Heat-Resistant Polymer Discovery by Utilizing Interpretable Graph Neural Network with Small Data

Abstract

Polymers with exceptional heat resistance are critically valuable in numerous domains, particularly as essential components of flexible organic light-emitting diodes. Among these, polyimides (PIs) demonstrate significant potential as substrate candidates for these next-generation flexible displays due to their robustness. However, traditional Edisonian approaches struggle to navigate the vast chemical space of PIs and also pose challenges of small data, which constrains the learnable chemical space for machine learning (ML). In this study, we propose a chemical-knowledge-based strategy to facilitate the design of PIs with high glass transition temperature (T g ) utilizing an atom-wise graph neural network and small data. Inspired by chemical intuition, our strategy leverages the available data on the same property (i.e., T g ) from other polymers, which is beneficial for expanding the chemical space used for ML. The trained ML model achieves an impressive performance in predicting T g of polymers. We have also investigated the impact of the chemical space encompassed by the data sets on the performance of ML models. Through interpretability analysis, it has been demonstrated that our ML model has learned more accurate chemical knowledge. Utilizing the ML model, 89 PIs were rapidly discovered from over 10 6 candidates, with experimental validation confirmed their exceptional heat resistance of the most promising PIs, which have been found to possess a T g exceeding 405 °C and even 450 °C. These results, along with the trained ML model, have the potential to accelerate the discovery of polymer substrate materials for next-generation flexible display devices.