Machine Learning-Assisted Multi-Target Coarse-Graining Strategy for Polystyrene

Coarse-grained (CG) molecular dynamics offers a powerful means to bridge atomistic simulations and macroscopic experiments, but constructing CG models that simultaneously preserve structural, thermodynamic, and dynamical consistency remains challenging. Here, we present a machine learning-assisted, multi-objective parameterization strategy for atactic polystyrene (PS) based on a 2:1 mapping scheme. By integrating Support Vector Regression (SVR) and Particle Swarm Optimization (PSO), we systematically optimize Lennard-Jones parameters to reproduce atomistic-level radial distribution functions, density, cohesive energy density, and self-diffusion coefficients at 600 $K$ and 1 $atm$ . Notably, the inclusion of the diffusion coefficient as an optimization target enables the construction of a dynamically consistent CG model. The resulting CG force field achieves remarkable agreement with all-atom (AA) simulations across multiple observables, establishing a robust framework for predictive polymer modeling. This methodology provides a framework that could be extended to materials discovery and rational polymer design in future studies.