This software package implements the Crystal Graph Convolutional Neural Networks (CGCNN) that takes an arbitary crystal structure to predict material properties.
The package provides two major functions:
- Train a CGCNN model with a customized dataset.
- Predict material properties of new crystals with a pre-trained CGCNN model.
The following paper describes the details of the CGCNN framework:
Please cite the following work if you want to use CGCNN.
@article{PhysRevLett.120.145301,
title = {Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties},
author = {Xie, Tian and Grossman, Jeffrey C.},
journal = {Phys. Rev. Lett.},
volume = {120},
issue = {14},
pages = {145301},
numpages = {6},
year = {2018},
month = {Apr},
publisher = {American Physical Society},
doi = {10.1103/PhysRevLett.120.145301},
url = {https://link.aps.org/doi/10.1103/PhysRevLett.120.145301}
}
This package requires:
If you are new to Python, the easiest way of installing the prerequisites is via conda. After installing conda, run the following command to create a new environment named cgcnn
and install all prerequisites:
conda upgrade conda
conda create -n cgcnn python=3 scikit-learn pytorch torchvision pymatgen -c pytorch -c conda-forge
*Note: this code is tested for PyTorch v1.0.0+ and is not compatible with versions below v0.4.0 due to some breaking changes.
This creates a conda environment for running CGCNN. Before using CGCNN, activate the environment by:
source activate cgcnn
Then, in directory cgcnn
, you can test if all the prerequisites are installed properly by running:
python main.py -h
python predict.py -h
This should display the help messages for main.py
and predict.py
. If you find no error messages, it means that the prerequisites are installed properly.
After you finished using CGCNN, exit the environment by:
source deactivate
To input crystal structures to CGCNN, you will need to define a customized dataset. Note that this is required for both training and predicting.
Before defining a customized dataset, you will need:
- CIF files recording the structure of the crystals that you are interested in
- The target properties for each crystal (not needed for predicting, but you need to put some random numbers in
id_prop.csv
)
You can create a customized dataset by creating a directory root_dir
with the following files:
-
id_prop.csv
: a CSV file with two columns. The first column recodes a uniqueID
for each crystal, and the second column recodes the value of target property. If you want to predict material properties withpredict.py
, you can put any number in the second column. (The second column is still needed.) -
atom_init.json
: a JSON file that stores the initialization vector for each element. An example ofatom_init.json
isdata/sample-regression/atom_init.json
, which should be good for most applications. -
ID.cif
: a CIF file that recodes the crystal structure, whereID
is the uniqueID
for the crystal.
The structure of the root_dir
should be:
root_dir
├── id_prop.csv
├── atom_init.json
├── id0.cif
├── id1.cif
├── ...
There are two examples of customized datasets in the repository: data/sample-regression
for regression and data/sample-classification
for classification.
For advanced PyTorch users
The above method of creating a customized dataset uses the CIFData
class in cgcnn.data
. If you want a more flexible way to input crystal structures, PyTorch has a great Tutorial for writing your own dataset class.
Before training a new CGCNN model, you will need to:
- Define a customized dataset at
root_dir
to store the structure-property relations of interest.
Then, in directory cgcnn
, you can train a CGCNN model for your customized dataset by:
python main.py root_dir
You can set the number of training, validation, and test data with labels --train-size
, --val-size
, and --test-size
. Alternatively, you may use the flags --train-ratio
, --val-ratio
, --test-ratio
instead. Note that the ratio flags cannot be used with the size flags simultaneously. For instance, data/sample-regression
has 10 data points in total. You can train a model by:
python main.py --train-size 6 --val-size 2 --test-size 2 data/sample-regression
or alternatively
python main.py --train-ratio 0.6 --val-ratio 0.2 --test-ratio 0.2 data/sample-regression
You can also train a classification model with label --task classification
. For instance, you can use data/sample-classification
by:
python main.py --task classification --train-size 5 --val-size 2 --test-size 3 data/sample-classification
After training, you will get three files in cgcnn
directory.
model_best.pth.tar
: stores the CGCNN model with the best validation accuracy.checkpoint.pth.tar
: stores the CGCNN model at the last epoch.test_results.csv
: stores theID
, target value, and predicted value for each crystal in test set.
Before predicting the material properties, you will need to:
- Define a customized dataset at
root_dir
for all the crystal structures that you want to predict. - Obtain a pre-trained CGCNN model named
pre-trained.pth.tar
.
Then, in directory cgcnn
, you can predict the properties of the crystals in root_dir
:
python predict.py pre-trained.pth.tar root_dir
For instace, you can predict the formation energies of the crystals in data/sample-regression
:
python predict.py pre-trained/formation-energy-per-atom.pth.tar data/sample-regression
And you can also predict if the crystals in data/sample-classification
are metal (1) or semiconductors (0):
python predict.py pre-trained/semi-metal-classification.pth.tar data/sample-classification
Note that for classification, the predicted values in test_results.csv
is a probability between 0 and 1 that the crystal can be classified as 1 (metal in the above example).
After predicting, you will get one file in cgcnn
directory:
test_results.csv
: stores theID
, 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 inid_prop.csv
, which is not important.
To reproduce our paper, you can download the corresponding datasets following the instruction.
This software was primarily written by Tian Xie who was advised by Prof. Jeffrey Grossman.
CGCNN is released under the MIT License.