Property-guided generation of complex polymer topologies using variational autoencoders

Abstract The complexity and diversity of polymer topologies, or chain architectures, present substantial challenges in predicting and engineering polymer properties. Although machine learning is increasingly used in polymer science, applications to address architecturally complex polymers are nascent. Here, we use a generative machine learning model based on variational autoencoders and data generated from molecular dynamics simulations to design polymer topologies that exhibit target properties. Following the construction of a dataset featuring 1342 polymers with linear, cyclic, branch, comb, star, or dendritic structures, we employ a multi-task learning framework that effectively reconstructs and classifies polymer topologies while predicting their dilute-solution radii of gyration. This framework enables the generation of polymer topologies with target size, which is subsequently validated through molecular simulation. These capabilities are then exploited to contrast rheological properties of topologically distinct polymers with otherwise similar dilute-solution behavior. This research opens avenues for engineering polymers with more intricate and tailored properties with machine learning.