Word embedding vectors for materials, learned from Robocrystallographer descriptions

This code generates structure based feature vectors derived from Robocrystallographer text descriptions of crystal structures using word embeddings. It uses mat2vec to learn these embeddings from Robocrystallographer descriptions for all the elements in periodic table.

Table of Contents

• Installation
• Reproduction of publication results
• Publications/How to cite
• Mainteiners

Installation and basic usage

Clone or download this GitHub repository

Do any one of the following:

• Clone this repository to a directory of your choice on your computer.
• Download an archive of this repository and extract it to a directory of your choice on your computer.

Install dependencies

• Make sure you have the pip module installed.
• Navigate to the root folder of this repository and run pip install --ignore-installed -r requirements.txt. Note: If you are using a conda environment and any packages fail to compile during this step, you may need to first install those packages separately with conda install package_name.
• You are ready to go!

Reproduction of publication results

• To get robocrystallographer description for stable materials from materials project, navigate to the directory get_embeddings and run mpid_to_robo_descriptions.py.
• You can train mat2vec on your downloaded corpus of robocrystallographer description and generate word embeddings for all the elements of periodic table. get_embeddings.py in get_embeddings directory contains directions for that.
• To download matbench dataset navigate to training directory and run get_matbench_data.py.
• You can train ML models on matbench tasks while featurizing the data using our word embedding vectors by running train_models.py from the directory training. We used composition based feature vector technique for the featurization. You need to keep the .csv file that contains word embedding in the directory training/cbfv/cbfv/element_properties.
• To plot figures as shown in the paper run get_plots.py in the same directory. This will save the figures in the directory results.

Publications/How to cite

Mainteiners

hasan-sayeed (main maintainer)