Polymer Solubility Prediction Using Large Language Models

Traditional approaches in polymer informatics often require labor-intensive data curation, time-consuming preprocessing such as fingerprinting, and choosing suitable learning algorithms. Large language models (LLMs) represent a compelling alternative by addressing these limitations with their inherent flexibility, ease of use, and scalability. In this study, we propose a novel approach utilizing fine-tuned LLMs to classify solvents and nonsolvents for polymers, a property critical to polymer synthesis, purification, and diverse applications. Our results show that fine-tuned GPT-3.5 achieves predictive performance comparable to or exceeding traditional machine learning methods, even with limited data sets. The model achieved predictive accuracies of 0.90 and 0.83 for identifying soluble and insoluble solvent–polymer pairs, respectively. Remarkably, these models accurately classify solvents and nonsolvents in entirely unseen scenarios, indicating that they are able to effectively leverage the components embedded in their base models. The operational simplicity and accuracy of LLMs highlight their potential for advancing polymer research.