qujiaxing / icsg3d

Official implementation of "3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning" (JCIM 2020)

Inorganic Crystal Structure Generation in 3D (ICSG3D)

A deep learning pipeline for generation of 3D crystal structures and prediction of their properties

---

All source code and images are associated with the paper:
C. J. Court, B. Yildirim, A. Jain, J. M. Cole,
"3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning",
J. Chem. Inf Model. (accepted for publication) (2020).
Paper HTML

---

Examples of generated structures

Representation learning of crystal structures

Architecture

Our pipeline consists of 3 components.

1. A Conditional Deep Feature Consistent Variational Autoencoder
2. A UNet semantic segmentation network
3. A Crystal Graph Neural Network

VAE

Encoder: 4x 3D convolutions, BatchNorm, ReLU and Maxpooling

Bottleneck: 3D convolution, LeakyReLU, Dense (256), 2x Dense (256) ( and )

Decoder: 4x 3D convolutions, BatchNorm, ReLU and upsampling

UNET

Downward: 4 x 2 x 3D convolutions, ReLU, BatchNorm, and pooling

Bottleneck: 2 x 3D convolutions, ReLU BatchNorm

Upward: 4 x 2 x 3D convolutions, ReLU, BatchNorm and UpSampling

CGCNN

Installation

1. Clone the git repository

   git clone https://github.com/by256/icsg3d

2. Install requirements

   python3 -m pip install -r requirements.txt

Requirements

Tensorflow2.0 is not currently supported, tested with tensorflow-gpu==2.1.0 and Keras==2.3.1 Requires keras-contrib

Getting Data

The system works on crystallographic information files (CIFs) to train the deep learning pipeline. In theory these can be from any source, but by default we use the materialsproject API.

For example, to retrieve all CIFs for cubic perovskites (ABX3):

python3 query_matproj.py --anonymous_formula="{'A': 1.0, 'B': 1.0, 'C':3.0}" --system=cubic --name=perovskites

This will create a data/perovskites folder containing the cifs and a csv with associated properties

Creating the network inputs

The various network input matrices can be created by

mpiexec -n 4 python3 create_matrices.py --name=perovskites

Train the UNET

Train the unet for as many epochs as needed

python3 train_unet.py --name perovskites --samples 1000 --epochs 50

Train the VAE

Make sure you train the VAE second (as it uses the unet as a DFC perceptual model)

python3 train_vae.py --name perovskites --samples 1000 --epochs 250

View some results

1. Interpolations in vae latent space

   python3 interpolate.py --name perovskites

2. Whole pipeline plots

   python3 view_results.py --name perovskites

3. Evaluate coordinates and lattice params

   python3 eval.py --name perovskites

Generate new samples

Attempt to generate 1000 new samples arund a base compound CeCrO3 with variance 0.5

python3 generate.py --name perovskites --nsamples 1000 --base CeCrO3 --var 0.5

This will create a new directory where you will find Cifs, density matrices, species matrices and properties for all generated compounds.

Using pre-trained models

The pre-trained models in the models directory. To use these, you'll need to move/rename the models to a corresponding directories:

saved_models/unet/mode/unet_weights_mode.h5 saved_models/unet/mode/class_weights.npy saved_models/vae/mode/vae_weights_mode.hdf5 saved_models/cgcnn/

where mode is the name of your data directory.

Citation

C. J. Court, B. Yildirim, A. Jain, J. M. Cole, "3-D Inorganic Crystal Structure Generation and Property Prediction via Representation Learning", J. Chem. Inf. Model. 2020 (accepted for publication).

Funding Statement

This project was financially supported by the Engineering and Physical Sciences Research Council (EPSRC, EP/L015552/1), Science and Technology Facilities Council (STFC) and the Royal Academy of Engineering (RCSRF1819\7\10).