Aerodynamic Optimization of Turbine Airfoils Using Multi-Fidelity Surrogate Models

Aerodynamic shape optimizations of turbine airfoils typically require numerous time-consuming numerical flow simulations. By using accurate high-fidelity methods, these optimization processes can extend over days or weeks. However, alongside these high-fidelity methods, there exist faster, albeit less accurate, low-fidelity methods. Combining simulation results from both fidelities in multi-fidelity optimizations can minimize effort while retaining high-fidelity accuracy.
This research examines a combination of 3D-CFD RANS (high-fidelity) and quasi-3D-CFD Euler (low-fidelity) simulations. First, it assesses the discrepancies between these two methods for typical industrial gas turbine vanes and blades, which exhibit different flow regimes. It is found that low-fidelity simulations consistently yield similar error values for relevant response variables across all examined vanes and blades.
In addition, this work examines and develops several single-fidelity and multi-fidelity surrogate models that integrate results from different fidelity levels, with a focus on models based on proper orthogonal decomposition (POD). These models, on average, provide better prediction accuracy than the single-fidelity Kriging models that serve as a reference.
The highest-performing modeling method is then applied to the aerodynamic shape optimization of gas turbine vanes and blades on an industrial scale, achieving up to 30% performance advantages. The proposed optimization strategy offers cost-efficient design space exploration and rapid identification of feasible optimization members. Comparing initial and final airfoil designs reveals the reduction of typical losses, including shock losses, flow separation, boundary layer growth, and secondary flows.