Improving Accuracy Of Quantum Chemical Calculation: Extreme Learning Machine Neural Network

The quantum chemical calculation methods have shown great advantages whenexplaining and predicting the experimental results of the small-sized and median sizedmolecules. However, the calculation errors of quantum chemical calculation methods arealways unavoidable, especially for those large-sized molecules of irregular structure, beacauseof the inherent approximation of these calculation methods. In order to obtain accurate resultsusing less machine time and computing resources, or to obtain results that cannot be calcuatedunder the the existing calculation condition by tradtional calculation methods, we combinedquantum chemical calculation methods with intelligent calculation method to find somesimple and efficient ways to improve the accuracy of quantitative calculations, or calculatethe properties those aredifficult to calculate by existing methods..The main contributions of this thesis are as follows:First, we aim to predict the first hyperpolarizability (β0) values at MP2method of36alkalides with the excess electrons. As we know, MP2method can yield a highly accurateresult of nonlinear optical properties, while the dramatic increase in computational cost withthe size of systems limits its applications to calculate the large molecules. On the other hand,Hartree-Fock method can calculate the first hyperpolarizability in a reasonable computationaltime, while the solution qualities are not always satisfactory. Therefore, in this work aneffective intelligent computing method, as called extreme learning machine-neural network(ELMNN), is proposed to predict accurately the first hyperpolarizability (β0) of alkalides fromlow-accuracy first hyperpolarizability. Results show that we can build a sort of quantitativerelationship between high-accuracy β0(MP2)values and low-accuracy β0(HF)values. In this way,we can use low-accuracy β0(HF)to predict the high-accuracy β0(HF)with less machine time andcomputing resources. Concretely, by using the proposed method, the RMS (root mean square)deviation of the predicted MP2calculated values is0.02a.u., which is better than the RMSdeviation of predicted value by NN and GANN (0.17and0.08a.u.), respectively. Obtainingthe high accuracy level calculated with less computing cost is another excellent point of theproposed method. Experimental results show that the computing time for MP2is2.4to4times of the computing time for our method.Moreover, the electronic excitation energies of90BODIPY derivatives were studied inthis thesis. The electronic excitation energy is a significant physical property for BODIPYdyes, and implies inherent structure information and electronic properties. Thus, it isimportant to calculate the electronic excitation energy accurately in the field of computationalchemistry. However, the B3LYP method has some limitations and shows increasing energy errors with the increase of the molecular size. These errors are connected to the approximatedexchange-correlation functional and there is no systematic way to construct the exactfunctional. In this work, the ELMNN method is used successfully to predict the experimentalelectronic excitation energies of BODIPYs.In order to investigate the intrinsic relation between the descriptors and the electronicexcitation energy, four different groups of descriptors are considered in our method.Experimental results show that the RMS deviation of the predicted electronic excitationenergies of our method with quantum chemical descriptors can be reduced to0.13eV. Thismeans that the group of quantum chemical descriptions plays the most important role inpredicting electronic excitation energies with the smallest RMS deviation. Finally, a webserver (EEEBPre: Prediction of electronic excitation energies for BODIPY dyes) is built forresearchers to obtain the predicted electronic excitation energy values of their BODIPY dyes,which is high consistent with the experimental values. This web server can be accessedpublically at http://202.198.129.218.