Directed Message Passing Neural Networks for Accurate Prediction of Polymer–Solvent Interaction Parameters

Accurate prediction of polymer–solvent interactions is essential for applications such as polymer processing, drug delivery, and membrane separations. The Flory–Huggins interaction parameter (χ parameter) serves as a key descriptor for polymer–solvent compatibility; however, its experimental determination is often costly and time-consuming. In this study, we develop a machine learning framework based on directed message passing neural networks (D-MPNNs) to predict χ parameters directly from molecular structures, temperature, and volume fraction. Our approach systematically evaluates different feature representations, pooling methods, and empirical equation integrations to optimize prediction accuracy. Among the tested models, D-MPNN-TC, which incorporates both temperature and volume fraction, demonstrates strong predictive performance (MAE = 0.092, RMSE = 0.162, and R2 = 0.926), outperforming descriptor-based models that rely on precomputed molecular fingerprints and handcrafted chemical features. Additionally, integrating the Flory–Huggins equation into the classification framework enables highly accurate miscibility predictions, with F1 scores of 0.915. Further analysis using t-SNE visualization reveals that D-MPNNs effectively capture key structural features, such as aromaticity and cyclic structures, that influence polymer–solvent interactions. Our findings highlight the advantages of graph-based molecular representations over traditional fingerprinting methods and underscore the importance of volume fraction information in predicting polymer–solvent compatibility. This study provides a scalable and interpretable framework for leveraging machine learning in polymer science, facilitating data-driven solvent selection and polymer design.