Periodic Set Transformer

This is a repository for the implementation of the Periodic Set Transformer. The starter code was taken from CGCNN: https://github.com/txie-93/cgcnn. The general usage is largely the same as their model.

File Descriptions

• main.py: Main file for command line utility
• model.py: Contains the full model implementation for the PST
• matbench_parameters.py Defines the parameters used in the Matbench experiments
• mb.py: Main file for running the Matbench test suite
• data.py: Defines all PyTorch datasets used in main.py and mb.py
• train.py: Implements the training and validation cycles
• plots.py: Defines the functions for creating the plots in the article
• gpr.py: The Gaussian Process Regression implementation used in the experiments
• pdd_helpers.py - Helper functions for creating PDDs from CIFs
• atom_init.json: Defines CGCNN atom features in a one-hot encoded manner
• mat2vec.csv: Defined Mat2Vec atom features
• mf: This folder contains the Jarvis IDs that were in the train, validation and test set when Matformer was run
• jarvis_dft_2d_2021_pymatgen_structures.zip: Zip file containing the data to run the Jarvis-DFT dataset

General Usage

First, install the dependencies via pip install -r requirements.txt.

To run the model there needs to be data folder containing the following:

1. A set of [CIF] files with the following names convention id.cif
2. A [CSV] files named id_prop.csv with two columns: the first is the name of the corresponding [CIF] filename without the extension and the second is the property value

Please make sure that the atom_init.json file is in the current directory along with main.py.

The model can then be run by using the following:

python main.py data

A number of model options are listed in main.py that can be adjusted as needed. For example:

python main.py --epochs=200 data

Matbench

In order to run the test suite for Matbench please run the following: pip install -r requirements.txt to install the necessary packages, then python mb.py.

The results.tar.gz file will appear in the same directory.

Jarvis-DFT

To run the Jarvis-DFT dataset first, unzip the data file jarvis_dft_3d_pymatgen_structure.zip. The resulting binary will automatically be read in after running python run_jarvis.py.