Multiscale Modeling of Polyurethane Properties via Latent Variables with Hierarchical Machine Learning

The objective of the following dissertation was to develop Hierarchical Machine Learning models to accurately predict bulk properties of polyurethanes with domain knowledge from small datasets. Polyurethanes find application in coatings, foams, and solid elastomers but the range of processing conditions and the diversity of monomers results in limited datasets in practice and as a result statistical modeling has not been pursued extensively in place of high-throughput experiments, bias brute force modeling, or assumptive analytical treatments. By introducing experimental and computational data describing the multiscale forces in polyurethanes with Hierarchical Machine Learning, interpretable and high predictive-strength models were studied. Four major property types were the subject of modeling—thermal, rheological, mechanical, and failure. Each has unique hypotheses as to what underlying physical trend accurately predicts the property. Testing was performed on withheld sets of training data and secondary validation sets of unobserved chemical formulations were used to further analyze the generalizability of prediction. Through model performance and feature selection/analysis, material and property design principles could be extracted from model outputs.