Graph-Based Machine Learning for Multicomponent Polymer Solution Phase Diagram Predictions

The prediction of multicomponent polymer phase diagrams is pertinent to the processing, self-assembly, and purification of polymer materials, allowing precise control in the design and morphology of polymer films, coatings, and composites. Previous work has focused on accelerating the prediction of experimental polymer solution cloud point temperatures using machine learning and Flory-Huggins theory but have been restricted to binary polymer solutions and linear polymer architectures. Recently, we have implemented a graph-based, transfer learning approach to account for various polymer architectures. Here, we present a graph-based approach that can be used to generalize the featurization of a multicomponent mixture to predict phase boundaries for ternary polymer solution mixtures. We show how a graph neural network (GNN) approach can allow for the inclusion of binary polymer solution data to improve the prediction for ternary phase diagrams. Specifically, we find that the GNN benefits the most from the inclusion of binary data to predict out-of-domain polymers in the ternary mixture and lowers the prediction error slightly when predicting new phase boundaries at new temperatures. Lastly, we discuss the introduction of two novel batch selection algorithms for active learning to accelerate the mapping of unseen ternary phase diagrams, demonstrating more efficient mapping than traditional variance-based selection algorithms.