Development of Machine Learned Potentials for Coarse Grained Simulations of Polymers

In this work, we adopted Graph Convolutional Neural Networks (GCNN) to develop Coarse Grained (CG) Machine Learned potentials for bulk polymer systems, implementing a scheme that includes a force-matching procedure. The GCNN models were used to perform CG Molecular Dynamics (MD) simulations of polyethylene and of a polymer of intrinsic microporosity (PIM-1). The structural and thermodynamic properties of the CG systems were compared with the underlying atomistic reference, examining the effect of the CGNN model size and hyperparameters on the simulation results. The models obtained showed transferability to longer chain lengths than the ones use for training. The open-source python package SchNetPack was extended to support the study of macromolecular systems with connectivity information and inter- and intra-molecular neighbours distinctions. The extended version is are available on GitHub at the following link: https://github.com/ml-multimem/schnetpack-for-bulk-systems. Moreover, the GCNN models were interfaced with the MD engine LAMMPS. This methodology has the potential to streamline the generation of CG force fields, enabling systematic multiscale studies of complex macromolecular systems.