CubicGAN:

HIGH-THROUGHPUT DISCOVERY OF CUBIC CRYSTAL MATERIALS USING DEEP GENERATIVE NEURAL NETWORKS

Created by Yong Zhao

Introduction

This repository contains the implementation of CubicGAN. You can visit Carolina Materials Database for more info.

Prerequisites

python 3.7
tensorflow-gpu==2.2.0
pymatgen==2020.3.13
numpy==1.16.4
pandas==0.25.1
scikit-learn==0.23.1

we recommend that you build a virtural environment running the code.

How to use the code

Step 1

Once you download the code to your machine, you need to make three folders to the project root directory. The three folders are logs, models, and generated_mat. The folder logs saves training loss of discriminator and generator. The folder models saves trained models. The folder generated_mat saves your generated materials information. We keep 100 cleaned CIFs in folder data/trn-cifs/, you can use them for trying this code. Just calling

python chain.py

can let you experience whole process. If you want to build your own training work, you might need below steps.

Generate materials with a pre-trained model

Under directory pre-trained/, we uploaded the ternary generator model. If you want use the pre-trained model, you can simply change the load model in generate_crystal.py and then you can directly start from Step 4 to generate samples.

Step 2

Before you start to run the code, you have to collect the training data by yourself from open databases (e.g., OQMD). You have to be sure that you convert the downloaded crystal strucutes to symmetrized CIFs and select materials only with atom position of multiplicative factor of 0.25. Then, you can replace the given CIFs in folder data/trn-cifs/ and run

python extract_crystal.py

Note: We recommend that you change the space group ratio in util.py if you want your generated mateirals to follow the similar space group ratio as training data.

Step 3

Train the generator model by runing

python wgan-v2.py

The model will be saved in the folder models.

Step 4

Generate materials by runing

python generate_crystal.py --n_samples=1000

Argument n_samples controls how many materials you will generate and the data will be saved under directory generated_mat/sample-1000.

Step 5

Runing

python pymatgen_valid.py --n_samples=1000

will generate two folders under directory generated_mat/sample-1000. The folder tmp-charge-cifs saves the valid CIFs recognized by pymatgen. The folder tmp-charge-cifs saves the crystal structures with zero charge.

Step 6

For further analysis, run

python build_repeating_cifs_stru.py --n_samples=1000

and

python build_unique_records.py --n_samples=1000

which build a tree structure for repeating CIFs and unique CIFs csv file under directory generated_mat/sample-1000.

Training data used in our work

In file oqmd-dat.csv, we show the IDs and other information of ternary materials in OQMD.

Total	Fm-3m	F-43m	Pm-3m	unique formula
375749	186344	184162	5243	249646

License

CubicGAN is released under the MIT License.