Machine learning of an implicit solvent for dynamic Monte Carlo simulations

Abstract

The Bond Fluctuation Model (BFM) is a highly efficient and versatile method for simulating polymers, membranes, and soft matter. Due to its coarse-grained nature, the BFM is employed to understand the universal properties of polymers. Solvent effects are often mediated by explicit solvent particles, while implicit solvent models have had limited use as they may lead to frozen states and, thus, ergodicity-related problems. In simulation setups, such as coagulated multiple homopolymers chains, explicit solvent models are computationally expensive because the region of interest can be localized in a small space compared to the dimension of the periodic box. We introduce an implicit solvent model based on an artificial neural network (NN) that was trained with BFM simulation data for single homopolymers in an explicit solvent. We demonstrate that NN-based simulations that take into account only the information of the local environment of monomers reproduce the expected universal macroscopic properties of the polymer under varying solvent conditions. The homopolymer chains simulated using the NN reproduce the coil–globule transition, the static and dynamic bond autocorrelation, and the mean square displacement of chain monomers. We show that the learned parameters from a single chain system can be transferred to a system containing multiple homopolymers, indicating that the learned parameters are transferable to considerably different systems.