cnncnnzh / cgnn

A Graph Convolutional Neuron Network Designed for Material Properties

CGNN

A graph convolutional neuron network designed for predicting crystalline material properties.

Quick Run

• To train a model:

  python main.py --root_dir=./tests_train

• To predict:

  python main.py --predict --root_dir=./test_predict

• To change the settings of training:

  python main.py --root_dir=./test_train --epochs=100 --train_ratio=0.6 --val_ratio=0.2 --test_ratio=0.2`
  
  python main.py --root_dir=./test_train --dropout_rate=0.5 --loss_func=CrossEntropyLoss --optimizer=RMSprop
  

Input Format

In the 'training' mode, the directory root_dir should include .cif files of all the crystals, and a targets.csv file in which the first column contains the names of .cif files, and the second column contains the target values.

In the 'predict' mode, only .cif files should be provided. DO NOT include a targets.csv file.

Examples can be found in tests_train and test_predict folders.

References

https://www.cell.com/patterns/fulltext/S2666-3899(22)00076-9#%20

https://pubmed.ncbi.nlm.nih.gov/29694125/