Efficient Parallel Decomposition of Dynamical Sampling in Glass-Forming Materials Based on an "On the Fly" Definition of Metabasins

In this work, we propose a highly parallelizable sampling scheme designed for atomistic simulations of glassy materials in the vicinity of the glass-transition temperature T-g, based on the idea of inherent structures (IS). Glassy dynamics is envisioned as a combination of two types of motions: (a) an "in basin" vibrational motion in the vicinity of a potential energy minimum (IS), and (b) transitions from one basin to another. In order to perform efficient dynamical sampling in the vicinity of T-g, we propose an on the fly" definition of metabasins (i.e., collections of basins communicating via fast transitions in which the system spends a sufficient time before moving on to a neighboring collection). Our criterion for defining metabasins is based on the rate of identification of new basins in the course of a canonical molecular dynamics (MD) run. In order to compute individual rate constants between basins and metabasins, we propose to follow a swarm of microcanonical MD trajectories initiated at phase-space points sampled by a canonical MD run that is artificially trapped within a metabasin. The execution time required by this highly parallelizable scheme is reduced dramatically, since no information exchange takes place between the microcanonical trajectories. Results from our parallel methodology are compared against results from artificially tapped canonical MD runs, in terms of the evaluated rate constants, and found to be in very good agreement. Parallel simulations have been conducted on up to 250 processors, achieving almost linear scaling. The validity of our definition of metabasins is confirmed by analysis of the resulting network of basins.