Reactive Machine Learning Force Field for Crosslinked Epoxy

Large-scale molecular dynamics (MD) simulations of crosslinked epoxy with quantum-level accuracy while capturing complex reactivity is a compelling yet unrealized challenge. In this work, through the construction of a chemical-environment-directing dataset, a reactive machine learning force field that accurately captures both reactive events and thermos-mechanical properties is developed. The force field achieves energy and force root-mean-square errors of 1.3 meV/atom and 159 meV/Å, respectively, and operates approximately 1200 times faster than ab initio molecular dynamics. MD simulations demonstrate excellent predictive capabilities across multiple critical thermos-mechanical properties (radial distribution function, density, and elastic modulus), with results being well consistent with experimental values. In particular, the force field can provide accurate prediction of the bond dissociation energies for typical bonds with a mean absolute error of 7.8 kcal/mol (<8%), which enables the simulation of tensile-induced failure caused by chemical bond breaking. Our work demonstrates the capability of the machine learning force field to handle the extraordinary complexity of crosslinked epoxy systems, providing a valuable blueprint for future development of more generalized reactive force fields applicable to most polymers.