Metamodel-Based Optimization of the Labyrinth Seal

• Main Authors: Rulik Sebastian, Wróblewski Włodzimierz, Frączek Daniel
• Format: Article
• Language: English
• Published: Polish Academy of Sciences 2017-03-01
• Series: Archive of Mechanical Engineering
• Subjects: labyrinth seal; metamodel optimization; neural network; genetic algorithms; evolutionary algorithms; CFD optimization
• Online Access: http://www.degruyter.com/view/j/meceng.2017.64.issue-1/meceng-2017-0005/meceng-2017-0005.xml?format=INT

The presented paper concerns CFD optimization of the straight-through labyrinth seal with a smooth land. The aim of the process was to reduce the leakage flow through a labyrinth seal with two fins. Due to the complexity of the problem and for the sake of the computation time, a decision was made to modify the standard evolutionary optimization algorithm by adding an approach based on a metamodel. Five basic geometrical parameters of the labyrinth seal were taken into account: the angles of the seal’s two fins, and the fin width, height and pitch. Other parameters were constrained, including the clearance over the fins. The CFD calculations were carried out using the ANSYS-CFX commercial code. The in-house optimization algorithm was prepared in the Matlab environment. The presented metamodel was built using a Multi-Layer Perceptron Neural Network which was trained using the Levenberg-Marquardt algorithm. The Neural Network training and validation were carried out based on the data from the CFD analysis performed for different geometrical configurations of the labyrinth seal. The initial response surface was built based on the design of the experiment (DOE). The novelty of the proposed methodology is the steady improvement in the response surface goodness of fit. The accuracy of the response surface is increased by CFD calculations of the labyrinth seal additional geometrical configurations. These configurations are created based on the evolutionary algorithm operators such as selection, crossover and mutation. The created metamodel makes it possible to run a fast optimization process using a previously prepared response surface. The metamodel solution is validated against CFD calculations. It then complements the next generation of the evolutionary algorithm.