autoXRD is a python package for automatic XRD pattern classification of thin-films, tweaked for small and class-imbalanced datasets. The main application of the package is high-throughput screening of novel materials.
autoXRD performs physics-informed data augmentation to solve the small data problem, implements a state-of-the-art a-CNN architecture and allows interpretation using Average Class Activation Maps (CAMs), according to the following publications:
"Oviedo, F., Ren, Z., Sun, S., Settens, C., Liu, Z., Hartono, N. T. P., ... & Buonassisi, T. (2019). Fast and interpretable classification of small X-ray diffraction datasets using data augmentation and deep neural networks. npj Computational Materials, 5(1), 60." Link: https://doi.org/10.1038/s41524-019-0196-x
"Fast classification of small X-ray diffraction datasets using data augmentation and deep neural networks, (2019), Felipe Oviedo, Zekun Ren, et. al. Link: arXiv:1811.08425v
Accepted to NeurIPS 2018 ML for Molecules and Materials Workshop. Final version published npj Computational Materials 2019
To install, just clone the following repository:
$ git clone https://github.com/PV-Lab/autoXRD.git
Just run space_group_a_CNN.py , with the given datasets. Note that this performs classification for patterns into 7 space-groups. Dimensionality data is not included in the code, please contact authors if interested.
The package contains the following module and scripts:
| Module | Description |
|---|---|
space_group_a_CNN.py |
Script for XRD space-group classification with a-CNN |
autoXRD |
Module dedicated to XRD pattern preprocessing and data augmentation |
autoXRD_vis |
Visualizer module for class activation maps (CAMs) |
Demo / XRD_dimensionality_demo.ipynb |
Notebook containing a demo for physics-informed data augmentation. This is a version with a modified CNN and no CAM to speed up the computation |
Felipe Oviedo and "Danny" Zekun Ren
| AUTHORS | Felipe Oviedo and "Danny" Ren Zekun |
| VERSION | 1.0 / May, 2019 |
| EMAIL OF REPO OWNER | foviedo@mit.edu |
This work is under an Apache 2.0 License and data policies of Nature Partner Journal Computational Materials. Please, acknowledge use of this work with the apropiate citation.
@article{oviedo2019fast,
title={Fast and interpretable classification of small X-ray diffraction datasets using data augmentation and deep neural networks},
author={Oviedo, Felipe and Ren, Zekun and Sun, Shijing and Settens, Charles and Liu, Zhe and Hartono, Noor Titan Putri and Ramasamy, Savitha and DeCost, Brian L and Tian, Siyu IP and Romano, Giuseppe and others},
journal={npj Computational Materials},
volume={5},
number={1},
pages={60},
year={2019},
publisher={Nature Publishing Group}}