De novo multiscale method for nonequilibrium molecular dynamics

In this work, we report a de novo multiscale method that can significantly reduce the number of atoms required to simulate an atomistic system for nonequilibrium molecular dynamics (NEMD) problems. In our approach, first, we form a dynamical matrix that breaks an atomistic system into real and virtual domains, so as real and virtual atoms, Then, the dynamical matrix corresponding to the virtual domain is transformed into the time-history kernel function (THKF), where all the dynamic behavior is just the output of the kernel function taking the behavior of the bonded real atoms as the input so that the virtual atoms no longer need to participate in the simulation. In the practical approach, the real atoms are kept for typical NEMD ensembles, such as thermal states, to retrieve physical properties, while the virtual atoms not bonded with real atoms can be neglected, or even removed. Furthermore, with our semi-analytical approach, the THKF can be derived from eigenvalue and eigenvectors on the dynamical matrix of the virtual domain in arbitrary geometries. To examine whether the dynamic behavior is properly reserved, we first introduce 1-D wave propagation of atom chains with our multiscale method. In addition, we adopt this method to simulate different 3-D silicon-based nanowires with or without twin boundaries and point defects. All the results via our multiscale method agree with the counterpart from fully atomistic systems, which indicates the ability to obtain physical properties with fewer atoms in simulation models than in NEMD. Overall, this method has great potential to perform simulations to collect physical properties with only a small number of atoms at higher length scales, which is important for future studies on the physical meanings over higher length scales.