Session: 36-04 Neural-Network based approaches (3)

Paper Number: 128792

128792 - Generative Model Based Parameterization for More Efficient Aerodynamic Optimization of Non-Axisymmetric Endwall 

With the ever-increasing blade loadings, the technique of non-axisymmetric endwall contouring (NEC) is proposed, which has been proved its effectiveness in reducing the secondary loss of turbine cascades. However, the parameterization of NEC would face the following problems, i.e., a tradeoff among the issues such as sample variability in design space, the number of design variables and the proportion of abnormal shapes as well, should be made, particularly when the spline-based parameterization techniques such as Non-Uniform Rational B-Spline (NURBS), etc. are used. To address the above issue and thus achieve more efficient NEC optimization, a generative model based parameterization method is proposed in this paper. Specifically, to build a large design space that contains sufficient variability of NEC shapes, the samples of a set of conventional classical NEC parameterization methods are collected, which are combined by representing them with scatter coordinates in a uniform format. Then, to reduce the number of variables to the minimum extent, the Least-Square GAN (LS-GAN) is used to learn the intrinsic dimensions of parameterized NEC design space, which maintains the generation capability of vanilla GAN while improving the training stability by using modified adversarial loss terms. Additionally, to ensure the smoothness of the NEC profiles and thus avoid abnormal shapes, a layer of NURBS is incorporated into the neural network architecture of LS-GAN. To verify the effectiveness of the proposed method, this proposed model was used for the NEC parameterization of a high-pressure turbine stage with 8 variables. Through visualization by using t-SNE plot, the shape variability of LS-GAN based NEC design space is shown to be far better than conventional approaches. Moreover, the proportion of abnormal NEC shapes is reduced to almost 0%. Furthermore, an aerodynamic optimization of the non-axisymmetric endwall is carried out by using LS-GAN and other compared NEC parameterization approaches. It is shown that the optimization with the generative model based parameterization method can achieve significantly better aerodynamic performance with the same limit on the number of performance evaluations. While achieving similar optimized solutions, the computational cost of generative model based approach can be reduced by at least 50%. And a detailed analysis is also carried out to validate the performance of optimized solutions. With the above, the effectiveness of our proposed generative model based parameterization method is well demonstrated.

Presenting Author: Cunxi Li Xi'an Jiaotong University

Presenting Author Biography: Cun xi is a master student in Xi'an Jiaotong University, which is majored in generative model based parameterization and optimization methods for turbomachinery design

Authors:

Cunxi Li Xi'an Jiaotong University
Liming Song Xi'an Jiaotong University
Zhendong Guo Xi’an Jiaotong University
Zhao Yang Xi'an Jiaotong University
Jun Li Xi'an Jiaotong University
Zhenping Feng Xi'an Jiaotong University