Crystal Graph Convolutional Neural Networks

This experimental study was carried out at AstraZeneca's Department. However, the project's focus isn't solely on machine learning; rather, it's on applying machine learning to the subject of materials science. To be more specific, it has to do with the field of material discovery, namely the improvement of properties. The research centered on two fundamental process:

• Train a Crystal Graph Convolutional Neural Network model with a customized dataset.
• Predict material properties of selected crystals with the CGConv Neural Network model.

AIM

The goal of this study is to see if graph neural networks can be utilized to predict crystal characteristics and hence speed up the crystal structure prediction framework. As a result, we concentrated on examining the proposed strategy from two perspectives:

• Computational speed and financial cost.
• How far we can go with the graph neural network model in material science? We checked the model's transferability on two different targets: Formation Energy and Intensity values of x-ray powder diffraction (XRPD) technique.

Prerequisites

This packages requires:

• PyTorch
• scikit-learn
• pymatgen
• pytorch-lightning

The easiest way of installing the prerequisites is via 'conda'. If you are using the JupyterHUB, run the following command to create a new environment and then install all prerequisites.

# create new environment
$ ml Miniconda3
$ conda create -n jcml19 python=3.8.12
# activate new environment
$ conda activate jcml19

# create definition file 
cat << EOF > jcml19.yml
---
display_name: 'JCML19_python_3.8'
language: 'python'
manager: 'conda'
conda_env: 'jcml19'
EOF
# run the script
ml jupyterlab_kernel_generator
CONDA_PREFIX="/opt/scp/software/Anaconda3/2019.10" jupyterlab_kernel_generator --definition ./jcml19.yml

After creating a new environment then install all needed prerequisites.

# Pytorch install
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html

# pymatgen install
pip install pymatgen==2022.2.7

# pytorch-lightning install
pip install pytorch-lightning=1.5.9

Usage

Preprocessing

• The notebook collect-data.ipynb collects structures, and their intensity values from the folder database and save them as CSV format in folder data.

• The dataset folder shows:

The train.csv, valid.csv, and test.csv after preprocessing (standardization) step.

The org-train.csv, org-valid.csv, and org-test.csv without preprocessing (standardization) step.

• The atom-init.json file that stores the initialization vector for each element.
• The folder content and log contain the model's checkpoints and metrics [Mean Absolute Error (MAE) and Mean Square Error (MSE)], respectively.

Phase One: Formation Energy

• The notebook gnn-qm.ipynb that displays the CGConv Neural network model's training, testing, and prediction stage on formation energy.

Phase Two: Intensity values (XRPD Technique)

The folder XRPD, contains:

• The notebook gnn-xrpd.ipynb that displays the CGConv Neural network model's training, testing, and prediction stage on Intensity values after XRPD technique.

Citation

If you use this code or the models in your research, please cite the following:

@misc{marziehphi,
  author = {Farahani, Marzieh},
  title = {Accelerate Prediction of Crystal Material Properties Using Deep GNNs},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/marziehphi/CrystalGNN}},
}