This paper illustrates how multi-fidelity metamodels can be efficiently applied to save time and costs in parametric design optimization, which normally requires simulating numerically a large number of designs. Datasets of high fidelity (HF) and middle-low fidelity (LF) simulations, obtained for instance from computational models solved by grid discretizations of different accuracy, can be used together to feed the surrogate model, improving the accuracy of the response function prediction and reducing the overall computational cost at the same time. The methodologies proposed in this paper include adaptive Design of Experiments algorithms to define the optimal dataset of design simulations, and efficient multi-fidelity surrogate methods for scalar fields (Cokriging) and vector fields (Reduced Order Models). All the methods are tested and applied to CFD test cases.
