Generative Multi-Property Refinement of Polymer Chemistries

We present a generative, multiobjective optimization method for polymer chemistry. By leveraging monomer-level properties that are correlated with polymer properties, we design step-growth polymers with targeted glass transition temperatures (T_g), band gaps (E_g), and Flory–Huggins interaction parameters with water (χ_water) across a broad chemical space. Generative design is accomplished using a variational autoencoder integrated with linear property prediction heads. Linear organization of the latent space enables the identification of a single latent vector to steer the simultaneous optimization of multiple polymer property objectives. Subsequent Bayesian optimization within the latent space allows further enhancement of T_g, E_g, and χ_water relative to a reference polymer chemistry. We then apply the generative model to design per- and polyfluoroalkyl substances (PFAS)-based polymers with reduced fluorine content but comparable physical properties. Overall, this work establishes a generative, multiobjective approach for navigating early stage polymer design, prior to experimental validation or computationally expensive simulations.