Deep Generative Models for Dense Polymer Systems

We employ a variational autoencoder (VAE) to efficiently sample independent configurations of polymer systems in the high-density phase. Using molecular dynamics simulations of a polyethylene chain, the VAE is trained to generate canonically distributed distance matrices, which are subsequently embedded in three-dimensional space following an energy minimization procedure.

The results show that the model accurately learns the system's free energy landscape, although the embedding procedure is necessary to fully recover the system’s physical properties. Additionally, we show that the latent representation retains information about the polymer's topological features. This was confirmed by encoding knotted polymers within the same VAE architecture and generating new samples conditioned on their topological state.

These findings suggest the potential of topology-aware deep models to explicitly control the generation of samples with complex entanglement structures and tailored material properties.