Global Optimization of H-Passivated Si Clusters at the Ab Initio Level via the GAM1 Semiempirical Method

A genetic algorithm (GA)-based procedure for finding the global minimum of medium-sized SixHy clusters at the ab initio level is presented. Semiempirical calculations are used to prescreen the relative energies of the many locally optimized SixHy structures generated by the GA. The AM1 semiempirical method used in our original global optimization strategy produces cluster energies appreciably different from the ab initio ranking. To improve the AM1 energy ranking we recently explored using for each specific SixHy stoichiometry two coupled GAs: (1) a parametrization GA (PGA) for finding better semiempirical GAM1 parameters for Si and (2) a cluster global optimization GA (CGA) that uses the GAM1 parameters in the cluster energy prescreening. The two GAs method is a more reliable approach for global optimization of different SixHy stoichiometries because the GAM1 parameters adjust to whether the Si atoms are partially or completely passivated by H atoms. Unfortunately, even for medium-sized SixHy clusters we find the two GAs method is too computationally demanding. In this paper we examine under what conditions the GAM1 parameters obtained for a specific small SixHy stoichiometry could be successfully transferred to larger clusters so that only the CGA global search need be performed. We choose Si7H14 as the starting stoichiometry since it can adopt various possible structures which model a Si7 core completely passivated by H atoms. We find the resulting GAM1 parameters obtained from the Si7H14 training set produce relative cluster energies matching well wit