A generalized exchange-correlation functional: the Neural-Networks approach

Abstract

A Neural-Networks-based approach is proposed to construct a new type of exchange-correlation functional for density functional theory. It is applied to improve B3LYP functional by taking into account of high-order contributions to the exchange-correlation functional. The improved B3LYP functional is based on a neural network whose structure and synaptic weights are determined from 116 known experimental atomization energies, ionization potentials, proton affinities or total atomic energies which were used by Becke in his pioneer work on the hybrid functionals [J. Chem. Phys. ${\bf 98}$, 5648 (1993)]. It leads to better agreement between the first-principles calculation results and these 116 experimental data. The new B3LYP functional is further tested by applying it to calculate the ionization potentials of 24 molecules of the G2 test set. The 6-311+G(3{\it df},2{\it p}) basis set is employed in the calculation, and the resulting root-mean-square error is reduced to 2.2 kcal$\cdot$mol$^{-1}$ in comparison to 3.6 kcal$\cdot$mol$^{-1}$ of conventional B3LYP/6-311+G(3{\it df},2{\it p}) calculation.