To confront current market changes, Set-Based Design (SBD) should help carmakers improve time, cost and quality in early design phase. In a development process using ever more computer experiments, building predictive models of the solution space is key to implementing SBD. The present article proposes a new algorithm to build high quality predictive metamodels efficiently and tests it on several benchmark problems. Promising results are obtained. We believe such models could be used for several purpose in design: implementing SBD, optimisation and generating feasible concepts, among others.