TY - GEN
T1 - TWOFOLD ADAPTIVE DESIGN SPACE REDUCTION FOR CONSTRAINED BAYESIAN OPTIMIZATION OF TRANSONIC COMPRESSOR BLADES
AU - Pretsch, Lisa
AU - Arsenyev, Ilya
AU - Raponi, Elena
AU - Duddeck, Fabian
N1 - Publisher Copyright:
Copyright © 2024 by ASME.
PY - 2024
Y1 - 2024
N2 - As turbomachinery designs become more complex, shape optimization tasks involve computationally expensive simulations and many constraints and design variables. Bayesian optimization (BO) is a class of adaptive surrogate-based methods for global optimization. It can efficiently utilize a small budget of high-fidelity evaluations and handle a large number of constraints. However, it suffers from a hampered convergence rate for problems with a large number of design variables. Adaptive design space reductions via principal component analysis (PCA) and trust region (TR) approaches have been shown to improve the scalability in different phases of the optimization. Extending existing methods, we implement the ability to benefit from parallel high-fidelity evaluations and to handle constraints. We then sequentially combine our PCA-BO and TR-BO in a hybrid method to profit from their respective complexity reduction strategies. We assess the performance of our hybrid algorithm by comparing it to vanilla BO on two problems: a 40D analytical test function and a 55D aerodynamic compressor blade design. The empirical results show that PCA-BO enhances the convergence rate in the initial optimization phase, while TR-BO allows for further improvements in the later iterations. Moreover, the algorithm computation time is more than halved. For the compressor blade case, our approach yields an equally good design as vanilla BO after only 20 instead of 96 iterations. The proposed approach has the potential to extend the good performance of BO to even higher-dimensional constrained problems, including multi-stage turbomachinery optimizations.
AB - As turbomachinery designs become more complex, shape optimization tasks involve computationally expensive simulations and many constraints and design variables. Bayesian optimization (BO) is a class of adaptive surrogate-based methods for global optimization. It can efficiently utilize a small budget of high-fidelity evaluations and handle a large number of constraints. However, it suffers from a hampered convergence rate for problems with a large number of design variables. Adaptive design space reductions via principal component analysis (PCA) and trust region (TR) approaches have been shown to improve the scalability in different phases of the optimization. Extending existing methods, we implement the ability to benefit from parallel high-fidelity evaluations and to handle constraints. We then sequentially combine our PCA-BO and TR-BO in a hybrid method to profit from their respective complexity reduction strategies. We assess the performance of our hybrid algorithm by comparing it to vanilla BO on two problems: a 40D analytical test function and a 55D aerodynamic compressor blade design. The empirical results show that PCA-BO enhances the convergence rate in the initial optimization phase, while TR-BO allows for further improvements in the later iterations. Moreover, the algorithm computation time is more than halved. For the compressor blade case, our approach yields an equally good design as vanilla BO after only 20 instead of 96 iterations. The proposed approach has the potential to extend the good performance of BO to even higher-dimensional constrained problems, including multi-stage turbomachinery optimizations.
KW - Bayesian optimization
KW - Design space reduction
KW - Dimensionality reduction
KW - Efficient global optimization
KW - High-pressure compressor
KW - Principal component analysis
KW - Shape optimization
KW - Trust region
UR - http://www.scopus.com/inward/record.url?scp=85204296825&partnerID=8YFLogxK
U2 - 10.1115/GT2024-121848
DO - 10.1115/GT2024-121848
M3 - Conference contribution
AN - SCOPUS:85204296825
T3 - Proceedings of the ASME Turbo Expo
BT - Turbomachinery - Multidisciplinary Design Approaches, Optimization, and Uncertainty Quantification; Radial Turbomachinery Aerodynamics; Unsteady Flows in Turbomachinery
PB - American Society of Mechanical Engineers (ASME)
T2 - 69th ASME Turbo Expo 2024: Turbomachinery Technical Conference and Exposition, GT 2024
Y2 - 24 June 2024 through 28 June 2024
ER -