Multi-Fidelity Aerodynamic Optimization of a Low-Reynolds-Number Outlet-Guide-Vane Cascade

Paper in proceeding, 2026

This study presents an efficient aerodynamic optimization framework that integrates multi-fidelity surrogate modeling with advanced dimensionality reduction techniques. The study focuses on the shape optimization of a low-Reynolds-number Outlet-Guide-Vane (OGV) cascade, with the objective of minimizing pressure losses at the aerodynamic design point and two off-design conditions, within the context of the ongoing Horizon Europe Sci-Fi-Turbo project. Parameterization of the OGV cascade is enabled using the BladeGen tool developed at the DLR Institute of Propulsion Technology. To construct multi-fidelity Kriging and Neural Network-based surrogate models, Reynolds Averaged Navier-Stokes (RANS) Simulations and the MIT Inverse Solver for Euler Simulations (MISES) are respectively employed as high- and low-fidelity datasets. Optimization results demonstrate that both multi-fidelity Kriging and Neural Network-based models yield Pareto fronts closely matching those generated by high-fidelity-only optimization, despite using just 22 high-fidelity simulations. This highlights the potential of multi-fidelity surrogate modeling to significantly reduce computational cost while maintaining accuracy. The results also show that the multi-fidelity Kriging model slightly outperforms the multi-fidelity Neural Network model, likely due to the limited number of high-fidelity training samples (22 in this study), which necessitated the use of a small neural network architecture, specifically, a single hidden layer with four neurons. To further improve surrogate performance in high-dimensional design spaces, Proper Orthogonal Decomposition (POD) and Autoencoders are applied for dimensionality reduction of the cascade design parameter space. Results show that the Autoencoder achieves better dimensionality reduction than POD, with the reduced-space Pareto front more closely aligning with the full-space counterpart. The dimensionality of the parameter space is successfully reduced from 15 to 5, facilitating more efficient surrogate modeling. Overall, this study demonstrates a robust and computationally efficient framework for the multi-fidelity aerodynamic optimization of turbomachinery components, supporting the development of next-generation, low-emission aero engines.