Design Optimization of Internal Flow Devices

Methods of Computational Fluid Dynamics (CFD) have matured to a stage, where it is possible to gain substantial insight into uid ow processes of engineering relevance. However, the motives of uid dynamicists typically go beyond improved understanding to the de nitive aim of improving the performance of the engineering systems in consideration. It is in recognition of these circumstances that the present thesis investigates the use of automated design optimization methodologies in order to boost the power of CFD for engineering design purposes. Optimum design problems require the merit or performance of designs to be measured explicitly in terms of an objective function. At the same time, it may possibly be required that one or more constraint functions should be satis ed. To describe allowable variations in design, a shape parameterization using basic geometrical entities such as straight lines, arcs, and spline curves is employed. The developed tools have been implemented in the computer-aided design environment of the Optimum DESign SYstem ODESSY. Two fundamentally di erent optimization techniques were considered. Firstly, iterative search optimization based on design sensitivities, i.e. gradients of the objective and constraint functions w.r.t. the design variables. Secondly, optimization via the construction of response surface approximations has been considered. The paramount complication encountered in developing tools for derivative-based optimization is to devise robust and e cient ways of computing sensitivities of the ow eld. The simplistic overall nite-di erencing approach by repeated analyses is computationally expensive and needs tuning of perturbation sizes to be accurate. A more e cient semi-analytical method based on direct di erentiation of the discrete equations of uid motion, i.e. the Reynolds-averaged NavierStokes equations, has therefore been implemented. This implementation was rather cumbersome and prone to errors, since it involved the manual calculation of the Jacobian matrix of derivatives for the system of uid equations. In contrast, response surface approximations were found substantially more straight-forward to apply. However, the cost of building response surface approximations appear prohibitive when the number of design variables exceed about ten. In contrast, search optimization using semi-analytical sensitivities is an approach well-suited for large-scale optimization problems. In the choice of design cases presented in the thesis, it has been attempted to cover a representative spectrum of problem types. This includes loss minimization in di user ows, tailoring of velocity distributions by design of guide vanes, and optimization of a valve with consideration of structural loads imposed by the uid ow.