Quantitative structure-property relationship of standard enthalpies of nitrogen oxides based on a MSR and LS-SVR algorithm predictions

Abstract

Quantum chemistry method is used to calculate 1444 descriptors of each nitrogen oxide molecule. Variables are screened by using multiple stepwise regression (MSR). The optimal ten-element regression equation is derived, its R2 = 0.934 and Q2 = 0.927. The support vector regression (SVR) model between ten molecular descriptors and standard formation enthalpy is established, and the support vector machine is optimized by using the least squares method (LS). The training set consisting of 78 compounds have R2 = 0.969 and Q2 = 0.954. The other 22 compounds constitute the test set to verify the external prediction ability of the model, and its R2 = 0.958. It demonstrates that the LS-SVR model has good stability and prediction ability, and overcomes the problem of over-fitting. The conclusion of quantitative structural property relationship (QSPR) shows that the molecular descriptor and the standard enthalpy of formation are non-linear second-order functions. We speculate that the electronegativity of atoms is the key to determine the standard enthalpy of formation of compounds