Long-time molecular dynamics simulations on massively parallel platforms: A comparison of parallel replica dynamics and
- PDF / 479,297 Bytes
- 10 Pages / 584.957 x 782.986 pts Page_size
- 38 Downloads / 251 Views
Molecular dynamics (MD) is one of the most widely used techniques in computational materials science. By providing fully resolved trajectories, it allows for a natural description of static, thermodynamic, and kinetic properties. A major hurdle that has hampered the use of MD is the fact that the timescales that can be directly simulated are very limited, even when using massively parallel computers. In this study, we compare two time-parallelization approaches, parallel replica dynamics (ParRep) and parallel trajectory splicing (ParSplice), that were specifically designed to address this issue for rare event systems by leveraging parallel computing resources. Using simulations of the relaxation of small disordered platinum nanoparticles, a comparative performance analysis of the two methods is presented. The results show that ParSplice can significantly outperform ParRep in the common case where the trajectory remains trapped for a long time within a region of configuration space but makes rapid structural transitions within this region.
I. INTRODUCTION
Molecular dynamics (MD), the numerical solution of atomistic equations of motion, is an established workhorse in the computational physical sciences. Its popularity—as of 2017, more than 3 million papers refer to MD—relies on its predictive power: because MD generates an unbiased sample from the ensemble of valid trajectories given some dynamics and some initial conditions, they can, in principle, be used to compute any thermodynamic, or perhaps even more importantly, dynamic quantity of interest. These remarkable advantages, however, come at a rather steep computational price that often prohibits direct simulations in relevant experimental regimes of size, time, temperature, strain rate, etc. An important cause of this limitation is rooted in constraints imposed by the complex, multiscale nature of the potential energy surface, i.e., the (3N 1 1)dimensional surface that assigns a potential energy to every point in configuration space. In the case of hard materials, such a surface is typically composed of a very large number of local minima, separated from one
Contributing Editor: Vikram Gavini a) Address all correspondence to this author. e-mail: [email protected] b) This author was an editor of this journal during the review and decision stage. For the JMR policy on review and publication of manuscripts authored by editors, please refer to http://www.mrs. org/editor-manuscripts/. c) Permanent address: Department of Physics, Xiamen University, Xiamen 361005, China. DOI: 10.1557/jmr.2017.456
another by energetic and/or entropic barriers. In this setting, the maximum integration time step dt has to be a small fraction of the period of the fastest vibrational mode of the system (which is controlled by the highest curvature around minima; for hard materials dt ; 1 fs). On the other hand, “interesting” trajectories for materials science application should capture micro/nano-structural evolution; i.e., they should contain many jumps from one basin to another. The so-
Data Loading...
