(709b) Generative Modeling of Polymer Melts for Solving Inverse Design Problems

The ability to design novel polymeric materials with specified properties requires solving an, often, ill-posed problem via thorough exploration of a high-dimensional design space. Inverse design becomes especially challenging for polymeric systems as they exhibit for each design parameter, structural arrangements and order/disorder at various length scales that together lead to the desired material property. Choosing to use traditional molecular dynamics (MD) simulations alone to sample all of the relevant polymer chain configurations and various length scales of structural arrangements for every design parameter can be computationally expensive due to long relaxation times of chains. In this work, we introduce a framework to effectively explore the configurational space of polymer melts using machine learning and MD simulations. We utilize a variational autoencoder (VAE) trained on MD simulation trajectories of melts of coarse-grained polymer chains to learn chain configurations and embed them onto a continuous distribution of low-dimensional latent space variables. Our methodology can generate many chain configurations including ones (e.g., non-equilibrium extended configurations upon processing) that may be computationally difficult to sample via traditional MD simulations. Using a simple A-B coarse-grained bead-spring diblock copolymer and its design parameter space as an example, we demonstrate our approach. We will conclude our talk with a brief discussion of potential extensions of this framework specifically for design of next generation of novel, high-performance polymeric materials.