Accurate interatomic force field for molecular dynamics simulation by hybridizing classical and machine learning potentials

Abstract

Full atom simulations have demonstrated the scalability for billions of atoms, but still suffered from the transferability of semi-empirical interatomic potentials. We propose a dynamic multiscale molecular dynamics (MD) simulation method with both high accuracy and efficiency in interatomic force field calculation by hybridizing both classical and machine learning (ML) potentials. A dynamic procedure has been adopted by evaluating the centro-symmetry parameter of evolving microstructures during MD simulations and accordingly modifying the highly distorted regions depicted by ML potentials. Atomic force field calculation in near-perfect or perfect lattices remain sticking to the fast EAM potential, which precisely captures the long range elastic interactions. A handshaking region is introduced in order to enforce the continuity in atomic interactions. The MD simulations using a dynamic multiscale scheme can achieve the ab-initio accuracy without raising considerable computational cost. The foundation of this approach deeply roots in the facts that the ML method has comparable accuracy to ab-initio MD simulations and possesses the same order of computation complexity to the classical MD (O(N)). The proposed multiscale method, attaining both high accuracy and efficiency simultaneously, will pave the way for solving many material science problems involving complex interactions of both long range elastic forces and short range chemical bonding/debonding processes.