PolyID^TM provides a framework for building, training, and predicting polymer properities using graph neural networks. The codes leverages nfp, for building tensorflow-based message-passing neural networ, and m2p, for building polymer structures. The example notebooks demonstrate how to build polymer structures, train a message-passing neural network ensemble, and evaluate its predictions, following the methodology used in the publication.

1. Quick train and predict: examples/quick_train_and_predict.ipynb — an end-to-end example that loads a polymer dataset, splits it into train/holdout, trains a k-fold ensemble of message-passing neural networks, saves/loads the models (.keras format), and evaluates predictions on the holdout set.
2. Checking domain of validity: examples/domain_of_validity.ipynb — determine the domain of validity for a set of predictions.

For more details, see the manuscript PolyID: Artificial Intelligence for Discovering Performance-Advantaged and Sustainable Polymers, Macromolecules 2023.

Cite

If you use PolyID in your work, please cite

@article{wilson2023polyid,
  title={PolyID: Artificial Intelligence for Discovering Performance-Advantaged and Sustainable Polymers},
  author={Wilson, A Nolan and St John, Peter C and Marin, Daniela H and Hoyt, Caroline B and Rognerud, Erik G and Nimlos, Mark R and Cywar, Robin M and Rorrer, Nicholas A and Shebek, Kevin M and Broadbelt, Linda J and Beckham, Gregg T and Crowley, Michael F},
  journal={Macromolecules},
  volume={56},
  number={21},
  pages={8547--8557},
  year={2023},
  publisher={ACS Publications}
}