Deep learning based multi-cavity blade tip seal optimization

Tip leakage flow significantly affects both tip loss and the aerodynamic efficiency of turbines. This study presents a novel method for generating high-pressure turbine tip structures using Voronoi diagrams to mitigate tip leakage flow. This geometric strategy, coupled with an efficient optimization framework and a U-Net-based neural network trained on computational fluid dynamics (CFD) data, serves as a rapid surrogate model for predicting tip leakage flow. The surrogate model facilitates efficient optimization of the complex Voronoi geometry using a physics-based genetic algorithm. Comparisons with CFD results indicate that the neural network model exhibits higher prediction accuracy for blade tip static pressure, velocity, and leakage velocity but lower accuracy for cascade passage vorticity, suction surface static pressure, and shear stress. Aerodynamic analysis shows that the optimized tip structure produces targeted cavity formations at different locations along the blade tip. These tailored cavities intensify internal vortex structures and effectively obstruct the leakage flow transport from the pressure side to the suction side. The Voronoi-based blade tip cavity design reduces leakage mass flow by 3.1 % relative to conventional honeycomb blade tips.