Session: 36-05 Robust Design and response surface methods (1)

Paper Number: 121848

121848 - Twofold Adaptive Design Space Reduction for Constrained Bayesian Optimization of Transonic Compressor Blades 

As turbomachinery designs become more complex, shape optimization tasks contain more constraints and design variables. Bayesian optimization (BO) is a class of adaptive surrogate model-based methods for global optimization, which can efficiently utilize a small budget of high-fidelity evaluations. This is beneficial for the commonly non-convex problems and expensive simulations in turbomachinery design. BO scales well with an increasing number of constraints but suffers from a hampered convergence rate for a larger number of design variables, because the design space volume grows exponentially with its dimension.

To improve the scalability, we build on existing adaptive design space reduction methods for BO: PCA-BO, TRIKE, and TuRBO. PCA-BO operates on a subspace spanned by a reduced number of principal components and thus reduces the design space dimension. We combine TRIKE and TuRBO in a trust region-based BO (TR-BO), which localizes the search by reducing the design variable ranges. Moreover, we extend both approaches to handle constrained problems and parallel high-fidelity evaluations, which is required for industrial applications. Applying PCA-BO and TR-BO sequentially, we profit from their respective advantages over the course of the optimization.

We assess the performance of our hybrid algorithm by comparing it to a vanilla BO on an analytical and an industrial problem. First, we minimize a 40-dimensional constrained Rastrigin function. Second, we optimize the blade shape in a single stage of an aircraft engine high-pressure compressor. More precisely, we seek for maximum isentropic efficiency by optimizing 55 design variables subject to several constraints. The empirical results show that PCA-BO enables a higher convergence rate in the initial optimization phase, while TR-BO allows for further improvements in the later iterations. In combination, they yield better designs for a fixed budget of high-fidelity evaluations, with an especially large improvement for smaller budgets. The proposed approach has the potential to make the benefits of BO available for the optimization of even higher-dimensional constrained problems, including multi-stage turbomachinery configurations.

Presenting Author: Lisa Pretsch Technical University of Munich

Presenting Author Biography: Lisa Pretsch received her M.Sc. in Mechanical Engineering from the Technical University of Munich (TUM) in 2021 with distinction. The focus of her studies was on computational methods for aerodynamics, structural mechanics, and acoustics. In the same year, she joined the TUM School of Engineering and Design as a research associate and doctoral candidate. Her research topic are surrogate-based optimization methods for aircraft engine blade design. The work is part of a research project in cooperation with MTU Aero Engines and the German Aerospace Center (DLR), funded by the free state of Bavaria.

Authors:

Lisa Pretsch Technical University of Munich
Ilya Arsenyev MTU Aero Engines
Elena Raponi Leiden University
Fabian Duddeck Technical University of Munich