Neural network correction for heats of formation with a larger experimental training set and new descriptors

A neural-network-based approach was applied to correct the systematic deviations of the calculated heats of formation for 180 organic molecules and led to greatly improved calculation results compared to the first-principles methods [J. Chem. Phys. 119 (2003) 11501]. In this work, this neural network approach has been improved by using new descriptors obtained from natural bond orbital analysis and an enlarged training set including organic, inorganic molecules and radicals. After the neural network correction, the root-mean-square deviations for the enlarged set decreases from 11.2, 15.2, 327.1 to 4.4, 3.5, 9.5 kcal/mol for the B3LYP/6-31G(d), B3LYP/6-311G(2d,d,p) and HF/6-31G(d) methods, respectively.