TY - JOUR
T1 - Interpretable Tg prediction for polyurethane elastomers under restricted-distribution conditions
AU - Jin, Lu
AU - Wei, Zhiyan
AU - Zhuang, Zihan
AU - Shang, Jialu
AU - Sun, Shuping
AU - Xue, Dichang
AU - Fu, Debiao
AU - Zhu, Yindi
AU - Li, Xiaodong
AU - Jiang, Hao
AU - Zou, Meishuai
N1 - Publisher Copyright:
© 2026 Elsevier B.V.
PY - 2026/11/15
Y1 - 2026/11/15
N2 - Small and imbalanced datasets remain a critical challenge for property prediction in complex polymer systems. This issue is particularly pronounced in polyurethane elastomers (PUEs), where the glass transition temperature (Tg) is governed not only by chemical structure, but also by formulation composition, processing history, and testing protocol. In this work, we develop an interpretable machine-learning framework for Tg prediction in PUEs under small-sample and imbalanced-data conditions. A literature-curated dataset comprising 121 PUE samples was constructed by integrating chemical structures, formulation information, processing parameters, and testing methods reported over the past two decades. A hybrid feature space was established by combining RDKit-derived physicochemical descriptors with encoded experimental variables, together with type-aware preprocessing and collinearity control. Ablation analysis identified testing protocol as a first-class feature, while processing descriptors provided auxiliary information. A gradient boosting regression model was then developed, and systematic bias in the underrepresented high-Tg regime was mitigated through targeted data supplementation and weighted learning strategies. The optimized model achieved a test-set R2 of 0.904. SHAP and multidimensional partial dependence analyses provided post hoc insight into Tg-related model relationships across structure, composition, processing, and testing. Finally, prediction-guided experiments on representative formulations showed that the deviations between predicted and measured Tg values remained within 10%. These results provide a practical strategy for interpretable Tg prediction and formulation screening in PUEs, and demonstrate an effective approach for modeling small and imbalanced datasets in polymer materials.
AB - Small and imbalanced datasets remain a critical challenge for property prediction in complex polymer systems. This issue is particularly pronounced in polyurethane elastomers (PUEs), where the glass transition temperature (Tg) is governed not only by chemical structure, but also by formulation composition, processing history, and testing protocol. In this work, we develop an interpretable machine-learning framework for Tg prediction in PUEs under small-sample and imbalanced-data conditions. A literature-curated dataset comprising 121 PUE samples was constructed by integrating chemical structures, formulation information, processing parameters, and testing methods reported over the past two decades. A hybrid feature space was established by combining RDKit-derived physicochemical descriptors with encoded experimental variables, together with type-aware preprocessing and collinearity control. Ablation analysis identified testing protocol as a first-class feature, while processing descriptors provided auxiliary information. A gradient boosting regression model was then developed, and systematic bias in the underrepresented high-Tg regime was mitigated through targeted data supplementation and weighted learning strategies. The optimized model achieved a test-set R2 of 0.904. SHAP and multidimensional partial dependence analyses provided post hoc insight into Tg-related model relationships across structure, composition, processing, and testing. Finally, prediction-guided experiments on representative formulations showed that the deviations between predicted and measured Tg values remained within 10%. These results provide a practical strategy for interpretable Tg prediction and formulation screening in PUEs, and demonstrate an effective approach for modeling small and imbalanced datasets in polymer materials.
KW - Glass transition temperature
KW - Imbalanced data
KW - Interpretable machine learning
KW - Polyurethane elastomers
KW - SHAP analysis
KW - Small-sample learning
UR - https://www.scopus.com/pages/publications/105050215052
U2 - 10.1016/j.cej.2026.181677
DO - 10.1016/j.cej.2026.181677
M3 - Article
AN - SCOPUS:105050215052
SN - 1385-8947
VL - 548
JO - Chemical Engineering Journal
JF - Chemical Engineering Journal
M1 - 181677
ER -