Two new exchange-correlation functionals are developed using an optimi
zation procedure involving experimental and ab initio data fbr a chose
n training set of systems. The limited first-row training sets used in
our preliminary studies are significantly expanded to include 68 atom
s and molecules involving both first- and second-row atoms. By trainin
g on (i) exchange-correlation potentials computed from ab initio data,
(ii) experimental molecular atomization energies, and (iii) near-exac
t atomic total energies, the TH3 GGA functional is developed. The func
tional 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 convention
al GGA functionals. To this cause the least-squares algorithm is amend
ed to incorporate (iv) the exchange-correlation energy gradient vector
; then the TH4 GGA functional is derived. This yields comparable energ
etic 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 orb
ital exchange. The derivation of the TH3 and TH4 functionals highlight
s the importance of a dominant exchange contribution if a functional i
s to be applicable to a wide range of molecular systems.