Atomic Hirshfeld Surface Fingerprint

This software package takes an arbitary crystal structure to predict material properties.

Table of Contents

• Prerequisites
• How to Use
• How to Cite
• Data
• Authors
• License

Prerequisites

• Pandas
• Keras (with Tensorflow backend)
• Scikit-Learn
• Seaborn (for visualization)

How to use:

TODO

Training the model

Using a pre-trained model:

Download the package using the following code:

git clone https://github.com/arpanisi/mof_single_atom_hs.git

Use conda to create an environment as:

conda create -n ahsfp python=3.7 scikit-learn keras tensorflow

This creates a conda environment along with installing the prerequisites. Activate the environemnt by:

conda activate ahsfp

Navigate to the folder /ahsfp and type:

(ahsfp) working_directory/ahsfp$ python predict.py

Alternately it can be used by one of the following codes:

python predict.py --parameter old_lattice and python predict.py --parameter new_lattice. The output of both the commands give the following output

How to cite

TODO

Data

TODO

Architecture

The complete network architecture of ahsfp is given below:

Name	Output Shapes	Parameters
InputLayer	(None, 50, 50, 1)	0
Conv2D	(None, 50, 50, 32)	320
Conv2D	(None, 50, 50, 32)	9248
MaxPooling2D	(None, 25, 25, 32)	0
Dropout	(None, 25, 25, 32)	0
BatchNormalization	(None, 25, 25, 32)	128
Conv2D	(None, 25, 25, 64)	18496
Conv2D	(None, 25, 25, 64)	36928
Conv2D	(None, 25, 25, 64)	36928
MaxPooling2D	(None, 12, 12, 64)	0
Dropout	(None, 12, 12, 64)	0
BatchNormalization	(None, 12, 12, 64)	256
Conv2D	(None, 12, 12, 128)	73856
Conv2D	(None, 12, 12, 128)	147584
Conv2D	(None, 12, 12, 128)	147584
Conv2D	(None, 12, 12, 128)	147584
MaxPooling2D	(None, 6, 6, 128)	0
BatchNormalization	(None, 6, 6, 128)	512
Flatten	(None, 4608)	0
Dense	(None, 128)	589952
Dropout	(None, 128)	0
BatchNormalization	(None, 128)	512
Dense	(None, 128)	16512
Dropout	(None, 128)	0
BatchNormalization	(None, 128)	512
Dense	(None, 1)	129

Authors

This software is written by Arpan Mukherjee. Data Collection and analysis by Logan Williams. Arpan and Logan were advised by Prof. Krishna Rajan

Please use the following article to cite our work:

Williams, Logan, Arpan Mukherjee, and Krishna Rajan. "Deep Learning Based Prediction of Perovskite Lattice Parameters from Hirshfeld Surface Fingerprints." The Journal of Physical Chemistry Letters 11.17 (2020): 7462-7468.

License

released under the MIT License