Theory-Informed Machine Learning of Star and Polymer-Grafted Nanoparticle Solution Phase Behavior

The design and processing of novel polymer materials hinges on the quantitative prediction of their solution phase behavior, but significant challenges hinder progress. These include navigating vast parameter spaces such as polymer-solvent chemistry or polymer architecture (star, ring, comb, etc.), the limited accuracy of existing thermodynamic models, and the cost of high-fidelity simulations. Prior machine learning (ML) approaches have demonstrated that Flory-Huggins (FH) theory can be integrated to improve predictions of polymer solution cloud point temperatures, but lack generalizability to other chemistries and architectures. To address this, we demonstrate how thermodynamics-informed ML can predict the phase behavior for binary solutions of linear and star polymers, and polymer-grafted nanoparticles (PGNs). We leverage FH theory derived for star polymers and extend it to include PGNs. Our theory-informed models integrate this modified FH model by either embedding it into the loss function or using a neural network to predict specific interaction terms. We show these approaches are superior to pure ML approaches when extrapolating to new architectures and chemistries, reducing training data requirements, and discuss how this methodology can be extended to multicomponent systems.