Accelerating Polyester Intelligence: Machine-Learning-Assisted Prediction of Glass Transition Temperature and Virtual Molecules Screening

Rapid development of the economy and society has resulted
in a
need for polyesters that are tailored to diverse performance requirements.
Unfortunately, the innovation of polyester materials is mainly dependent
on experience and intuitive guidance. Herein, we propose various interpretable
quantitative-structure–property relationship (QSPR) models
based on machine-learning-assisted approaches, which can accurately
predict polyesters’ glass transition temperatures (Tg) and facilitate the exploration of novel polyesters.
Initially, 695 polyesters with Tg values
are collected to establish multiple QSPR models using three different
algorithms, which undergo both internal and external validation. The
relative coefficient (R2) values of the
best deep neural network (DNN) model on the training set and testing
set reach 0.9588 and 0.9314, respectively, which is among the better
levels in related studies. The use of Morgan fingerprint with frequency
(MFF) descriptors and associated Shapley Additive Explanations analysis
does reveal a couple of interesting physical trends associated with
variation of Tg with the substructure
beyond what was reported before. To better widen the chemical space
of the existing polyester material family, a virtual polyester library
is constructed using a retrosynthetic strategy. Furthermore, this
workflow identifies 20 novel polyesters with low synthetic complexity
by high-throughput screening and validates these polyesters through
molecular dynamics simulations, which show an average absolute error
of 9.42 °C between the model-predicted and MD-simulated values.
Machine-learning-assisted approach not only improves the efficiency
of polyester material discovery but also provides a promising perspective
for understanding the thermal properties of polyesters from a microscopic
chemical structural viewpoint.