AI-Based Optimization of Internal Cooling Passages of a High–Pressure Vane

Gas Turbine (GT) performance greatly depends on Turbine Inlet Temperature (TIT), whose upper limit is determined by the necessity to keep the metal temperature of High–Pressure Vanes (HPV) below safety thresholds. Internal cooling systems are used to increase the TIT without incurring in turbine failure by thermal fatigue. Machine Learning (ML) has proved to be a fast and useful tool for optimizing turbine aerothermal performance, but most approaches are limited to scalar, averaged quantities and fail to capture the detailed thermal distribution on the vane surface. The goal of the current study is to overcome the limitations of ML methods while optimizing the internal cooling configuration of the well–established NASA C3X test case. Two different kinds of ML will be contextually employed. The Encoder–Decoder Convolutional Neural Network (EnDeNN) has been selected for its ability to produce the surface temperature distribution from an image of a spanwise section of the vane, detecting not only the average temperature, but also thermal hot spots. Moreover, the mass–flow rate of the internal channels is computed through a traditional Artificial Neural Network (ANN). Both the EnDeNN and the ANN methods are used as surrogate models for an optimization algorithm aimed at determining the best possible cooling channels configuration for the selected working conditions. The Design of Experiments (DoE) is populated by a number of Conjugate Heat Transfer (CHT) simulations performed using ANSYS FLUENT. Turbulence is modeled by using transitional SST γ – Reθ method, appropriately tuned to match the experimental temperature distribution. The outcomes of this work demonstrate that the optimization of cooled turbomachinery components can be significantly accelerated. The proposed approach provides numerical results consistent with CHT simulations, and identifies an optimized configuration characterized by improved thermal performance with respect to the baseline one.