This work focuses on a new approach to enhance the results obtained by a multi objective ptimization, especially tailored for computationally hard CAE models. The methodology combines Self-Organising Maps and Response Surfaces with evolutionary multi-objective optimization algorithms. The challenge is to improve the results spending only few extra computations. First, an optimization with an evolutionary algorithm explores the wide range of the possible solutions. Then, Self-Organising Maps has been used to detect local correlations, this way circumscribing the scope of the search in the design parameter space. After the definition of new bounds for the input variables, a new Design of Experiment was performed on the CFD model, and then interpolated with Response Surface Modeling techniques. Then, a virtual optimization has been applied to these meta-models, and the most interesting virtual designs were validated by means of new CFD simulations. Beyond the improvements with regard to the initial design, the methodology guaranteed a reduction of the computational resources needed to obtain such results, compared to a full direct optimization. Self-Organizing Maps allowed for selecting the most promising area for the objectives and gained new insights to the relations between input and output variables.