ABSTRACT
We devised a method, Tabu Search in Descriptor Space (TSDS), for global search of energy minima on potential surfaces of atomic clusters. In each cycle TSDS generates many structures at random, calculates their structural descriptors, and screens them using energy predictions based on descriptors. Only a small fraction (typically less than 10%) of the clusters are retained for energy evaluation. This cycle is repeated many times. In the final step, clusters are sorted and only the best few undergo local optimization. The TSDS method requires between ten and a hundred times fewer energy evaluations than a good genetic algorithm for locating the global minimum of n-atom clusters (n < 35) described by a Lennard-Jones potential. It is a promising method for global optimization of functions that are computationally expensive, for example, energy surfaces calculated by first-principles. We will discuss results obtained by combining TSDS with empirical potentials and Kohn-Sham theory that model clusters of Ar, Li, Si, and a few other elements.
