PolyID: Artificial Intelligence for Discovering Performance-Advantaged and Sustainable Polymers

• DOI: 10.1021/acs.macromol.3c00994
• Publication Date: 2023-10-19T21:02:40Z
• Authors: A. Nolan Wilson; Peter C. St John; Daniela H. Marin; Caroline B. Hoyt; Erik G. Rognerud; Mark R. Nimlos; Robin M. Cywar; Nicholas A. Rorrer; Kevin M. Shebek; Linda J. Broadbelt; Gregg T. Beckham; Michael F. Crowley
• Keywords: Chemical space

A necessary transformation for a sustainable economy is the transition from fossil-derived plastics to polymers derived biomass and waste resources. While renewable feedstocks can enhance material performance through unique chemical moieties, probing vast design space by experiment alone not practically feasible. Here, we develop machine-learning-based tool, PolyID, reduce of enable efficient discovery performance-advantaged, biobased polymers. PolyID multioutput, graph neural network specifically designed increase accuracy quantitative structure-property relationship (QSPR) analysis It includes novel domain-of-validity method that was developed applied demonstrate how gaps in training data be filled improve accuracy. The model benchmarked with both 20% held-out subset original 22 experimentally synthesized mean absolute error glass temperatures 19.8 26.4 °C achieved test experimental sets, respectively. Predictions were made on composed monomers four databases contain biologically accessible small molecules: MetaCyc, MINEs, KEGG, BiGG. From 1.4 × 106 polymers, identified five poly(ethylene terephthalate) (PET) analogues predicted improvements thermal transport performance. Experimental validation one PET demonstrated temperature between 85 112 °C, which higher than within range tool. In addition accurate predictions, show model's predictions are explainable individual bond importance nylon. Overall, aid polymer practitioner navigate number discover materials enhanced