One-shot methods for aerodynamic shape optimization

  • One-Shot-Verfahren in der aerodynamischen Formoptimierung

Özkaya, Emre; Gauger, Nicolas Ralph (Thesis advisor)

Aachen (2014)
Dissertation / PhD Thesis

Aachen, Techn. Hochsch., Diss., 2014


This thesis is concerned with the aerodynamic shape optimization based on the one-shot strategy, which aims at performing an optimization simultaneously with the simulation process. Combined with the consistent discrete adjoint method based on Automatic Differentiation (AD), one-shot methods enable performing a shape optimization at a small multiple of the time required for a single flow simulation. In the present work, we first investigate the preconditioning techniques and constructive conditions that are required to satisfy the contractivity of the coupled one-shot iterations. Further, an analysis concerning the quantification of the retardation rate, which is an indication of the computational cost of the overall optimization compared to a single flow simulation, is presented. We also discuss the implementation aspects concerning the aerodynamic design and optimization chains, and present a review of the shape parameterization techniques commonly used in aerodynamic shape optimization. An assessment of the most common sensitivity evaluation methods with respect to accuracy, computational cost and robustness is performed. Among all the methods, AD based consistent discrete adjoint method is the most robust and flexible method, which enables computing accurate sensitivity information at a fixed computation cost independent of the number of design parameters used in the optimization. The application of this strategy as well as the AD techniques to generate adjoint solvers from the in-house state equation solvers are discussed in detail. Furthermore, advanced techniques to improve the computational performance are introduced. In the last section, first validation results obtained from the developed adjoint Euler and Navier-Stokes solvers are presented. The run-time and memory requirements of the adjoint solvers using different grid levels are provided to demonstrate the efficiency of the chosen adjoint methodology. Then, optimization results that are obtained by applying the one-shot method to three different airfoil optimization scenarios are presented. Furthermore, a comparison between the one-shot method and the nested quasi-Newton method is performed using one of the test cases. The run-time measurements are presented to prove the efficiency of the one-shot method.