Structural optimization of atomic clusters by tabu search in descriptor space

We devised a new algorithm - tabu search in descriptor space - for searching the global energy minimum structure of atomic clusters. In each cycle, the algorithm generates many cluster structures randomly, transforms them to a standard form by "symmetrization of interatomic distances", and calculates structural descriptors for each. Clusters are then screened according to a model energy obtained by interpolation in descriptor space, and only a small fraction (10% or less) are retained for energy evaluation. This cycle is repeated many times. In a final step, clusters are sorted by increasing energy and optimized by conjugate gradient. This method requires between 10 and a 100 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 seems a very promising method for global optimization on energy surfaces calculated by first-principles.