SEEDPoly: A Latent Embedding-Based Screening Framework for Polymer Design

Machine learning (ML) has enabled efficient polymer design through virtual screening and inverse molecular generation. However, de novo design of functional polymers remains challenging due to the vast chemical space and limited availability of labeled polymer data. To address this, we introduce SEEDPoly (Screening of Embeddings for Extending into Designable Polymers), a latent space search framework for identifying structurally dissimilar molecules with similar target properties. SEEDPoly integrates deep neural networks, latent chemical distance metrics, and rule-based polymerization to explore molecular libraries in the learned latent space and convert selected candidates into designable polymers. We apply SEEDPoly to gas separation tasks by coupling it with a customized graph neural network (GNN) to search the molecular space and generate millions of hypothetical polymers. Their gas permeability performances are predicted using feedforward neural networks (FNNs), XGBoost (XGB), and random forest (RF) models. Top-performing candidates exceeding the Robeson upper bound are further validated using molecular dynamics (MD) simulations. Our results demonstrate that SEEDPoly can efficiently identify structurally diverse small molecules with desired properties and transform them into synthetically accessible polymers via known reaction rules. Overall, SEEDPoly provides a powerful and interpretable framework for data-efficient and scalable polymer discovery, paving the way toward rational design of next-generation functional materials