Machine Learning-Driven Discovery of Thermoset Shape Memory Polymers With High Glass Transition Temperature Using Variational Autoencoders

Abstract

The discovery of high-performance shape memory polymers (SMPs) with enhanced glass transition temperatures (Tg) is of paramount importance in fields such as geothermal energy, oil and gas, aerospace, and other high-temperature applications, where materials are required to exhibit shape memory effect at extremely high-temperature conditions. Here, we employ a novel machine learning framework that integrates transfer learning with variational autoencoders (VAE) to efficiently explore the chemical design space of SMPs and identify new candidates with high Tg values. We systematically investigate the effect of different latent space dimensions on the VAE model performance. Several machine learning models are then trained to predict Tg. We find that the SVM model demonstrates the highest predictive accuracy, with R ² values exceeding 0.87 and a mean absolute percentage error as low as 6.43% on the test set. Through systematic molar ratio adjustments and VAE-based fingerprinting, we discover novel SMP candidates with Tg values between 190°C and 200°C, suitable for high-temperature applications. These findings underscore the effectiveness of combining VAEs and SVM for SMP discovery, offering a scalable and efficient method for identifying polymers with tailored thermal properties.