Hierarchical Machine Learning Model for Mechanical Property Predictions of Polyurethane Elastomers from Small Datasets

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