Session: 36-03 MDO & Multi-Fidelity

Submission Number: 178703

Optimization of an OGV Cascade Using Multi-Fidelity Surrogate Modeling and Active Learning With Scale-Resolving Simulations 

The ambitious goals set by the European Union to achieve climate neutrality by 2050 require a drastic reduction in greenhouse gas emissions across all sectors. Within the transportation sector, civil aviation represents a particular challenge, as it is identified by the European Environment Agency as the fastest-growing source of carbon dioxide emissions in Europe. Consequently, achieving a carbon-neutral European Union will necessitate substantial reductions in the emissions generated by the aviation industry. Enhancing propulsion efficiency directly reduces fuel consumption, therefore lowering the environmental impact of air travel. Typically, this can be pursued through aerodynamic shape optimization of engine components, such as compressor and turbine blades, where the objective is to identify geometries that maximize aerodynamic performance while minimizing total pressure losses.

Scale-resolving simulations (SRS), particularly Large-Eddy Simulations (LES), have the potential to play a central role in aerodynamic shape optimization by providing accurate predictions of the flow scales of interest for turbomachinery flows. However, their use in computational optimization frameworks for engineering applications is significantly limited by their computational cost, which significantly restricts the exploration of the design space. To overcome this limitation, it is therefore essential to combine these high-fidelity simulations with lower-fidelity models within a multi-fidelity framework. This approach balances computational cost and predictive accuracy, enabling the optimization process to leverage the accuracy of SRS while maintaining computational tractability.

In this work, we present an automated Bayesian surrogate-based multi-objective optimization framework and we apply it to the optimization of an outlet guide vane (OGV) cascade under geometrical and outflow constraints at an inlet Reynolds number of 150,000. The optimization aims to minimize the pressure loss at the design inlet flow angle and at two off-design angles, namely ±5° with respect to the design condition. The blade geometry is parameterized using a manufacture-ready representation (Christian Voß et al., 2025) with more than ten design parameters, which are subsequently reduced through Proper Orthogonal Decomposition. The surrogate model employed in the optimization is a multi-fidelity Gaussian process, namely co-kriging, which combines a large number of cheap low-accuracy Reynolds-Averaged Navier–Stokes (RANS) simulations with a limited number of expensive high-resolution LESs to make prediction with high-fidelity accuracy at points where high-fidelity data is not available. RANS simulations are performed with the hybrid finite volume/finite element Navier-Stokes solver WOLF (Alauzet et al., 2024), whereas LESs are performed with the high-order finite difference CFD solver MUSICAA (A. Bienner et al., 2024). The surrogate model is initially trained with 50 low-fidelity and 5 high-fidelity evaluations. An infill procedure leveraging three different infill criteria is then employed to add 150 low-fidelity and 15 high-fidelity samples and find the optimal blade design. The infill criteria are selected to balance exploration and exploitation of the design space. To improve the robustness of the design space exploration and ensure the feasibility of the optimal geometry, both geometric and flow constraints are incorporated into the infill procedure. Geometric constraints are enforced by restricting the evaluation of the acquisition functions to feasible regions of the design space. Flow constraints, on the other hand, are framed in a probabilistic manner. A co-Kriging model is fitted to the outflow angle and used to promote the acquisition of design points in regions of the design space that maximize the probability of satisfying the outflow angle constraints. Finally, to assess capabilities of the proposed framework, we compare the results of the multi-fidelity optimization with a reference fully LES-based and a conventional RANS-based optimization.

Presenting Author: Mattia Fabrizio Ciarlatani Sorbonne University

Presenting Author Biography: Mattia Ciarlatani received his B.Sc. and M.Sc. degrees in Aeronautical Engineering from Politecnico di Milano, and his Ph.D. in Civil and Environmental Engineering from Stanford University, where his research focused on LES-based multi-fidelity modeling and wind loading prediction for the design of high-rise buildings. He is currently a postdoctoral researcher at the Institut Jean le Rond d’Alembert, Sorbonne University.

Authors:

Mattia Fabrizio Ciarlatani Sorbonne University
Marc Schouler Sorbonne University
Anca Belme Sorbonne University
Xavier Gloerfelt École nationale supérieure d'Arts et Métiers
Paola Cinnella Sorbonne University