Machine-learned closure for polymer liquid state theory

Polymer reference interaction site model (PRISM) theory is a powerful computational technique to efficiently predict the structure and thermodynamics of liquid-like polymer systems. However, the use of PRISM theory can pose practical challenges due to issues with accuracy and numerical stability for some classes of polymer systems. Typically, the analytical closure relation used to solve the PRISM equations is considered the key barrier toward broader applicability of the theory. In this talk, we describe our efforts to develop a machine learning (ML)-based closure relationship for PRISM theory. We prepared a dataset for model training using coarse-grained molecular dynamics simulations of homopolymer melts and solutions across a range thermodynamic conditions and developed an ML-based closure to predict 2-body structural correlation functions within the PRISM solution loop. Our preliminary results show that the ML closure performs favorably in comparison to widely-used analytical closures (e.g., Percus-Yevick). We then applied the ML closure to model the results of small-angle neutron scattering experiments on polymer solutions, again showing superior performance compared to traditional approaches.