Predicting Glass Transition Temperature of Polyimide Homopolymers and Copolymers Using Interpretable QSPR-Based Machine Learning: The MFKit Framework
- 1Bauman Moscow State Technical University, 2S1 Baumanskaya St., 5/1, Moscow, Russia, 105005
Abstract
In this work, we address the problem of predicting the glass transition temperature of polyimides, a class of high-performance engineering thermoplastics, with an emphasis on both homopolymer and copolymer systems. While machine learning approaches have been widely applied to polymer property prediction, their extension to copolymers remains limited and often lacks the interpretability required for rational materials design. To overcome these limitations, we develop an interpretable QSPR-based framework (MFKit) that combines physically grounded descriptors for monomer featurization with Gaussian Process Regression (GPR). The model is trained on a dataset of polyimide homopolymers from PoLyInfo and achieves performance comparable to state-of-the-art approaches (R2 = 0.856, MAE = 20.7 °C on the test set). Causal inference and SHAP analysis reveal key structure–property relationships, identifying aromatic content, backbone flexibility and rotation, side chain volume and hydrogen bonding as dominant factors governing Tg. The framework is subsequently extended to polyimide copolymers by constructing copolymer representations through combination of monomer descriptor vectors using several aggregation schemes. The best-performing approach yields R2 = 0.832 and MAE = 23.3 °C on an external dataset compiled from the literature. Furthermore, guided by the identified structure–property relationships, polyimide copolymers are designed, synthesized, and experimentally characterized, each exhibiting increased Tg compared with the parent homopolymer. Validation against DSC measurements demonstrates strong agreement with predictions (MAE = 16.3 °C). Overall, the proposed approach provides a unified and interpretable framework for Tg prediction for both homopolymer and copolymer systems, enabling physically grounded and data-driven design of polyimide-based materials.
Related
- Interpretable Machine Learning-Based QSPR Modeling for Predicting Hydrogen Adsorption Performance of Ester Oil Impregnating Agents
- Predicting Blood–Brain Barrier Permeability from Experimental Data: An Interpretable and Externally Validated Machine Learning Framework
- Interpretable Machine Learning for Predicting Cytochrome P450 Mechanism-Based Inhibition
- A Physics-Informed TCN-LSTM Approach for Interpretable Internal Temperature Estimation of NMC811 Lithium-Ion Batteries
- Predicting room temperature phosphorescence of new organic molecules by combining physical models and data-driven methods
BibTeX
@misc{averochkin2026predicting,
title = {{Predicting Glass Transition Temperature of Polyimide Homopolymers and Copolymers Using Interpretable QSPR-Based Machine Learning: The MFKit Framework}},
author = {Gleb Averochkin and Yulia Kirpal and Alexander Khina and Denis Bulkatov and Semen Fedorov and Elena O. Platonova and Ilya V. Tretyakov and Ivan P. Storozhuk and David G. Melikyants and Eugeny V. Alexandrov},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010512/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010512/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Predicting Glass Transition Temperature of Polyimide Homopolymers and Copolymers Using Interpretable QSPR-Based Machine Learning: The MFKit Framework
AU - Gleb Averochkin
AU - Yulia Kirpal
AU - Alexander Khina
AU - Denis Bulkatov
AU - Semen Fedorov
AU - Elena O. Platonova
AU - Ilya V. Tretyakov
AU - Ivan P. Storozhuk
AU - David G. Melikyants
AU - Eugeny V. Alexandrov
PY - 2026
DA - 2026/10/15
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010512/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010512/v1
AB - In this work, we address the problem of predicting the glass transition temperature of polyimides, a class of high-performance engineering thermoplastics, with an emphasis on both homopolymer and copolymer systems. While machine learning approaches have been widely applied to polymer property prediction, their extension to copolymers remains limited and often lacks the interpretability required for rational materials design. To overcome these limitations, we develop an interpretable QSPR-based framework (MFKit) that combines physically grounded descriptors for monomer featurization with Gaussian Process Regression (GPR). The model is trained on a dataset of polyimide homopolymers from PoLyInfo and achieves performance comparable to state-of-the-art approaches (R2 = 0.856, MAE = 20.7 °C on the test set). Causal inference and SHAP analysis reveal key structure–property relationships, identifying aromatic content, backbone flexibility and rotation, side chain volume and hydrogen bonding as dominant factors governing Tg. The framework is subsequently extended to polyimide copolymers by constructing copolymer representations through combination of monomer descriptor vectors using several aggregation schemes. The best-performing approach yields R2 = 0.832 and MAE = 23.3 °C on an external dataset compiled from the literature. Furthermore, guided by the identified structure–property relationships, polyimide copolymers are designed, synthesized, and experimentally characterized, each exhibiting increased Tg compared with the parent homopolymer. Validation against DSC measurements demonstrates strong agreement with predictions (MAE = 16.3 °C). Overall, the proposed approach provides a unified and interpretable framework for Tg prediction for both homopolymer and copolymer systems, enabling physically grounded and data-driven design of polyimide-based materials.
ER -