An aerodynamic optimization system for centrifugal fans consisting of an impeller and a diffuser has been developed. A fan's three-dimensional shape is represented by a non-uniform rational B-spline curve, which enhances the freedom of the blades' leading and trailing edge profiles. The fan's aerodynamic performance is predicted using commercial software, "STAR-CD." The optimization system uses a new algorithm that is a hybrid of a simulated annealing algorithm and a neural network. The hybrid algorithm not only can learn the design-space to predict better solutions, but also can analyze the space's statistical characteristics, such as its sensitivity and non-linearity. Applying this system to centrifugal fan design problems, we obtained novel fan shapes in which the leading edge of the impellers and the trailing edge of the diffusers were bent unconventionally. Important design parameters were simultaneously revealed using design-space analysis.