Cross-scale feature fusion with gradient-enhanced attention for accurate prediction of film cooling

Abstract

Accurate prediction of film cooling effectiveness is critical for optimizing gas turbine blade durability under extreme thermal conditions. This study proposes a novel deep learning framework integrating a gradient-enhanced attention mechanism with U-Net architecture to establish an end-to-end mapping from four key parameters (injection angle $\alpha$, lateral expansion angle $\beta$, forward expansion angle $\gamma$, and blowing ratio $M$) to two-dimensional cooling effectiveness distributions. The UCGA model employs cross-scale feature fusion through adaptive pooling and a Sobel operator-based gradient attention module to enhance edge perception in flow field reconstruction. The model is based on 125 Computational Fluid Dynamics numerical simulations covering different geometrical parameter configurations. The validation demonstrates exceptional prediction accuracy (Coefficient of Determination (R²) > 0.99, Structural Similarity Index Measure (SSIM) > 0.97). SHapley Additive exPlanations (SHAP) analysis shows that $\beta$ and $M$ are the dominant parameters, with $\alpha$ and $\gamma$ having a relatively small average effect on model predictions. This is consistent with the physical mechanisms controlling coolant coverage and momentum balance. This work provides valuable insights into the film cooling effectiveness distribution. Most importantly, the developed UCGA model is a very promising tool for fast, high-fidelity iterative optimization of the geometric parameters of the fan-shaped holes, with the aim of providing a reference for accelerating the design cycle.