Interpretable Machine Learning of Phase Separated Microstructures in Polyurethane Block Copolymers

We employ interpretable machine learning as a tool for investigating the relationship between the chemical composition of a polyurethane block copolymer and the resulting morphology of phase separated microstructures. We carry out atomistic MD simulations with various choices of chemistry for soft segments, hard segments, and chain extenders, as well as varying the total molecular weight and molecular weight ratio between hard and soft segments. We use the temperature-dependent structure factor to characterize phase separation in our systems. We measure the Flory-Huggins interaction parameter χ to compare with a phase field model. We use the structure factor at small wave numbers as an objective function to determine interaction parameters for a coarse-grained bead-spring chain model to prioritize the accuracy of large-scale phase separated structures in the coarse-grained model. We investigate the use of interpretable machine learning to discover tractable relationships between polymer design parameters and phase separated structures beyond existing theories and empirical trends.