Machine learning of polyurethane prepolymer viscosity: a comparison of chemical and physicochemical approaches

Abstract

Polyurethane prepolymers are essential intermediates in the production of polyurethane foams, films, and elastomers, with viscosity playing a critical role in formulation, processing, and manufacturing. Despite its importance, there are no models that quantitatively predict the viscosity of a given polymer as a function of monomer chemistry. Traditional empirical models can effectively capture viscosity trends but often require extensive experimental datasets and provide limited interpretability, particularly when applied to novel formulations. Here, we explored regression options for representing polymer chemistry and for modeling the form of the temperature dependence. Monomers can be represented as a formulation in which they are labeled by monomer name or in a physicochemical framework where they are labeled by molecular characteristics derived from experimental measurements or computational methods. The overall form of the temperature-dependent viscosity can be modeled through a generic regressor function or by assuming the empirical form of the Andrade equation. A 39-sample training library was used to evaluate both approaches, with machine learning models achieving a coefficient of determination (r²) of 0.71 for the chemical model testing data and 0.86 in predicting the Andrade equation parameter, which provides interpretable access to the continuous viscosity-temperature curve, for previously untested compositions. While chemically defined models offer a direct path to high-accuracy predictions within known compositional spaces, physicochemical informed models provide deeper insight into structure–property relationships, facilitating extrapolation to novel materials. This work underscores the tradeoffs between empirical and physics-informed modeling strategies and offers a structured approach to integrating domain knowledge into predictive frameworks for complex material systems.