Discovery of thermosetting polymers with low hygroscopicity, low thermal expansivity, and high modulus by machine learning

Abstract

Traditional material design principles for discovering thermosetting polymers, such as polycyanurates with a combination of low hygroscopicity, low thermal expansivity, and high modulus, are inefficient due to the intrinsic conflict among properties. Machine learning (ML) can be employed to overcome this conflict by exploring the vast polymer space in silico but it is limited by the lack of polymer data. This work demonstrates that through multi-fidelity learning, limited experimental data can be well utilized with the assistance of all-atomic simulation data for establishing robust ML-based property estimation models. We screen the virtual candidates obtained from combinations of “polymer genes” through well-trained ML-based models. Several promising polycyanurates are demonstrated to have remarkable properties, as verified by theoretical simulations and proof-of-concept experiments. Furthermore, structural analysis is performed to deduce the underlying physics. The proposed framework can be extended to accelerate the rational design of crosslinked polymers with desired properties and reveal fundamental physical rules governing the design.