U-Net convolutional architecture for squealer tip film cooling prediction and multi-objective optimization

The optimization of turbine blade squealer tip cooling systems is limited by the high computational cost of Computational Fluid Dynamics simulations, which restricts the exploration of large design spaces and the identification of efficient cooling layouts. This work proposes a machine-learning-based framework for the cooling system of the 𝐺 ⁡ 𝐸 − 𝐸 3 high-pressure turbine blade tip, with fixed squealer cavity geometry and variable number, diameter, and position of film cooling holes. A dataset of 2000 Computational Fluid Dynamics simulations is used to train two Artificial Neural Networks for predicting area-averaged adiabatic effectiveness and coolant mass flow rate, and a U-Net for reconstructing the full tip temperature field. The trained surrogate models are coupled with a Particle Swarm Optimization algorithm using four objective functions representing different cooling design priorities. The framework rapidly identifies optimized configurations that vary according to the selected objective. The best solution simultaneously improves average tip adiabatic effectiveness, coolant mass flow rate, and temperature-field deviation from the coolant temperature relative to the baseline design. The U-Net reconstructs the full temperature field with maximum prediction errors below approximately 5.92 kelvin, while reducing design-evaluation time from several hours to a few seconds. For the investigated geometry and operating conditions, the optimized layouts indicate that coolant is more effectively distributed through multiple small-diameter holes, whereas holes located in high-pressure tip regions may provide limited benefit because of the local surface pressure. These results show that data-driven surrogate models combined with evolutionary optimization can enable rapid and effective exploration of turbine blade cooling designs.