Improving The Accuracy Of Density Functional Theory Calculation For Homolysis Bond Dissociation Energies Of Y-NO Bond: Neural Network And Support Vector Machine Methods

The chemical bond energy of a molecule is the one of the its thermochemical properties. The bond strength can measure the stability of the molecule. The bond energies can control the rate of many reactions and determine the chemical reaction mechanism. Therefore, the accurate prediction the bond energy is one of the important topics in computational chemistry. The Hartree Fock approach can calculate the accurate bond energy, while it is hard to describe the large system due to expensive computation. As we know, density functional theory （DFT） approach has the advantage of possessing the low spin contamination and fast compuational efficiency with respect to ab initio approach, and in particular, the accuracy of its results can sometimes be comparable to that of highly-correlated approach such as multiconfiguration or multireference. However, the DFT result usually underestimates the bond energy. To resolve this, the efficient way to correct such errors is desired.In the present work, mean impact value, grey relational analysis, principal component analysis, neural network and least squares support vector machine approaches have been applied to improve the calculation accuracy of quantum chemical approaches for homolysis bond dissociation energies （BDE） of 92 organic molecules. With general physical parameters, these combined approaches can greatly eliminate the systemic errors of theoretical calculation due to ignoring the electron correlation and using small basis set, and will be a novel tool for predicting the properties of the molecules. Our work has been focus on following aspects:1. The back propagation neural network （BPNN） approach based on mean impact value （MIV） （MIV-BPNN） was used to improve the accuracy of density functional theory （DFT） calculation for homolysis BDE of Y-NO bond. Quantum chemistry calculations and MIV-BPNN are used jointly to calculate the homolysis BDE of 92 Y-NO organic molecular systems. The results showed that compared to a single density functional theory B3LYP/6-31G（d） approach, full parameters BPNN approach reduces the root-mean-square （RMS） of the calculated homolysis BDE of 92 organic molecules from 5.31 to 0.45 kcal/mol and MIV-BPNN approach further reduces the RMS to 0.33 kcal/mol. The corrected results of combined B3LYP/6-31G（d） and MIV-BPNN approach are in good agreement with experimental results.2. We propose a generalized regression neural network （GRNN） approach based on grey relational analysis （GRA） and principal component analysis （PCA） （GP-GRNN） to improve the accuracy of density functional theory （DFT） calculation for homolysis BDE of Y-NO bond. GRA is used to select the appropriate physical parameters. PCA is used to optimize the selected physical parameters. Finally, GRNN is used to establish nonlinear model. Quantum chemistry calculations and GP-GRNN are used jointly to calculate the homolysis BDE of 92 Y-NO organic molecular systems. We compare the GP-GRNN correction results with the B3LYP/6-31G （d） correction results, the correction results of the full-descriptor GRNN without GRA and PCA （F-GRNN） and with GRA （G-GRNN）, respectively. The results show that the F-GRNN and with G-GRNN approaches reduce the RMS of the calculated homolysis BDE of 92 organic molecules from 5.31 to 0.49 and 0.39 kcal mol^-1 for the B3LYP/6-31G （d） calculation. Then the newly developed GP-GRNN approach further reduces the RMS to 0.31 kcal mol^-1. Thus, the GP-GRNN correction on top of B3LYP/6-31G （d） can improve the accuracy of calculating the homolysis BDE in quantum chemistry and can predict homolysis BDE which cannot be obtained experimentally.3. The least squares support vector machine （LS-SVM） approach based on self-organizing feature map （SOM） （SOM-LS-SVM） was used to improve the calculation accuracy of DFT theory. SOM is used to select the appropriate physical parameters. K-CV is used to select the parameters of LS-SVM. And LS-SVM is used to establish nonlinear model.LS-SVM and SOM-LS-SVM approaches reduce the RMS of the calculated homolysis BDE of 92 organic molecules from 5.31 to 0.33 and 0.28 kcal mol^-1 for the B3LYP/6-31G （d） calculation. The results showed that the feasibility and effectiveness of the SOM-LS-SVM approach. And, the SOM-LS-SVM correction on top of the B3LYP/6-31G（d） results is a better approach to predict homolysis BDE and can be used as the approximation of experimental results when the experimental results are limited to measurement with very high accuracy. SOM-LS-SVM greatly extends the reliability and applicability of the B3LYP/6-31G（d） approach.