Welcome to the repo for classifying crystal structures & space groups from 1D X-ray diffraction (XRD) patterns.

Can machine learning identify crystals in light diffraction patterns?
Check out our paper for more details and information, and be sure to cite us.

---

Paper:

@article{Crystals,
title   = {XRDs with deep learning (pending actual name)},
author  = {Jerardo Salgado; Sam Lerman; Zhaotong Du; Chenliang Xu; and Niaz Abdolrahim},
journal = {pre-print:Nature Communications},
year    = {2023}
}

---

☝️ Setup

1. Clone Current Project

Use git to download the XRDs repo:

git clone [email protected]:slerman12/XRDs.git

Change directory into the XRDs repo:

cd XRDs

2. Install UnifiedML

This project is built with the UnifiedML deep learning library/framework.

Download UnifiedML

git clone [email protected]:agi-init/UnifiedML.git

Install Dependencies

All dependencies can be installed via Conda:

conda env create --name ML --file=UnifiedML/Conda.yml

Activate Conda Environment

conda activate ML

ⓘ If your GPU doesn't support the latest CUDA version, you may need to redundantly install Pytorch with an older version of CUDA from pytorch.org/get-started after activating your Conda environment. For example, for CUDA 11.6:

pip uninstall torch torchvision torchaudio
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

ⓘ CUDA is needed to run the deep learning code on GPUs rather than CPUs. UnifiedML will automatically select GPUs when a working CUDA is available.

---

Reproducing The Work

To run, we have 3 model variants for predicting 7-way crystal types:

Model 1: No-pool CNN

python XRD.py task=NPCNN

Model 2: Standard CNN

python XRD.py task=SCNN

Model 3: MLP

python XRD.py task=MLP

💡 To predict 230-way space groups instead, add the num_classes=230 flag.

python XRD.py task=NPCNN num_classes=230

Plots automatically save to ./Benchmarking/<experiment>/.

The above scripts will launch training on the "souped" synthetic + random 50% RRUFF experimental data, & evaluation on the remaining 50% RRUFF data. The trained model is saved in a ./Checkpoints directory and can be loaded with the load=true flag.

All model and dataset code can be found in XRD.py

Custom datasets can be evaluated with the Dataset= flag and train_steps=0 load=true from a saved model.

Differences from and additions to paper

Synthetic data

This repo automatically downloads the public CIF database as opposed to ICSD as in the paper. If you’d rather use ICSD and have access, you can download it to the Data/Generated/CIFs_ICSD/ directory, and this code will automatically use that instead as in the paper. If you’d like to use both, add the open_access=true flag.

Souping and evaluation data

This GitHub provides the experimental real-world data RRUFF. It will be detected and used for souping as described in the paper. That is, reserving a random 50% subset of the real-world data for training and the remaining 50% for evaluation. If you’d like to disable souping, use the soup=false flag. If you’d like to train only on a 0.9/0.1 split of the synthetic data, you can use rruff=false.

---

Citing

If you find this work useful, be sure to cite us:

@article{Crystals,
title   = {XRDs with deep learning (pending actual name)},
author  = {Jerardo Salgado; Sam Lerman; Zhaotong Du; Chenliang Xu; and Niaz Abdolrahim},
journal = {pre-print:Nature Communications},
year    = {2023}
}

---

All UnifiedML features and syntax are supported.

This code is licensed under the MIT license.