Leveraging representations extracted from language models pretrained on material science literature for material discovery and property prediction.
🚧 Code still under construction 🚧
- To install the dependencies via Anaconda:
- Clone the repo:
git clone https://github.com/ertekin-research-group/Material_Recommender.git - Create conda virtual env:
conda env create -f environment.yml - Activate virtual env:
conda activate matrec
-
Download composition and structure embeddings: The embeddings for 116K materials obtained in this work can be found here
composition_embeddings_116k.h5: embeddings on material compositions.structure_embeddings_116k.h5: sentence embeddings on automatically generated material descriptions. Place the downloaded files under the main directory. -
Download pretrained weights: Follow the instructions on MatBERT repo to download pretrained weights and tokenizer for the uncased model. Place the folder under matbert_model_files directory.
- Search material candidates in the representation space.
📖TODO - Ranking candidates for materials with similar TE performance.
📖TODO - Training MMoeE models on material representations.
📖TODO
@misc{qu2023leveraging,
title={Leveraging Language Representation for Material Recommendation, Ranking, and Exploration},
author={Jiaxing Qu and Yuxuan Richard Xie and Elif Ertekin},
year={2023},
eprint={2305.01101},
archivePrefix={arXiv},
primaryClass={cond-mat.mtrl-sci}
}