Aiming to speed up and improve the aerodynamic design of compressor blades, an automatic Optimizer AutoOpti was developed at our institute. Since the aerodynamic design always demands a compromise between several objectives (e.g. efficiency, total pressure ratio, mass flow, flow turning, surge margin, ...), that need to be considered at all relevant operating points in the operating range of the compressor, in developing the AutoOpti one priority is its capability of simultanously optimizing several objective functions (multi-objective Optimization).
The core of the optimization algorithm is an evolutionary strategy, which according to nature's example, allows for an improvement in certain objective functions or desired features by the iterative use of the operators selection and inheritance. The rating of single individuals is done via so-called Pareto rank, that yields a measure of the fitness of a member by ranking it among the already evaluated data base. In the ranking all objective functions are considered.
In addition to these classical characteristics a large number of verifications were implemented in AutoOpti, all of them serving the goal of accelerating the optimization process. The core optimizer for example was parallelized so that it can run on a computer cluster. To guarantee optimal use of all processors taking part in the optimization, the processes are run asynchronously. Given these hardware-related methods of increasing the pace of the process, the main speed up of the evolutionary strategy inside AutoOpti (software-related) is based on the use of efficient meta models (or response surfaces). This means that already evaluated geometries, consisting of design parameters and objective function values, are read in from a data base and approximated and/or interpolated by neural networks and/or Kriging models. This step is followed by a temporary optimization on the response surfaces and then only the best data sets on the response surfaces are sent to the numerically costly process chain:
Blade Generation - Grid Generation - Geometric Constraints - Navier-Stokes Solver - Post - Objective Function Evaluation
In this sense the response surfaces constitute a new inheritance operator creatingpromising geometries.
The images show an automatic optimization result of a counter-rotating fan (left: initial fan, right: efficiency-optimized fan). Clearly, the local backflow area after the shock of the first rotor is avoided by complex stacking, improving the isentropic efficiency of the entire stage by about 1.5%.
The AG Turbo co-financed department 3D Optimization cooperates with thefollowing partners: