Parallel bond order potentials for materials science simulations
C. Teijeiro Barjas, T. Hammerschmidt, R. Drautz, G. Sutmann.
The computation of interatomic interactions in materials science is a challenging problem, because of the need for an accurate description of different bonding situations. Density functional theory (DFT) and tight binding (TB) provide good approximations to the problem but have high computational complexity, which limits the size of the systems to be studied. Analytic bond-order potentials (BOPs) provide a coarse-grained computation of interatomic interactions derived from DFT and TB in order to obtain satisfactory approximations, with an order-N increase in the simulation time as the system size grows. Even though BOPs are significantly less expensive than first principle methods, analytic BOPs require an efficient implementation in order to obtain good scalability for large systems. This paper presents a performance evaluation of a parallel implementation of a BOP code, with a description of the most time consuming tasks, and basic concepts for a parallelisation of the simulation. The main contributions of this paper are (1) the analysis of an optimized simulation code in terms of its different routines, (2) the implementation of parallel algorithms that take advantage of the nature of the simulation to obtain high scalability, (3) a performance evaluation of the parallel code on average-sized systems and the proposal of best practices for future developments, and (4) the example of integration of the routine for the precise computation of energies and forces in a molecular dynamics (MD) code.
Change of potential energy during relaxation of a 1/2 screw dislocation in bcc-W computed by parallel computations with analytic bond-order potentials. The colour coding indicates atom-resolved binding energy.