Hierarchical distributed metamodel‐assisted evolutionary algorithms in shape optimization

Abstract

AbstractIn aerodynamic shape optimization, the availability of multiple evaluation models of different precision and hence computational cost can be efficiently exploited in a hierarchical evolutionary algorithm. Thus, in this work the demes of a distributed evolutionary algorithm are ordered in levels, with each level employing a different flow analysis method, giving rise to a hierarchical distributed scheme. The arduous task of exploring the design space is undertaken by demes consisting the lower hierarchy level, which use a low‐cost flow analysis tool, namely a viscous–inviscid flow interaction method. Promising solutions are directed towards the higher level, where these are further evolved based on a high precision/cost evaluation tool, viz. a Navier–Stokes equations solver. The final, optimal solution is obtained from the highest hierarchy level. At each level, metamodels, trained on‐line on the outcome of evaluations with the level's analysis tool, are used. The role of metamodels is to allow a parsimonious use of computational resources by filtering the poorly performing individuals in each deme. The entire algorithm has been implemented so as to take advantage of a parallel computing system. The efficiency and effectiveness of the proposed hierarchical distributed evolutionary algorithm have been assessed in the design of a transonic isolated airfoil and a compressor cascade. Remarkable superiority over the conventional evolutionary algorithms has been monitored. Copyright © 2006 John Wiley & Sons, Ltd.