Study On Physical Properties Of Materials By Support Vector Regression And Feature Selection

In 1960s, Vapnik and co-workers proposed statistical learning theory (SLT), which is a statistics theory for the analysis of a small-sample database. Support vector machine (SVM) is based on structural risk minimization principle and statistical learning theory and well-known as a paragon to learn in case of a small number of samples. Compared with Multivariate Nonlinear Regression (MNR), SVM has a stronger learning and generalization ability and has been successfully applied to solve classification and regression problems in many fields.In this thesis, sensitivity analysis is applied to select the optimal subset firstly. Based on optimal subset, SVR was utilized to fit and predict superconducting transition temperature for superconducting elements. And then the prediction results of support vector regression (SVR) are compared with those achieved by MLR without optimal subset. In addition, SVR was directly utilized to estimate physical properties of materials, including the softening point of bitumen in producing, superconducting transition temperature Tc for superconductors of doping MgB2 system and adhesive strength for plasma spraying FeO4 powder in Ni/Al bottom coat. At the same time,the optimal synthesis parameters were searched out by particle swarm optimization (PSO) based on the optimal models and factor analysis was also conducted.The outline of this thesis is as below:â‘ The current methods of feature selection were reviewed. The advantages and disadvantages of several algorithms including Grey Relational Analysis(GRA), Sensitivity Analysis(SA), Signal-to-Noise Ratio (SNR), Entropy Criterion (EC), Genetic Algorithm (GA), Principal Component Analysis (PCA), Independent Component Analysis (ICA), PSO and Simulation Annealing (SA), were introduced.â‘¡The principle,algorithm,implementation and development of SVR were described in detail. The regression principles of popular regression methods were reviewed briefly,such as General Regression Neural Network (GRNN), MLR, Ridge Regression (RR), etc.? ? ? ?â‘¢Based on the experimental datasets of superconductor transition temperature of superconducting elements, sensitivity analysis was applied to select the optimal subset of the descriptors. SVR was employed to modeling and predicting the Tc value for the superconducting elements. Then the prediction results of SVR are compared with those achieved by MLR; According to the experimental dataset on different superconductors of doping MgB2 system, SVR estimated the Tcs of the superconductors of doping MgB2 system by using their topological descriptors; In terms of the experimental dataset on bitumen in the production, SVR predicted the soft point of bitumen effectively; Based on the experimental dataset on adhesive strength of FeO4 coating, SVR was used to forecast the adhesive strength to plasma spraying FeO4 power in Ni/Al bottom.â‘£The optimal process parameters were searched out by PSO based on the established SVR model, and then the influence of multifactor were further analyzed.?The studies of above demonstrated that sensitivity analysis is helpful to improve the prediction precision. Furthermore, the generalization ability and accuracy of SVR were all superior to that of MLR and MNR methods. The results suggest that SVR is an effective and powerful technique; it may be further developed to be a potential application tool in the field of materials, such as material computer assistant design and material process parameter optimization, etc.