ORS10-0074
AI-Assisted Multiscale Polymer Simulation Workflow: From the Atomistic Scale to Phase-Separated Morphology Representation
When and Where
Nov 30, -0001
12:00am - 12:00am
Presenter(s)
HAEIN LEE (Resonac Corporation)
Co-Author(s)
Abstract
Phase-separated morphologies strongly influence polymer material properties, but direct multiscale simulations that connect polymer chemical structure, mesoscale phase-separated morphology, and macroscopic properties remain computationally impractical. To address this challenge, we present a multiscale, AI-assisted coarse-grained molecular dynamics (CGMD) framework that combines all-atom molecular dynamics (AAMD) reference data, Martini3-based CGMD, and multi-objective Bayesian optimization (MOBO) to develop transferable CG polymer models. Within a fixed Martini3 mapping, bonded parameters were optimized against AAMD-derived density and radius of gyration. For polyethylene, polystyrene, and poly(methyl methacrylate), the BO-optimized topologies improved agreement with AAMD compared with the corresponding unoptimized Martini3-based models and achieved errors below ~10% for both target properties across multiple degrees of polymerization, while retaining CGMD efficiency. We further evaluated whether models trained only on single-component structural properties can transfer to multi-component thermodynamic behavior. The Bayesian-optimized CG models reproduced phase separation in immiscible PS-PMMA and homogeneous mixing in miscible PVDF-PMMA, including a newly parameterized PVDF model. Density profiles and Flory-Huggins interaction parameters estimated from cohesive energy densities correctly ranked the relative miscibilities, supporting the use of the framework for polymer compatibility screening.
As a downstream direction, we also introduce morphology descriptors based on rotationally standardized two-point spatial correlations. These descriptors provide compact representations of mesoscale microphase-separated structures and can serve as inputs for future data-driven property prediction. This workflow demonstrates how Bayesian optimization can assist molecular simulation and how AI-assisted simulation can be used to design polymer materials across scales.
As a downstream direction, we also introduce morphology descriptors based on rotationally standardized two-point spatial correlations. These descriptors provide compact representations of mesoscale microphase-separated structures and can serve as inputs for future data-driven property prediction. This workflow demonstrates how Bayesian optimization can assist molecular simulation and how AI-assisted simulation can be used to design polymer materials across scales.
