Applications of improved SVM framework in modeling in mechanics

Abstract

The problem of empirical data modeling is pertinent to several mechanics domains. Empirical data modelling involves a process of induction to build up a model of the system from which responses of the system can be deduced for unobserved data. Machine learning tools can model underlying non-linear function given training data without imposing prior restriction on the type of function. In this paper, we show how Support Vector Machines (SVM) can be employed to solve design problems involving optimizations over parametric space and parameter prediction problems that are recurrent in engineering domain. The problem considered is diffuser design where the optimal value of pressure recovery parameter can be obtained very efficiently by SVM based algorithm even in a large search space. In addition, locating the position of points on a string vibrating in a damped medium serves as an appropriate prediction problem. A gridsearching algorithm is proposed for automatically choosing the best parameters of SVM, thus resulting in a generic framework. The results obtained by SVM are shown to be theoretically sound and a comparison with other approaches such as spline interpolation and Neural Networks shows the superiority of our framework.