Study On The Physical Properties Of Carbon And Its Composites Materials By Using Support Vector Regression

Carbon and its composites are widely used in various fields due to their excellent electrical properties,mechanical properties and thermal properties.Although the various types of carbon-based materials have been studied and widely used,but the impact of different factors on the physical properties of materials are not very clear.The relationship between the experimental process parameters or the molecular structure and the physical properties of the materials is complex,and it is very important to study and understand the regularity between them and to accurately predict the physical properties of carbon-based materials.In such circumstances,if there is no viable way to guide,you need a lot of manpower,material and financial resources to carry out a large number of experiments to find the right material.How to carry out the experimental design reasonably and analyze the experimental data effectively and reduce the experimental cost becomes an unavoidable scientific and technical problem.Therefore,it is very helpful to establish a stable prediction model for the synthesis of materials with excellent physical properties.In this thesis,the mechanical properties of carbon nanotubes/epoxy composites,the mechanical properties of graphene/poly lactic acid composites,the adsorption of organic compounds by carbon nanotubes,and the solubility of fullerenes in different organic solvents were studied by using support vector regression(SVR)combined with particle swarm optimization.The results compared with other regressive methods.At the same time,factor analysis,process parameter optimization and statistical deviation analysis are carried out based on the established SVR model and so on.The main work is as follows:(1)The experimental process parameters: the volume fraction of epoxy,the length of composite film and the cross-sectional area of composite film were taken as input variables.The tensile strength,elongation and elastic modulus were used as output variables.Support vector regression(SVR)combined with particle swarm optimization(PSO)for its parameter optimization,and integrating leave-one-out cross validation(LOOCV)was employed to construct mathematical model for prediction of the mechanical properties of the 15 carbon nanotubes/epoxy composites.The mean absolute percentage error(MAPE)for tensile strength,elongation and elastic modulus are 3.96%,3.14% and 2.62%,the correlation coefficient(2R)is as high as 0.991,0.990 and 0.997 respectively of SVR-LOOCV model.The predictive value of the SVR model is consistent with the experimental value indicating that the model is valid.Multifactor analysis is adopted for investigation on the significances of each experimental factor and their influences on the mechanical properties.Sensitivity analysis shows that the volume percentage of Epoxy has the greatest effect on tensile strength,and the length has great influence on elongation and elastic modulus.(2)The experimental process parameters: the percentage of graphene,the temperature,the experiment time and the stirring speed were taken as the input variables,and the tensile strength was used as the output variable.support vector regression(SVR)combined with particle swarm optimization(PSO)for its parameter optimization,and integrating leave-one-out cross validation was employed to establish a mathematical model for prediction of the tensile strength of 30 poly(lactic acid)/graphene nanocomposites,and the results were compared with the results of response surface method(RSM).The root mean square error(REMS),mean absolute error(MAE)and mean absolute percent error(MAPE)of SVR-LOOCV model for tensile strength(0.276,0.167,0.295%)are smaller than those(0.446,0.368,0.653%)calculated via RSM model,respectively.The correlation coefficients for tensile strength(0.99)achieved by SVR-LOOCV is also bigger than that(0.97)calculated by RSM model.Multifactor analysis and sensitivity analysis showed that temperature has the greatest impact on tensile strength,and the optimum process are volume percentage of Epoxy(0.10%),time(15min),stirring speed(50rpm)and temperature(170?),the tensile strength reaches the maximum 61.63 MPa which is 20000 Pa larger than the experimental value.(3)The molecular descriptor: the molecular polarizability and the mean value of the negative electrostatic potential of the molecular surface as the input variables,the Henry's constant was used as the output variable.Support vector regression(SVR)combined with particle swarm optimization(PSO)for its parameter optimization,and integrating leave-one-out cross validation was employed to establish a model to predict the Henry constants of carbon nanotubes(CNTs)adsorption of 35 volatile organic compounds(VOCs).The prediction performance of SVR was compared with those of the model of theoretical linear salvation energy relationship(TLSER).The RMSE,MAE,MAPE of SVR-LOOCV model for Henry constants(0.080,0.056,1.19%)are both smaller than those(0.507,0.375,9.15%)calculated via TLSER model.The correlation coefficients for Henry constants(0.997)achieved by SVR is also bigger than that(0.874)calculated by TLSER model.Factor analysis showed that the two molecular descriptors had a significant effect on the Henry's constant,but the negative electrostatic potential of the molecular surface has a greater effect on Henry constant.Both the SVR model and the TLSER model conclude that the adsorption of VOCs by carbon nanotubes is mainly due to the interaction between van der Waals forces and hydrogen bonds.Finally,the adsorption mechanism of carbon nanotubes to organic matter is briefly explained.(4)Based on the molecular descriptors: TI2,X1 Sol,FDI and H-052 as input variables and the solubility as the output variables,SVR-LOOCV combined with particle swarm optimization algorithm was used to construct the mathematical prediction of the solubility of fullerene C60 in various organic solvents Model,and the prediction results were compared with the multiple regression analysis(MLRA)model.The test errors,MAE(0.24),MAPE(6.08%)and RMSE(0.32)of SVR are smaller than those of MLRA(0.29,7.36%,0.36),based on the same training set and test set.The correlation coefficients for solubility(0.940)achieved by SVR is also bigger than that(0.903)calculated by MLRA.For SVR-LOOCV model,the MAE(0.023),MAPE(0.62%),RMSE(0.045)for solubility are smaller than those calculated via MLRA(0.29,7.36%,0.36)and SVR(0.24,6.08%,0.32)model,respectively.The correlation coefficients for solubility(0.998)achieved by SVR-LOOCV is also bigger than those calculated by MLRA(0.903)and SVR(0.940)models.The predicted results show that the solubility of C60 is closely related to the molecular descriptor of the organic solvent.The predicted values of the SVR-LOOCV are consistent with the experimental values.The SVR model based on particle swarm optimization is more accurate than the method of multi-variable linear regression,response surface method and theoretical linear solution.SVR modeling prediction method provides a new idea for the design and synthesis of carbon-based materials with ideal physical properties for experimental design.It provides scientific prediction and theoretical guidance for artificial intelligence-assisted design of carbon-based materials,and has a good supporting role on the physical properties of the materials.