Exploring Multi-Type Crosslinked Architectures in Polymer Materials Using Graph Neural Networks

Crosslinked polymer networks provide a promising route for developing novel material composites with a variety of applications in areas such as aerospace and biomedicine. These systems are typically made up of different combinations of backbone polymer chains and crosslinking agents of different kind. Understanding the relationship between the polymer network architecture and its properties is an important topic for design and development of new generation of polymer composites with on-demand functionality. While the study of crosslinked polymer networks has so far been mainly limited to systems with one or two types of crosslinkers, the existing synthesis techniques can readily be extended to devise more complex polymer networks with a larger number of crosslinkers and backbone chain types. However, the high-dimensional parameter space associated with such complex systems makes it difficult to predict the resulting architectures and properties of the materials developed through experimental trial and error.<br/><br/>In this work, we explore new designs for crosslinked polymer materials with multiple types of backbone chains and crosslinkers using a computational framework established based on graph theory (GT) and graph neural networks (GNNs). We demonstrate that including three or four types of backbone polymer chains can improve the mechanical behavior of the polymer composite provided that suitable structural features are established. We develop a graph representation of various polymer networks where graph edge features are defined according to the nature of crosslinking in the system, i.e. covalent or noncovalent bonds. We use a GNN framework to investigate the relationship between experimentally-controlled structural parameters, such as crosslinker density and type, and GT-based structural descriptors, such as graph connectivity and Wiener index. Following this step, we develop another GNN framework to study structure-property relationships using GT-based descriptors and mechanical properties obtained from Molecular Dynamics (MD) simulations. The trained GNNs will be used to predict the behavior of new potential structural morphologies in order to identify ‘best-performing’ configurations. The results of this work are aimed to be utilized for future experimental development of new crosslinked polymer composites.