Study On Quantitative Structure Properties Relationship Of Organic Mixtures Based On Artificial Neural Network

A suitable quantitative structure-property relationship(QSPR)model can accurately model the relationship between molecular structure and compound properties.In reality,the experimental value of the mixture was not very accurate.And the steps are complicated and time consuming.Therefore,the QSPR method was used to establish the model of betweeen the mixture’s properties and structure,quantitatively describe the relationship between the two,and make up for the deficiency of the experiment.The specific research contents are:In the first chapter the research methods and steps of QSPR were introduced,and the research progress of QSPR of predicting the properties of organic mixtures was reviewed.In the second chapter,based on the basic principle and steps of QSPR,the QSPR models between hydrocarbon and density or octane number were establish respectively by using multiple linear regression(MLR)and multilayer perceptron artificial neural network(MLP-ANN).Leave one out cross validation(LOO-CV)and external test set validation were used to evaluate the prediction performance and robustness of the two models.The results show that the MLR and ANN model can be used to predict the density and octane number of alkanes,but ANN model is better than MLR model.In the third chapter,four QSPR models that predicted 288 kinds of FP model of binary organic mixture were built using the MLR,stepwise regression(SR)and radial basis function(RBF)ANN methods.All models were assessed with external test set verification and k folds cross validation.The results demonstrate that all four models could be used to predict the flash point value of these mixture samples.Five descriptors with significant correlation with flash point were selected by stepwise regression.Among them,the RBF-ANN model with five descriptors as input variables is the most accurate model.In the fourth chapter,molecular structure descriptors with significant correlation with toxicity were selected by stepwise regression.QSPR models between 106 mixture consisted of 12 kinds of benzene and its derivatives and toxicity were established by MLR and MLP-ANN methods.External validation test set and LOO-CV were used to evaluate the two models.Prediction results show that the ANN model is steady,the generalization ability is strong,the prediction error is small,the prediction effect is satisfactory,and it can be used topredict the toxicity of the mixture of benzene and its derivatives.In the fifth chapter,molecular structure descriptors with significant correlation with toxicity were selected by stepwise regression.QSPR models between 23 mixture consisted of12 kinds of perfluorinated oxygen acid and toxicity were established by MLR and MLP-ANN methods.External validation test set and LOO-CV were used to evaluate the two models.The results show that both MLR and ANN are feasible,and the error of ANN model is smaller and Prediction effect is more higher.