Study On Correction Models Of Time-dependent Density Functional Theory Calculations For Molecular Absorption Energies Based On Ensemble Learning

Artificial intelligence(AI)is a frontier science with the aim of studying the theory and methods of computer algorithms to simulate human intelligence.Base on that the technology and practical systems are developed to help people solve problems.So machine learning,as a very important branch of AI,is involved in a wide range of fields.It has evolved from the initial symbol learning to statistical machine learning,and has been applied to many scientific and industrial fields from theory to the practical applications.Molecular absorption energy is produced by the transition of electrons,reflecting the electronic properties and the internal structure,and it is an important optical property of excited states.It is possible to analyze,measure and deduce the component and structure of molecules if knowing the value of absorption energies and intensities.It is of great importance for designing optical materials,such as nonlinear optical materials,photovoltaic materials of solar cells and so on.Therefore,how to calculate or predict the molecular absorption energy accurately and efficiently,especially for large molecules,is worth to explore.in the past several decades,quantum chemical calculations have been very effective in various chemical fields.In the last decade,combining quantum chemistry calculations with machine learning algorithms has gained remarkable achievements.The time-dependent density functional theory(TDDFT)is one of the best choices to calculate the excited state of molecules because of its high efficiency and adaptability of various molecular systems,but the calculation accuracy and the size of molecular systems still need to be improved.An attempt is made in this thesis to apply ensemble learning methods to correct the calculation results of TDDFT.In this thesis,433 organic molecules,including 276 dye molecules,are investigated.Firstly,(TD)B3LYP in(TD)DFT is used with basis sets STO-3G,6-31G(d)and 6-311G(d,p),to calculate the structure of the ground state and the spectrum of the excited state to obtain three data sets;then use the sample set partitioning based on joint x–y distances(SPXY)and Kennard-Stone to divide data;afterwards,using Pearson,sequential forward selection(SFS)and least absolute shrinkage and selection operator(LASSO)algorithms to select features;finally,regression models are built by three single machine learning methods,support vector machine(SVM),general regression neural network(GRNN)and extreme learning machine(ELM),and two ensemble machine learning methods,gradient boosting decision tree(GBDT)and random forest(RF).With all of the efforts,a stable and effective model for predicting molecular absorption energy with high precision is obtained,which improves the accuracy of TDDFT the most among the established models.In this thesis,we have pursued the best machine learning model for correcting absorption energies calculated by TDDFT.The ensemble learning algorithm RF model based on results calculated by(TD)B3LYP with STO-3G basis set showed best results among all of models.This model corrects the root mean square error(RMSE)of the molecular absorption energy of TDDFT results from 0.97 eV to 0.14 eV,and the mean absolute error(MAE)from 0.71 eV to 0.11 eV.The prediction errors of model on the basis of other two basis sets are also reduced significantly,and ensemble learning methods outperform single machine learning methods.This suggests that effective correction models for predicting the absorption energy of molecules can be successfully established by the combination of ensemble learning and quantum chemical methods.