Data-Enabled Coarse-Graining of Confined Simple Liquids

Data-driven approaches have achieved considerable success in effectively parameterizing Coarse-Grained (CG) force fields to emulate the behavior of All-Atom (AA) systems. While much attention has been focused on homogenous systems, application to confined fluids remains a challenge. To address this, we present a data-driven framework aimed at determining coarse-grained Leonard-Jones (LJ) potential parameters for simple liquids within confined systems. Our approach entails leveraging a Deep Neural Network (DNN) trained to approximate solutions to the Inverse Liquid State (ILST) problem specifically for confined systems. Using transfer learning, we predict single-site LJ potentials for simple multiatomic liquids within a slit-like channel. Our findings demonstrate that the DNN-parameterized CG potential adeptly reproduces both, the fluid structure and molecular forces of the target AA system, particularly when the electrostatic interactions or multibody effects in the AA system are not dominant. To transcend this limitation and enhance model fidelity, we propose a hybrid approach that integrates a data-driven framework with a well-established coarse-graining method based on Relative Entropy minimization (RE). This hybrid paradigm leads to a symbiotic benefit where the DNN-generated potentials serve as an intelligent starting point for the RE iterations leading to rapid convergence as well as improving the model fidelity towards more complex systems.