Transfer learning-driven artificial intelligence model for glass transition temperature estimation of molecular glass formers mixtures

Abstract

Predicting binary mixtures' glass transition temperature (Tg) is crucial in various fields, particularly for industrial materials affected by this property during production processes and in service-life. On the other hand, from the fundamental point of view, this predictive capability is relevant for understanding the chemical interactions between the two components and how this affects the Tg of the mixture. In this sense, some models provide different approaches for describing the Tg of the mixture. Among them, the Gordon-Taylor approach has been widely used since it only relies on the relationship between the Tg of the pure components, their weight fraction, and only one fitting parameter. Although simple, this approach still requires measurements of Tg of the pure components and at least some intermediated composition for the fitting procedure. In a previous work, our research has focused on neural networks methods for predicting Tg values directly from the chemical structure of monomers and molecules, but the scarcity of experimental data for binary mixtures limits the application of a similar approach. To address this problem, we propose to use in this work a transfer learning method that relays on the previous acquired knowledge of the chemical structure - Tg relationship, for the prediction of the Tg of the binary mixtures. Therefore, pure component characteristics are derived from chemical fingerprints originated in a pre-trained network, and enables a training process focused on their behavior within the mixtures. This approach successfully estimated K with very low deviations, even allowing for the exploration of the embedded chemical structure's relation to previously unknown mixtures.