Known Unknowns: Out-of-Distribution Property Prediction in Materials and Molecules

Code for extrapolation in materials property prediction as proposed in Known Unknowns: Out-of-Distribution Property Prediction in Materials and Molecules.

Setup

Clone the repository

git clone https://github.com/learningmatter-mit/matex.git

Create and activate a virtual environment

conda create -n blt-matex python=3.9.16
conda activate blt-matex

Install requirements

pip install -r requirements.txt

Environment Setup

Update hyperparameters in blt/configs/materials.yml. Run the following command where path_to_dir is the parent directory of matex.

export PYTHONPATH="${PYTHONPATH}:path_to_dir"

Data

Run the following script to process the data. Raw data is provided in blt/data. Processed data will be saved under blt/data as pkl files.

bash data_modules/create_data.sh

Training and Evaluation

Run the following script to train, evaluate and save the model

cd blt
bash train_eval.sh

Run the following script to create and save distribution and correlation plots

python plot_maker/plots.py

Cite

If you use this code in your research, please consider citing

@inproceedings{segal2024known,
  title={Known Unknowns: Out-of-Distribution Property Prediction in Materials and Molecules},
  author={Segal, Nofit and Netanyahu, Aviv and Greenman, Kevin and Agrawal, Pulkit and Gómez-Bombarelli, Rafael},
  booktitle={Workshop on AI for Accelerated Materials Design at Advances in Neural Information Processing Systems},
  year={2024}
}

Acknowledgements

The implementation is derived from Bilinear Transduction. The datasets are derived from AFLOW, Matbench, Materials Project, and MoleculeNet. The data processing and feature extraction are derived from Can machine learning find extraordinary materials?, Modnet and Deepchem.