Abstract
Two new exchange-correlation functionals are developed using an optimization procedure involving experimental and ab initio data for a chosen training set of systems. The limited firstrow training sets used in our preliminary studies are significantly expanded to include 68 atoms and molecules involving both first- and second-row atoms. By training on (i) exchange-correlation potentials computed from ab initio data, (ii) experimental molecular atomization energies, and (iii) near-exact atomic total energies, the TH3 GGA functional is developed. The functional gives a mean absolute error in atomization and atomic energies of 2.6 kcal mol-1 for the 68 training systems, although it does not eliminate the significant bond length errors observed using conventional GGA functionals. To this cause the least-squares algorithm is amended to incorporate (iv) the exchange-correlation energy gradient vector; then the TH4 GGA functional is derived. This yields comparable energetic accuracy to TH3, but significantly improved bond lengths for the training set systems, indicating that structural predictions of simple GGA functionals can be improved without the introduction of exact orbital exchange. The derivation of the TH3 and TH4 functionals highlights the importance of a dominant exchange contribution if a functional is to be applicable to a wide range of molecular systems.