Least Square Support Vector Machines (LS-SVM), a new machine-learning tool has been employed for developing data driven models of non-linear processes. The method is firmly rooted in the statistical learning theory and transforms the input data to a higher dimensional feature space where the use of appropriate kernel functions avoid computational difficulty. Further, a pattern search algorithm, which explores multiple directions and utilizes coordinate search with fixed step size, is employed for selecting optimal LS-SVM model that produces a minimum possible prediction error. To show the efficacy and efficiency of the fully automated pattern search assisted LS-SVM methodology, we have tested it on several benchmark examples. The study suggests that proposed paradigm can be a useful and viable tool in building data driven models of non-linear processes. © 2005 Institution of Chemical Engineers.