Predicting glass transition temperatures with polymer chain segments

This work, for the first time, employs polymer chain segments with similar number of atoms to calculate 635 molecular descriptors for corresponding polymers, from which a subset of descriptors was selected as inputs for establishing random forest (RF) models for the glass transition temperatures (<i>T</i>_gs) of 315 polymers with diverse structures. The optimal RF model (RF Model I) achieved determination coefficient <i>R</i>² and root-mean-square (rms) error of 0.98 and 18.1 K, respectively, for the training set (236 polymers), and of 0.94 and 30.9 K, respectively, for the test set (79 polymers), which are accurate compared with quantitative structure–property relationship (QSPR) models for <i>T</i>_gs reported in the literature. Mechanism analysis shows that increasing rotatable bond fraction by introducing –Si–O–, –Si–, –C– or –O– flexible segments and topological distance among atoms in the backbone (or side) chain can, respectively, result in high flexibility of the molecular chain and large free volume between molecules, and then bring down <i>T</i>_gs. Conversely, increasing the number of imides, aromatic bonds, delocalised bonds, or topological polar surface area by introducing N, O, S, and P atoms can enhance chain rigidity, leading to elevated <i>T</i>_gs.