The problem of the automatic design optimization for multistage axial flow turbines is considered and a design strategy based on a 3D Navier-Stokes solver and a RSM (Response Surface Method) approach is described. A multi-objective optimization code based on non-dominated sorting genetic algorithm (NSGA-2) is used to drive the optimization process in order to maximize the specific power while keeping the massflow rate constrained. In the present work the meridional channel is kept unchanged while for each blade the spanwise distribution of the profile restaggering is considered together with the inclusion of compound lean. The performance from the multistage turbine for the optimization loop are obtained from surrogate models built through a set of artificial neural networks. The neural networks are trained and tested using large DoEs and are not updated during the optimization process. This aspect is considered important to guarantee that the optimization converges to an optimum. The use of the 3D flow solver with coarse meshes in order to validate large DoEs in short times is discussed in some details. The above strategy has been applied to a four stage axial turbine from the open literature.