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

Abstract

A Neural-Networks approach is employed to improve B3LYP exchange-correlation functional by taking into account of high-order contributions. The new B3LYP functional is based on a Neural-Network whose structure and synaptic weights are determined from 116 known experimental energy data [J. Chem. Phys. 98 (1993) 5648]. It leads to better agreement between the first-principles calculations and the experimental results. The new functional is further tested by applying it to calculate 40 ionization potentials and 40 enthalpies of formation in G2-2 test set [J. Chem. Phys. 109 (1998) 42] using 6-311+G(3df,2p) basis set. The root-mean-square errors are reduced from those of conventional B3LYP calculations. © 2004 Elsevier B.V. All rights reserved.