Transfer learning with graph neural networks to predict polymer solubility parameters

Abstract

Solubility plays a vital role in materials research and applications, with Hansen Solubility Parameters (HSP) serving as valuable descriptors for guiding material design and processing. However, due to the experimental difficulty of obtaining polymer HSP values, available data remains extremely scarce, thereby limiting the development of generalizable machine learning models. Furthermore, most existing models rely solely on monomer level information, resulting in inadequate representations of polymer structures. To address these challenges, we propose a data-driven method that leverages transfer learning to alleviate the shortage of labeled polymer data by utilizing readily available solvent HSP annotations. In addition, we introduce the Polymer Property Trend Approximation (PPTA) algorithm, which enhances structural representations by dynamically incorporating information from generated oligomers and fragments. Our model, GT-PolySol, achieves state-of-the-art (SOTA) performance in polymer HSP prediction and provides interpretable insights into the structure-property relationship, while offering improved computational efficiency compared to existing methods.