PolyT-GNN: A Graph Neural Network Framework for Data-Driven Discovery of High-Temperature Two-Way Shape Memory Polymers

Two-way shape memory polymers (2W-SMPs) represent a promising class of smart materials capable of reversible actuation under cyclic thermal stimuli. Despite their potential for high-performance applications in aerospace, geothermal, self-healing, biomedical, and soft robotics, the discovery of new 2W-SMPs remains limited by the lack of large, systematic data sets and predictive design strategies, especially for high-temperature systems. Here, PolyT-GNN is introduced as a graph neural network framework for data-driven discovery of high-temperature 2W-SMPs. A curated data set of 170 experimentally validated 2W-SMPs is compiled from the literature. PolyT-GNN integrates atomic, bond, and molecular descriptors with explicit monomer weight ratios to accurately predict polymer transition temperatures, achieving a test R2 of 0.84 even with limited data. Pretraining and fine-tuning together enhance prediction accuracy by about 38% compared to training from scratch, demonstrating strong cross-domain transferability. The framework further generates and screens over 80693 new 2W-SMP formulations, from which a polyethylene–dicumyl-peroxide system is synthesized and experimentally validated, showing a melting transition near 130 °C in close agreement with the predicted 113.25 °C. PolyT-GNN serves as a robust framework linking molecular structure, composition, and actuation behavior, enabling rational design of high-temperature two-way shape memory polymers for next-generation smart materials.