An adaptive optimization strategy based on mixture of experts for wing aerodynamic design optimization
Résumé
In the field of aircraft design, the last few decades have focused on the iterative improve- ment of conventional tube-and-wing designs to reduce cost, noise, and emission. Never- theless, the growing expectation in terms of environment impact for the next generation of aircraft pushes for more radical changes in the design. For unconventional aircraft configurations, the need to integrate more accurate data coming from higher fidelity analysis earlier in the design process becomes more and more necessary. However, high-fidelity tools require long computation times and usually are associated with high-dimensional problems, both in terms of design variables and constraints. Therefore, these optimizations are often done at higher computational cost (gradient-based algorithms) in order to decrease the number of necessary function evaluations. In addition, the use of the adjoint method is often implemented to accurately and efficiently compute derivatives for large numbers of design variables. At the same time, new methods have been investigated to obtain opti- mized configurations at a reasonable computational cost. The work presented in this paper focuses on SEGOMOE algorithm, a solution to tackle this kind of optimization process of complex design problem through the use of an enrichment strategy approach based on mixture of experts surrogate models. Two aerodynamic shape optimization test cases, derived from cases developed by the Aerodynamic Design and Optimization Discussion Group (ADODG) are addressed: one with a single global minimum, and another one with several local minima. Both problems are nonlinearly constrained problems that involve a large number of design variables. Results are compared to gradient-based optimizers. A hybrid approach combining the advantages of both SEGOMOE and gradient-based optimization is proposed and evaluated to reduce the number of function evaluations and to ensure the convergence to the global optimum.