Machine Learning Framework to Predict Glass Transition Temperature in Natural Deep Eutectic Solvents: A Step toward Green Functional Materials

Abstract

Natural Deep Eutectic Solvents (NADES) are a promising class of sustainable and environmentally-safe solvents with highly tunable physicochemical properties, including the glass transition temperature, which is critical for their functional performance, including ice control applications. Here, we present an interpretable machine learning (ML) framework to predict glass transition temperature (Tg) from the molecular structure of NADES combination, integrating descriptor-based feature engineering, unsupervised clustering, and ensemble regression. Combination of components and their mixing ratios for forming NADES were utilized to generate specific multi-component descriptors to describe NADES for ML modeling. A set of multicomponent descriptors was calculated based on individual descriptors from chemically diverse components of NADES. In result, a Random Forest (RF) model was developed to predict Tg values of NADES and the model achieved a very good performance with R² values in a range of 0.87-0.93, for both training and test sets. The analysis of contributing factors by Shapley Additive exPlanations (SHAP) analysis identified key features highlighting contributions of 3D geometry, atomic mass distribution and electronic effects. Finally, our results demonstrate that ML approaches combined with the mixture descriptors approach and interpretable modeling, enables accurate and chemically meaningful prediction of Tg, facilitating the rational design of NADES for applications in green chemistry and sustainable materials science applications.