Machine-learning assisted coarse-grained model for epoxies over wide ranges of temperatures and cross-linking degrees

Abstract

We present a practical computational framework for the coarse-graining of cross-linked epoxies by developing a machine-learning technique, which integrates molecular dynamics simulations with artificial neural network (ANN) assisted particle swarm optimization (PSO) algorithm. Key features of the framework include two aspects: (1) determining the bonded interactions via the iterative Boltzmann inversion method to emulate the local structures of the epoxies and, (2) optimizing the nonbonded interaction potentials through the machine-learning approach to reproduce the mechanical properties. Such machine-learning based technique is computationally efficient in searching for the optimal solution of nonbonded potential parameters and enables the CG model to become transferable within a wide range of cross-linking degrees. This is mainly attributed to the fact that ANN can give good predictions based on training database obtained from CG simulations and thus greatly accelerates the PSO algorithm in achieving the optimal solution. On the basis of the DOC-transferable CG model, the cohesive interaction strength is phenomenologically adjusted to preserve the temperature-dependent properties. The CG model allows the mechanical properties of cross-linked epoxies to be predicted with reasonable accuracy over wide ranges of cross-linking degrees and temperature. The proposed framework will become highly beneficial to the design of high performance epoxy-matrix nanocomposites.