Elem-XtalFormer (EXF)

This implements the Elem-XtalFormer (EXF) model to predict crystal symmetry space groups from X-ray diffraction and chemical composition data

Installation

• Clone the repository
• Use conda to install the environment by running:

conda env create -f env.yml

• Activate the environment by running:

conda activate exf

Dataset

We use the dataset curated by Xie. et al

Run the data_prep.ipynb notebook after downloading the relevant crystal structure data to create the required supplementary files

Model

Training Elem2Vec

The elem2vec.ipynb notebook contains the necessary code to train and save the element composition embeddings necessary for the main EXF model

Training EXF

The exf.ipynb notebook contains a simple example walkthrough on training EXF from scratch

data_utils.py: contains the PyTorch dataset class to load data for training and X-ray diffraction classes for experimental simulation of data

transformer_models: contains EXF model architechture classes

Experiments

regularization.py: contains experiments related to L1/L2 regularization

eval.ipynb & wo_comp_eval.ipynb: contain experiments related to noise levels and the composition module