Higher-level coarse-graining (CG) significantly reduces computational cost but often faces challenges in structural fidelity, such as chain crossing. As geometric details are oversimplified and short-range repulsions are averaged out, this can lead to unphysical overlap or crossing behavior. Anisotropic CG models, which preserve segmental geometry and directionality, offer a promising strategy to address these limitations. Polytetrafluoroethylene (PTFE), a linear and semi-rigid polymer with a Kuhn length of approximately 23.4 Å in the melt state, serves as an ideal candidate for such modeling. Our findings indicate that the anisotropic effect increases by over 200% when CG beads represent 4 to 12 CF₂ groups. To balance accuracy and efficiency, we develop structure-based anisotropic CG models at two resolutions, corresponding to 6 and 8 CF₂ groups per bead, in which nonbonded pairwise interactions are described by the RE-squared potential [1]. We further evaluate their performance against isotropic CG models [2] by comparing local and global structural properties, dynamic characteristics such as self-diffusivity, and density predictions across various chain lengths and temperatures. While the isotropic models offer better computational efficiency, the anisotropic models provide distinct advantages. For instance, they can accurately capture the peak positions in angular radial distribution functions, even when the potential parameters are optimized solely based on angle-averaged radial distribution functions. In addition, we highlight the benefits of using analytical potentials over numerical ones. Analytical forms provide physically meaningful parameters, enabling reliable constant-pressure simulations based on dimensionless unit analysis. Complementing the structure-based approach, a force-matching CG method using TorchMD-Net [3], which leverages graph neural network architectures, is also under development to enhance model transferability across different thermodynamic conditions and polymer systems.
[1] Hsu YT, Yu HY. Macromolecules, 2025. DOI: 10.1021/acs.macromol.4c02413
[2] Salerno KM, Bernstein N. J. Chem. Theory Comput., 2018. DOI: 10.1021/acs.jctc.7b01229
[3] Doerr S, Majewski M, et al. J. Chem. Theory Comput., 2021. DOI: 10.1021/acs.jctc.0c01343
