Aerodynamic Shape Optimization and Knowledge Mining of Centrifugal Fans Using Simulated Annealing Coupled With a Neural Network

An aerodynamic shape optimization method suitable for “inexpensive” centrifugal impellers and diffusers has been developed. The shapes are parameterized using non-uniform rational B-spline curves with special attention being paid to the blade’s edge profiles. A hybrid algorithm combining simulated annealing and a neural network is employed for collaborative optimization. The simulated annealing and neural network take turns in controlling the optimization processes, not only for maximizing the efficiency of global exploration, but also for minimizing the risks of automation failures or of reaching an incorrect optimum. A statistical analysis was also conducted using the neural network to extract design knowledge. By applying the proposed method to a centrifugal impeller and diffuser design problem, we obtained innovative shapes for the leading edge of the impeller and the trailing edge of the diffuser. Important design parameters related to the new shapes were identified through the design space analysis.