reaction.py --mode evaluation

---

training

• Details

for training a new reaction-based model.

• Optional arguments

Parameter	Default	Description
--train_size	0.8	ratio size of the training-set
--batch	128	batch size to use within experinment
--graph_size	small	graph encoding format by neighborhood size, either 12 (small) or 16 (large)
--layers	3	number of AGAT layers to use in model (default:3)
--neurons	64	number of neurons to use per AGAT Layer
--heads	4	number of Attention-Heads to use per AGAT Layer

• Usage example:

python voltage-reaction.py --mode training

---

cross-validation or CV:

• Details

for running a k-fold cross-validation training/ evaluation method. Running this mode creates k different prediction-results which are saved to RESULTS/{k}-voltage--prediction.csv; where k corresponds to the cross-validation iteration.

• Optional arguments

Parameter	Default	Description
--fold	10	number of folds
--train_size	0.8	ratio size of the training-set
--batch	128	batch size to use within experinment
--graph_size	small	graph encoding format by neighborhood size, either 12 (small) or 16 (large)
--layers	3	number of AGAT layers to use in model (default:3)
--neurons	64	number of neurons to use per AGAT Layer
--heads	4	number of Attention-Heads to use per AGAT Layer

• Usage example:

python voltage-reaction.py --mode cross-validation

Formation-energy based voltage

Upcoming soon

How to cite:

Louis, S. Y., Siriwardane, E., Joshi, R., Omee, S., Kumar, N., & Hu, J. (2022). Accurate Prediction of Voltage of Battery Electrode Materials Using Attention Based Graph Neural Networks.

References

1. Louis, S. Y., Zhao, Y., Nasiri, A., Wang, X., Song, Y., Liu, F., & Hu, J. (2020). Graph convolutional neural networks with global attention for improved materials property prediction. Physical Chemistry Chemical Physics, 22(32), 18141-18148.

2. Omee, S. S., Louis, S. Y., Fu, N., Wei, L., Dey, S., Dong, R., ... & Hu, J. (2021). Scalable deeper graph neural networks for high-performance materials property prediction. arXiv preprint arXiv:2109.12283.

3. Louis, S. Y., Nasiri, A., Rolland, F. J., Mitro, C., & Hu, J. (2021). NODE-SELECT: A Graph Neural Network Based On A Selective Propagation Technique. arXiv preprint arXiv:2102.08588.