Reaching extended length scales and time scales in atomistic simulations via spatially parallel temperature-accelerated dynamics

Abstract

We present a method for performing parallel temperature-accelerated dynamics (TAD) simulations over extended length scales. In our method, a two-dimensional spatial decomposition is used along with the recently proposed semirigorous synchronous sublattice algorithm of Shim and Amar [Phys. Rev. B 71, 125432 (2005)]. The scaling behavior of the simulation time as a function of system size is studied and compared with serial TAD in simulations of the early stages of $\mathrm{Cu}∕\mathrm{Cu}(100)$ growth as well as for a simple case of surface relaxation. In contrast to the corresponding serial TAD simulations, for which the simulation time ${t}_{ser}$ increases as a power of the system size $N$ $({t}_{ser}\ensuremath{\sim}{N}^{x})$ with an exponent $x$ that can be as large as three, in our parallel simulations the simulation time increases only logarithmically with system size. As a result, even for relatively small system sizes our parallel TAD simulations are significantly faster than the corresponding serial TAD simulations. The significantly improved scaling behavior of our parallel TAD simulations over the corresponding serial simulations indicates that our parallel TAD method may be useful in performing simulations over significantly larger length scales than serial TAD, while preserving all the atomistic details provided by the TAD method.