Random Forest Based Approach for Predictions of Glass Transition Temperatures in Polymers

The glass transition temperature ( T g ) is a critical parameter in determining polymer properties. In this paper, a quantitative structure–property relationship (QSPR) model was developed for predicting T g s using the random forest (RF) algorithm on a large dataset of 1320 polymers. A subset of descriptors comprising 20 Dragon descriptors and 13 quantum chemical descriptors was obtained from the repeating units of polymers for developing QSPR models of T g s. RF parameter combinations formed 192 models, from which an optimal model (400 trees, 16 features/split) was selected for T g prediction, by performing 10× repeated 10‐fold cross‐validation on the training set (1056 polymers). This model displays excellent predictive performance (test set R 2 = 0.858, RMSE = 27.45 K). The mechanistic analysis revealed that rigidity factors (e.g., aromatic rings via SpMax_B(v), molecular compactness (PJI2), and dipole moment (μ)) elevate T g by limiting chain mobility. Conversely, flexible elements like rotatable bonds (RBF) and bulky side groups (nR = Ct) lower T g . Electronic effects further modulate T g , with high E HOMO increasing rigidity and polarizable moieties (MATS7p) enhancing flexibility. The model excels in its broad applicability domain, high predictive accuracy, and interpretable structure–property relationships.