Design of polyimides with targeted glass transition temperature using a graph neural network

Abstract

Polyimide substrates used in flexible display devices need to withstand very high temperatures and be highly thermally stable. The discovery of polyimides that satisfy these requirements, especially with high glass transition temperatures (T_g), has been a hot research topic. However, conventional structural design of polyimides requires domain knowledge and expertise. In addition, the latent polyimide structures can reach millions, making it impractical to navigate their structure–property landscapes under conventional experimental screening. To address this matter, we utilized a graph neural network to build a regression model for predicting polyimides’ T_g and a classification model for two-step ensemble molecular screening. Eight polyimides were synthesized with T_g measured using differential scanning calorimetry (DSC) and these values were well matched to the predicted values of the regression model. Unlike common machine learning models, our models substantially learn the effect of specific functional groups on polyimides T_g and thus offer visual interpretation of the effect of functional group variations on T_g, providing an unknown performance metric (Rot2Ring) for designing molecules with specific T_g. Meanwhile, the model predicted T_g values of 8 205 096 polyimides and a comprehensive screening has been implemented. Finally, combined with chemical knowledge, we derived 110 alternative polyimide structures with T_g exceeding 400, which can be used by scientists for structure filtering. Furthermore, we propose an accessible toolkit – polyScreen, which enables predicting T_g from structures in one step even without programming experience and more prediction modules will be updated in the next releases. This work demonstrates the feasibility of accurate prediction of material properties, providing potential polyimides with excellent properties for flexible displays, membranes and fibers utilizing graph neural networks, and provides quantitative guidance for the design of polyimides.