XIE-SPP: Crystal Image Encoder for Synthesis & Property Prediction

This code has been developed by Ali Davariashtiyani. (LinkedIn, Scholar)

This code has been developed under the support of the National Science Foundation (NSF) award number DMR-2119308.

1. Reference

Please cite the following paper if you use any part of our code or data:

Davariashtiyani, A., Kadkhodaie, Z. & Kadkhodaei, S. Predicting synthesizability of crystalline materials via deep learning. Commun Mater 2, 115 (2021). https://doi.org/10.1038/s43246-021-00219-x

The data files for this paper is available at https://uofi.app.box.com/s/7o39i8jt4ve7s5loo0twlub06k3d7np3

Ali Davariashtiyani, Sara Kadkhodaei. Formation energy prediction of crystalline compounds using deep convolutional network learning on voxel image representation. Commun Mater 2, 105 (2023). https://doi.org/10.1038/s43246-023-00433-9

The data files for this paper is available at https://uofi.box.com/s/ka69pgrpgn9gkauz912urcurj31xmlfm

2. Installation

Tested on Ubuntu 20.04

Navigating to the desired directory in the terminal:

cd [/navigate/to/directory/]

Cloning the repository and entering the directory:

git clone https://github.com/kadkhodaei-research-group/XIE-SPP.git
cd XIE-SPP/

We highly encourage you to use a conda envirenment. If you don't have conda installed on your machine, you can skip this step:

conda create -n synthesizability python=3.7.3 -y
conda activate synthesizability

Installing TensorFlow on Ubuntu:

***Installing the GPU version:
conda install -c anaconda keras-gpu -y
***Installing the CPU version:
conda install -c anaconda keras -y

Installing TensorFlow on Mac M1:

miniforge3 is required.

conda install -c apple tensorflow-deps -y
pip install tensorflow-macos
pip install tensorflow-metal

Installing the Crystal Image Encoder for Synthesis & Property Prediction (XIE-SPP) package and it's dependencies:

python setup.py install

3. Using the model

3.1 Using the model for synthesizability prediction (Version 1)

Loading the model:

from xiespp.synthesizability_1 import synthesizability

samples = synthesizability.get_test_samples('GaN')
samples

Output:

['XIE-SPP/finalized_results/explore_structures/cif/GaN/GaN_9.cif',
 'XIE-SPP/finalized_results/explore_structures/cif/GaN/GaN_12.cif']

Evaluation: (Input can be CIF files or ASE Atoms Objects)

synthesizability.synthesizability_predictor(samples)

The output is the synthesizability likelihood of the input list:

2/2 [==============================] - 2s 783ms/step
array([0.00200533, 0.9643494 ], dtype=float32)

3.2 Using the model for synthesizability prediction (Version 2)

We highly encourage the use of this version of the model. This version is more robust, accurate and faster than the previous version.

from xiespp import synthesizability, get_test_samples
# or 
# from xiespp import synthesizability_2 as synthesizability

samples = get_test_samples('CSi')
# The input to the model can in formats like: CIF, POSCAR, PyMatGen Structure, ASE Atoms
model = synthesizability.SynthesizabilityPredictor()
model.predict(samples)

3.3 Using the model for formation energy prediction

from xiespp import formation_energy, get_test_samples
model = formation_energy.FormationEnergyPredictor()

# As an example we take the samples from MP database
from mp_api.client import MPRester
api_key = '__YOUR_API_KEY__'
mpr = MPRester(api_key)
search = mpr.summary.search(chemsys="Si-O")
data_input = search[:10]

# As practice instead of directly passing the data to the model we first generate the Image Generator Object
gen = formation_energy.prepare_data(
    data_input=data_input, 
    # input_format= 'vasp', # Use this if the format of the file is not CIF file
)
model = formation_energy.FormationEnergyPredictor()
yp = model.predict(gen, return_all_ensembles=True)

4. Reproducibility

4.1 Unpacking the data files

Data files available at https://uofi.app.box.com/s/7o39i8jt4ve7s5loo0twlub06k3d7np3

Untarring the files to the data folder:

tar -xvzf Data.joined.tar.gz

4.2 Data availability

All the CIF files are collected from the Materials Project, COD and CSPD databases. The selected data fed to the model is available in the unwrapped data folder.

4.3 Preparing data from the scratch

All the used data is already prepared in the unwrapped folder. However, for reproducing everything again, follow the steps: Set the data folders in the config.py file.

Prepare the positive set: Obtain the COD:

mkdir data/data_bases/
wget http://www.crystallography.net/archives/cod-cifs-mysql.tgz
tar -xvzf cod-cifs-mysql.tgz

Follow the guideline to create the SQL file.

Run the entire positive_data_preparation.ipynb file.

According to the prepared guideline to prepare the CSPD data set and run anomaly_generation.ipynb to prepare the negative set
Create the image files by running data_set_selections.ipynb

4.4 Re-training the models

Re-training formation energy prediction: model7_training.ipynb

5. Crystal Voxel Representation (CVR) 3D images

Image creation package: CVR