High-fidelity gradient-free optimization of low-pressure turbine cascades

Abstract

In this paper we demonstrate the ability to perform shape optimization of Low Pressure Turbine (LPT) cascades using a combination of Large Eddy Simulation (LES) and the Mesh Adaptive Direct Search (MADS) optimization algorithm. To our knowledge, this is the first instance of aerodynamic shape design of LPT cascade blades using LES. Current industry-standard optimization of LPT cascades is performed using the Reynolds Averaged Navier–Stokes (RANS) approach. However, this is limited by inherent inaccuracies of existing turbulence models, particularly in the presence of transitional and separated turbulent flows. To alleviate this, our approach utilizes LES in lieu of RANS, which has been shown to provide more accurate results in flow regimes where RANS often fails. We first validate our LES simulations for a baseline T106D LPT cascade against available experimental data, showing good agreement. We then perform shape optimization using two different objective functions. The first optimization cycle, specified to minimize total pressure loss coefficient while maintaining tangential force, shows an improvement of 16% relative to the baseline T106D. The second optimization cycle, specified to maximize tangential force while maintaining total pressure loss coefficient, shows an improvement of 29% relative to the baseline T106D. Based on these results we conclude that optimization of LPT cascades with LES is feasible, and can yield significant improvements in performance with reasonable computational cost. Hence, this work supports a transition to higher-fidelity LES simulations as the foundation for optimization of next-generation LPT cascades.