In this paper, a new methodology combining the Polynomial Chaos Expansions (PCE) and Nondominated Sorting Genetic Algorithm II (NSGA-II) is proposed to design inductive power transfer (IPT) systems. The relationship between the outputs (mutual inductance and leakage flux density) and structural parameters (coils and ferrite shape, size, thickness) is expressed by a PCE metamodel. Then, this metamodel is set as the objective function and the NSGA-II is used to obtain an optimal design with a trade-off between these outputs. In addition, a sensitivity analysis directly available with this PCE metamodel shows how the structural parameters influence the outputs. This method successfully provides the Pareto front and strongly saves the computation time.