Simple machine learning model for the glass transition temperatures of hydrated polymers

The glass transition temperature under dry conditions (Tg,dry) is an important factor for understanding the properties of polymeric materials; however, Tg,dry is often not a suitable factor for understanding the functions of biomaterials because these materials are used in water or are in contact with water. Therefore, the glass transition temperature under hydrated conditions (Tg,wet) is a crucial thermal property of polymers, particularly biomaterials. Traditional Tg,wet measurements require significant skill and are prone to error. To address this, we developed a machine learning (ML) model to predict Tg,wet from the polymer structures using a small dataset of 33 polymers. SMILES was used to generate Morgan fingerprints (MFPs) and SMILES-fragments (FLs), which serve as descriptors for the ML models. We used both random forest (RF) and ridge regression (RR) algorithms, and these algorithms were optimizing through grid search and cross-validation. The ML models using only chemical structure descriptors (MFP and FL) exhibited poor predictive performance and showed overfitting. However, when the values of Tg,dry were included as an explanatory variable, the RR model using MFP provided the best performance. These results highlight the importance of incorporating the data of Tg,dry to enhance the prediction of Tg,wet. Our model has the potential to facilitate the design of functional biomaterials. The glass transition temperature of the dry polymers (Tg,dry) is a useful parameter for predicting that of hydrated polymers (Tg,wet). By combining Tg,dry and chemical structures, simple machine learning models for Tg,wet can be constructed even with a small dataset.