Machine Learning-Driven Prediction and Interpretation of Glass Transition Temperature in Polyurethanes

The rational design of polyurethanes (PUs) with tailored
thermal
properties is often impeded by the complex interplay between molecular
architecture and microphase-separated morphology. While machine learning
(ML) offers a promising method to accelerate materials discovery,
existing models often lack mechanistic interpretability. Herein, we
develop a robust predictive model for the glass transition temperature
(Tg) by benchmarking a suite of ML algorithms
against a carefully curated and expanded PU data set. Among the evaluated
architectures, ensemble-based methods demonstrate better performance,
with the optimized eXtreme Gradient Boosting (XGBoost) model achieving R2 > 0.91 and RMSE ≈ 9.6 K, outperforming
Lasso regression, Random Forest and Support-Vector Regression. Beyond
accurate prediction, Shapley Additive Explanations (SHAP) methodology
is employed to elucidate the mechanistic underpinnings of feature
contributions in determining Tg. The SHAP
analysis has been shown to quantitatively identify the density of
quaternary carbon units within the hard segment (c HS), the hard segment
weight fraction (HS wt %) and the molecular volume of the hard segment
(HS MV) as the three dominant physicochemical descriptors governing Tg. Furthermore, all of the significant features
that were examined can be categorized into three distinct groups,
i.e., positive, negative, and dual effect groups, reflecting the competition
between intramolecular chain stiffening and the disruption of intermolecular
packing and the hydrogen-bond network. Overall, this study establishes
concise data-driven design rules for modulating the thermal behavior
of PUs, bridging the gap between statistical learning and fundamental
polymer physics and enabling more informed accelerated PU formulation.