Linear regression correction to first principle theoretical calculations - Improved descriptors and enlarged training set

The linear regression correction previously developed to reduce quantum chemical calculation errors [X.M. Duan, G.L. Song, Z.H. Li, X.J. Wang, G.H. Chen, K.N. Fan, J. Chem. Phys. 121 (2004) 7086] has been further improved by using new descriptors obtained from natural bond orbital analysis and an enlarged training set of 350 organic, inorganic molecules and radicals. The new scheme is better suited for correcting reaction barriers. Upon linear regression correction, the mean absolute deviation for the new set decreases from 284.1, 8.2, 12.4 kcal/mol to 7.3, 3.3, 2.7 kcal/mol for the HF/6-31G(d), B3LYP/6-31G(d), and B3LYP/6-311G(2d,d,p) methods, respectively, and the mean absolute deviation of 12 barrier heights for six hydrogen transfer reactions is reduced from 5.3 to 2.9 kcal/mol for the B3LYP/6-311G(2d,d,p) method.