Study On The Prediction Of Hazardous Properties For Chemicals Based On The QSPR Models

Hazardous characteristic data of chemicals are the basis of chemical safety designs and risk assessments.At present,experimental test is the main method to obtain all kinds of hazardous characteristic data,but this method often has many risks and limitations.Therefore,it is urgent to realize the theoretical predictions of dangerous properties of substances by means of theoretical calculations.Quantitative structure-property relationship(QSPR)can be used to seek the internal quantitative relationships between macro properties and micro structures based on the principle that the properties and molecular structures of substances are closely related to each other.If it is possible to predict the hazardous properties of substances by building QSPR models,we can solve the problem of data shortage and provide reliable and accurate properties' data for chemical designs and risk assessments in a relatively short time and under the premise of reducing the consumption of manpower,financial and material resources.Moreover,when the new substance is not synthesized,the validated QSPR model can be used to determine whether the target substance has the expected properties,thus shortening the synthesis time and reducing the exploration frequencies.Based on the principle and methodology of QSPR,in this paper,the internal quantitative relationships between self-accelerating decomposition temperature and minimum ignition energy and molecular structures and the internal quantitative relationships between flammability limits of mixtures and molecular structures were studied and accordingly,several QSPR models were successfully constructed.Necessary verification and evaluation of the models were also carried out and the application domains of the models were discussed.In the QSPR study of the self-accelerating decomposition temperatures for pure compounds,two kinds of molecular descriptors were calculated based on atomic type electrical topological state index and Gaussian 09 software and then they were used to build the models with two modeling methods,multiple linear regression(MLR)and support vector machine(SVM).Totally,four QSPR models were constructed.The parameters of the MLR model based on the atomic type electrical topological state index are not satisfied with the verification criteria so this model was given up and no further research was performed.Although the parameters of the SVM model based on the atomic type electrical topological state index are satisfied with the verification criteria,the residuals plot of it is not qualified.The MLR and SVM model based on Gaussian 09 software are qualified and the performance of the SVM model are better than the MLR model in the goodness-of-fit,the internal robustness,and the external predictive ability.But the SVM model is a black-box model while the MLR model gives a clear mathematical expression.Therefore,the MLR model is more convenient in the application process.In a word,the above two models all have qualified fitting ability,stability and prediction ability and have their advantages and disadvantages.For application,the more suitable model should be selected according to the practical situation.In the QSPR study of the minimum ignition energy for pure compounds,two kinds of molecular descriptors were calculated based on atomic type electrical topological state index and Gaussian 09 software and then they were used to build the models with two modeling methods,MLR and SVM.Totally,six QSPR models were constructed.Among them,the parameters of the MLR and SVM models based on the atomic type electrical topological state index(61 samples and 23 atomic types)do not meet the verification criteria.The causes of this phenomenon were analyzed and the models were reconstructed.The new MLR and SVM models based on the atomic type electrical topological state index both meet the requirements and have higher accuracy.The stability of the SVM model is slightly better than that of the MLR model,but the prediction ability of the MLR model is stronger than that of the SVM model.The biggest advantage of these two models is that they do not need any expensive software but the scope of application of this two model is relatively small.The parameters of the MLR and SVM models based on Gaussian 09 software both meet the requirements of the verification standard,which proves the superiority of the molecular descriptors based on Gaussian 09 software.The stability of the MLR model is better while the prediction ability of the SVM model is stronger.In a word,the above four models all have qualified fitting ability,stability and prediction ability and for application,the most suitable model should be selected according to the practical situation.In the QSPR study of the flammability limits for mixtures,the molecular descriptors of the pure compounds were calculated using Gaussian 09 software and then the molecular descriptors of the mixtures were calculated by the method of weighted molar mass percentage.Then for lower flammability limits and upper flammability limits,the linear(MLR)and the nonlinear(SVM)models were constructed respectively.In order to ensure the comprehensiveness and reliability of model verification,the sample set partition methods of "Points out","Mixtures out"and "Compounds out" were adopted at the same time during the model construction process.During the verification of the linear model,the value of(R2-R02)/R2 of the"Compounds out" model for lower flammability limits is equal to 0.136 while the value of(R2-R02)/R2 of the "Mixtures out" model for upper flammability limits is equal to 0.262,which do not meet the requirements.Therefore,the MLR modeling fails.During the verification of the nonlinear model,the parameters of the "Points out" model,the "Mixtures out" model and the "Compounds out" model for lower flammability limits and upper flammability limits all meet the requirements.Among them,"Points out" models and "Mixtures out" models have very superior fitting ability,internal stability and external prediction ability,and the "Compounds out"models,which are the most strict external verification models,also far exceed the verification standard,which indicates the reliability and accuracy of SVM model.In the application of the model,it is not necessary to know the flammability limit of the pure substances and only the molecular structures and the proportions of the pure compounds which constitute the mixture need to be provided.Finally,based on the accurate and reliable QSPR models mentioned above and the mature QSPR models that can be consulted in the literature,the QSPR prediction software for hazardous characteristics is designed and developed using the C#(C sharp)language through the Microsoft Visual Studio(VS)2015.With the further development of QSPR research,the new QSPR prediction model can be gradually added.Thus,a database of QSPR prediction models between chemical structures and hazard characteristic parameters will be gradually formed and through this database,the related personnel can obtain the dangerous characteristic parameters of substances quickly and easily,which makes the safety production of chemical industry enter a new stage.