Study On The Correction Model Of Non-covalent Interaction Based On Ensemble Learning

Intelligent computing is the method of simulating basic laws of nature and intelligent biosystems.It can simplify complex problems through specific models,and nowadays is widely applied in many research fields.In this thesis,the intelligent method is used to establish an ensemble learning model to correct the accuracy of the non-covalent interaction by quantum chemical calculations.Non-covalent interaction(NCI)is a complicated weak interaction among molecules.It is closely related to many physical and chemical phenomena,but it cannot be calculated by explicit equations.In order to get more accurate calculation results and save computational costs,we combine machine learning methods with quantum chemical methods to improve the accuracy of NCI through ensemble models.In this thesis,we use four types of density functional theory(DFT)methods,M062 X,B3LYP,PBE and ?b97XD with four kinds of basis sets,6-31G*,6-31+G*,AUG-cc-pvdz and cc-pvdz to calculate the NCI of molecule systems in vacuum.On one hand,the regression ensemble model is used to correct the calculated value of density functional theory to improve the accuracy of DFT calculations,and the root mean square error after correction is significantly reduced.The best result is 0.22kcal/mol reduced from 0.94kcal/mol.On the other hand,we build an ensemble model based on three classes of dominant interactions,hydrogen bond,dispersion force and mixed interaction.The highest classification precision can reach 98% and the lowest can reach 94%.In both regressions and classifications,the ensemble models show good generalization,strong prediction ability and stability,and can effectively correct the calculation of NCIs.Additionally,through comparing the results of regressions and classifications,the most important features related to NCIs are explored,which is quite useful for both theoretical and experimental investigations of NCIs.