Particle-swarm structure prediction on clusters.

We have developed an efficient method for cluster structure prediction based on the generalization of particle swarm optimization (PSO). A local version of PSO algorithm was implemented to utilize a fine exploration of potential energy surface for a given non-periodic system. We have specifically devised a technique of so-called bond characterization matrix (BCM) to allow the proper measure on the structural similarity. The BCM technique was then employed to eliminate similar structures and define the desirable local search spaces. We find that the introduction of point group symmetries into generation of cluster structures enables structural diversity and apparently avoids the generation of liquid-like (or disordered) clusters for large systems, thus considerably improving the structural search efficiency. We have incorporated Metropolis criterion into our method to further enhance the structural evolution towards low-energy regimes of potential energy surfaces. Our method has been extensively benchmarked on Lennard-Jones clusters with different sizes up to 150 atoms and applied into prediction of new structures of medium-sized Li(n) (n = 20, 40, 58) clusters. High search efficiency was achieved, demonstrating the reliability of the current methodology and its promise as a major method on cluster structure prediction.