Physics-guided generative prediction of film cooling effectiveness via conditional wasserstein GANs with attention

Accurate and rapid prediction of film-cooling effectiveness on high-pressure turbine blades is essential for advanced thermal management and engine efficiency. However, traditional computational fluid dynamics (CFD) is computationally expensive for large-scale design exploration, while existing data-driven surrogates often suffer from unstable training or sacrifice local accuracy. To address these challenges, we propose a Physics-Guided Conditional Wasserstein Generative Adversarial Network with Gradient Penalty and Convolutional Block Attention Modules (CWGAN‑GP‑CBAM). This framework stabilizes adversarial training using the Wasserstein-1 distance with gradient penalty to mitigate mode collapse. It enhances feature discrimination through a CBAM-augmented U-Net generator, explicitly focusing on critical flow structures like shear layers and jet cores. Additionally, it encodes deterministic and stochastic information via a composite six-channel input combining 2D geometry maps, broadcast boundary-condition fields, and latent noise. A physically consistent dataset of 1500 sampled geometries and blowing ratios is generated using a Hybrid Free-Form Deformation (HFFD) method, reducing data-preparation time by 19.8%. Evaluated across multiple random seeds on an independent test set, the proposed model achieves a mean absolute error of 1.72×10⁻², a structural similarity index (SSIM) of 0.969, and a peak signal-to-noise ratio (PSNR) of 37.0 dB. Region-level evaluations show mean cooling effectiveness errors under 3% and nonuniformity errors below 8%. Compared to the standard CGAN baseline, our approach reduces mean squared error by 36.7%, demonstrating its potential as an effective tool for turbine cooling layout optimization and broad design space exploration.