Crystal Graph Convolutional Neural Network (CGCNN) Tutorial 2

Step 1: Download and Install Anaconda

• Go here and enter your email. You'll receieve a donload link in your email.

• Once you've successfuly downloaded the Anaconda3-2024.10-1-Linux-x86_64.sh file, install it with,

  $ bash Anaconda3-2024.10-1-Linux-x86_64.sh

and follow the given prompts.

Step 2: Install and configure CGCNN

• Upgrade conda $ conda upgrade conda

• Install prerequisites and CGCNN $ conda create -n cgcnn python=3 scikit-learn pytorch torchvision pymatgen -c pytorch -c conda-forge

• Activate cgcnn conda environment $ conda activate cgcnn

  $ conda install -c conda-forge mp-api

• Clone the cgcnn repo from github; $ git clone https://github.com/txie-93/cgcnn.git

Step 3: Using cgcnn

Step 3.1: Prepare data

• Go to the cloned repository $ cd cgcnn

• Inside the data directory, create a new directory with anyname, in this case formation_energy

$ cd data

$ mkdir formation_energy-trained

$ cd formation_energy and inside it have the following files: 1. CIF files having the format ID.cif 2. id_prop.csv which contains material id and property we want to predict in the first and second column respectively. 3. atom_init.json a JSON file that stores the initialization vector for each element. This data/sample-regression/atom_init.json file should be good for most applications.

• The CIF and id_prop.csv files can be obtained by editing generateCIF.py file (which can be downloaded from here) appropriately to suit your needs and then running;

  $ python3 generateCIF.py

• The atom_init.json file can be obtained by; $ cp ../sample-regression/atom_init.json .

Step3.2: The Util directory

• Create a Util directory at the root of the cgcnn directory to store utility scripts and data.

  $ mkdir Util

• In the data/formation_energy directory, move the generateCIF.py and full_dataset.csv files to the newly created Util directory.

  $ mv generateCIF.py ../../Util/

  $ mv full_dataset.csv ../../Util/

Step 3.3: Train data

• To train the model;

  $ python main.py --train-ratio 0.6 --val-ratio 0.2 --test-ratio 0.2 path/to/data/formation_energy 2>&1 | tee training.log

• To proceed with the calculation after an interuption;

  $ python3 main.py --train-ratio 0.6 --val-ratio 0.2 --test-ratio 0.2 data/formation_energy/ --resume checkpoint.pth.tar 2>&1 | tee training.log

• After training, you will get three files in cgcnn directory;

  1. model_best.pth.tar: stores the CGCNN model with the best validation accuracy.
  2. checkpoint.pth.tar: stores the CGCNN model at the last epoch.
  3. test_results.csv: stores the ID, target value, and predicted value for each crystal in test set.

Step 3.4: Predict material properties

• In the data directory create a new directory (in this case formation_energy_prediction ) and inside it have the CIF files, the atom_init.json and the id_prop.csv file.

• Run the following command at the root of cgcnn directory;

  $ python3 predict.py post-processing/model_best.pth.tar data/formation_energy_prediction/ 2>&1 | tee formation_energy_prection.log

• All the output from running the training and prediction is contained in the post-processing directory.

• After predicting, you will get one file in cgcnn directory:

  1. test_results.csv: stores the ID, target value, and predicted value for each crystal in test set. Here the target value is just any number that you set while defining the dataset in id_prop.csv, which is not important.

For more info see the github repo by Tian Xie here and the article on arXiv .