Phase Diagram Predictions of Various Polymer Macromolecules in Solution Using Transfer Learning

Polymer solution processing requires an in-depth understanding of liquid-liquid phase behavior to design polymer materials with optimal performance. Most models, such as Flory-Huggins and extensions thereof, provide physics insight into macro-phase separation but require complex empirical formulas that are fit to experimental data. Our past work shows that purely data-driven machine learning (ML) models, as well as theory-informed ML models, can accurately predict cloud point curves for many common linear polymer solutions. This work aims to extend the former framework for ML and physics-interpretable models to that of other polymer architectures such as star-shaped and polymer-grafted nanoparticle (PGN) solutions. However, due to limited data availability for stars and PGNs, our ML models show poor interpolation and extrapolation. Therefore, we develop a generalized feature vector that can incorporate linear polymer solution data to allow transfer learning of the linear polymer phase behavior to other polymer architectures. For the simplest feature vector, we show a decrease in the prediction error when linear polymer data is included, as well as better interpolation with less training data due to the model learning linear data behavior. We then discuss other methods of encoding polymer architecture and their effect on model interpolation and extrapolation.