Machine Learning-Based Coarse Grained Interaction Potentials for Molecular Systems

The development and implementation of hierarchical multiscale methods is necessary for the molecular simulation of complex chemical systems such as organic fluids and soft matter systems, in order to reach longer length and time scales. Coarse-Graining (CG) is at the core of multiscale methods. Machine Learning (ML) techniques have been investigated in the recent years for the development of improved atomistic force fields based on quantum mechanical calculations. However, the integration of ML methods into the development of CG force fields required for hierarchical multiscale modelling schemes for bulk organic systems, on the basis of atomistic simulations, is still very scarce. In this work, Graph Convolutional Neural Network (GCNN) architectures were adopted to develop CG Machine Learned potentials for bulk amorphous systems, implementing a strategy that includes a force-matching scheme using benzene liquid as a test system. Two CG representations were implemented for the benzene molecule: a single bead and a three-bead representation. The ML-based CG force fields were utilized to perform molecular dynamics (MD) simulations of the systems at the CG level and the properties of the CG systems simulated with a neural network potential were compared against the underlying atomistic ones to quantify the effectiveness of the developed models. The effects of hyperparameters, loss function construction and GCNN architecture size were thoroughly investigated and discussed providing a wealth of information that can serve as a general basis for the reliable use of ML-based CG approaches in the wide field of bulk soft matter systems.