Optimization of a centrifugal compressor impeller using genetic algorithm coupled with artificial neural networks

Nowadays, obtaining a state of art design for any kind of turbo machinery is clearly an issue of optimization and efficient methodology, since the difference between a decent and unacceptable design is determined by very small changes to the design. It can be very though or demanding task to fulfill the performance criteria such as range and efficiency with trial and error. Consequently, automatization of the design process redounds significant amount of time and effort alongside ensuring a certain level of confidence. Taking into consideration the aforementioned, a design methodology for the centrifugal compressors has been developed using on the shelf commercial tools and put to use. In the scope of this work, impeller of an experimental compressor of a micro turbojet engine will be utilized as baseline for optimization. Only impeller part of the compressor is the concern for this study, however the same methodology can be applied to any part of a turbomachinery, given the adequate amount of time. Results are compared to demonstrate the power of the methodology over a trial and error design. Baseline impeller has a pressure ratio of 1:4 with approximate isentropic efficiency of %85. Target for the optimization was at least %1 increase in isentropic efficiency, at the design point, without any change to pressure ratio, mass flow. Optimized impeller is going to replace the baseline and must work with the same baseline diffuser. This means that any increase in the performance parameters of the compressor is due to the geometrical changes in impeller only. Geometry parametrization is the first step and performed using Bezier curves. Any other geometrical parameters are frozen throughout this thesis. The range of geometrical parameters is determined mostly by manufacturing reasons or practical purposes. Rest of the process includes setting up a database automatically with the help of ANSYS Workbench, training an artificial neural network which has been presented by MATLAB as a fast performance prediction tool and a genetic algorithm code for creating new individuals (designs) until optimization criteria are satisfied. Thanks to the developed methodology, baseline impeller's total to total isentropic efficiency has been improved by %1.5 at the end of the study without any significant change to the mass flow ratio and pressure ratio. Furthermore, only internal geometry of the impeller (blade angle distribution) has been modified, meaning that optimized impeller can be retrofitted to the baseline compressor without any inconvenience.