AUTHOR=Menon Aditya , Thompson-Colón James A. , Washburn Newell R. TITLE=Hierarchical Machine Learning Model for Mechanical Property Predictions of Polyurethane Elastomers From Small Datasets JOURNAL=Frontiers in Materials VOLUME=6 YEAR=2019 URL=https://www.frontiersin.org/journals/materials/articles/10.3389/fmats.2019.00087 DOI=10.3389/fmats.2019.00087 ISSN=2296-8016 ABSTRACT=

Polyurethanes are a broad class of material that finds application in coatings, foams, and solid elastomers. The urethane chemistry allows a diversity of monomers to be used, and prediction of mechanical properties, which are determined by complex interplay between monomer chemistry and chain architecture, is an unresolved challenge. Urethanes are based on aromatic or cyclic isocyanates and linear or branched polyols, and polymerization results in linear chains for bifunctional monomers or branched chains for multifunctional monomers. Strong intermolecular interactions between aromatic groups result in the formation of hard-segment domains that generate physical crosslinks between disorganized rubbery domains and anchor the material microstructure, contributing to resistance to deformation. Here, a general hierarchical machine learning (HML) model for predicting the stress-at-break, strain-at-break, and Tan δ for thermoplastic and thermoset polyurethanes is presented. The algorithm was trained on a library of 18 polymers with different diisocyanates, bifunctional or trifunctional polyols, and NCO:OH index. HML reduces data requirements through robust embedding of domain knowledge and surrogate data in a middle layer that bridges input variables (composition) and output responses (mechanical properties). In this work, the middle layer included information on overall polymer composition, predictions of chain architecture derived from Monte Carlo simulations of polymerization, information on interchain interactions from empirically derived molecular potentials and shifts in infrared (IR) spectroscopy absorbances. The HML predictions are shown to be more accurate than those from a random forest model directly relating composition and properties, suggesting that embedding domain knowledge provides significant advantages in predicting the properties of complex material systems based on small datasets.