Multiscale machine learning strategies for designing high-performance indaceno-bithiophene-based polymers

A machine learning (<i>ML</i>) assisted reverse polymer engineering approach was employed to design indaceno-bithiophene polymers with high glass transition temperature (<i>T_g</i>). A dataset of 482 indaceno-bithiophene-based polymers was collected and used to train several <i>ML</i> models, including random forest, decision tree, AdaBoost, and gradient boosting. The random forest model exhibited the highest performance with an <i>R</i>² value of 0.97 and a mean squared error (MSE) of 19.69. Feature importance analysis revealed that heavy atoms, Chi1, Chi0, LabuteASA, Chi0v, and estate_VSA8 were the most important descriptors contributing to the predicted <i>T_g</i> values. Using the trained random forest model, 306 new indaceno-bithiophene polymers were designed with predicted <i>T_g</i> values ranging from 224-667. The highest predicted <i>T_g</i> value of 667 was achieved, indicating that the designed polymers possess high thermal stability. The synthetic accessibility of the designed polymers was also evaluated, with scores ranging from 0.001 to 0.20. This work demonstrates the potential of <i>ML</i>-assisted reverse polymer engineering in designing high-performance polymers with tailored properties. The approach can be applied to other polymer systems, facilitating the rapid development of novel materials with improved thermal stability.