Abstract

Aerodynamic shape optimization of gas turbine blades is a very challenging task, given e.g. the flow complexity, the stringent performance requirements, the structural and manufacturing constraints, etc.... This work addresses the challenge by automating the optimization process through the development, implementation and integration of state-of-the-art shape parametrization, numerical optimization methods, Computational Fluid Dynamics (CFD) algorithms and computer architectures. The resulting scheme is successfully applied to single and multi-point aerodynamic shape optimization of several cascades involving two-dimensional transonic and subsonic, viscous and inviscid flow in compressor and turbine cascades. The optimization objective is to achieve a better aerodynamic performance, subject to aerodynamic and structural constraints, over the full operating range of gas turbine cascades by varying the blade profile. That profile is parameterized using a Non-Uniform Rational B-Splines (NURBS) representation, which is flexible accurate and capable of representing the blade profiles with a relatively small number of control points for a given tolerance. The NURBS parameters are then used as design variables in the optimization process. The optimization objective is determined from simulating the flow using an in-house CFD code that solves the two-dimensional Reynolds-Averaged Navier-Stokes (or Euler) equations using a cell-vertex finite volume method on an unstructured triangular mesh and turbulence is modeled using the Baldwin-Lomax model. To save computing time significantly, Artificial Neural Network (ANN) is used to build a low fidelity model that approximates the optimization objective and constraints. Moreover, to reduce the computing wall-clock time, the optimization scheme was parallelized on an SGI ALTIX 3700 machine using Message Passing Interface (MPI), resulting in a parallelization efficiency of almost 100%. Different numerical optimization methods (genetic algorithm, simulated annealing and sequential quadratic programming) were developed, tested and implemented for the different parts of this work. The present choice of objective function and optimization methodology results in a significant improvement in performance for all the cascades that were optimized, without violating the design constraints. The use of ANN results in a ten-fold speed-up of the design process and the scheme parallelization allows for further reduction of the wall-clock time