Predicting distribution of adiabatic film cooling effectiveness using generative adversarial networks with mist/air mixture as coolant

Film cooling technology is critical for protecting the core components of aero-engines and gas turbines from high-temperature damage. However, high-fidelity computational fluid dynamics (CFD) simulations for mist-assisted film cooling are computationally expensive, especially during multi-parameter design and optimization processes. To address this issue, this paper proposes a deep-learning-based surrogate framework using a Conditional Generative Adversarial Network (CGAN) to rapidly predict the two-dimensional adiabatic film cooling effectiveness distribution on a flat plate with a mist/air mixture as the coolant. Unlike existing deep learning models that primarily focus on traditional single-phase air cooling or employ standard regression architectures that struggle to capture high-gradient flow details, the proposed CGAN model can more accurately characterize the complex, nonlinear multiphase flow features in mist/air cooling. The model takes the blowing ratio, mist concentration, and droplet diameter as input parameters and learns the nonlinear mapping from one-dimensional operating conditions to high-resolution effectiveness contours. Furthermore, by combining dataset expansion, transfer learning, and regularization, this paper proposes a generalization-enhancement strategy geared toward extrapolated operating conditions. The results show that the proposed CGAN model can achieve a computational speedup of several thousand times relative to CFD, while the maximum absolute errors of pixel values for both the training and test sets are maintained below 0.07. Additionally, it is found that the model's error distribution characteristics reflect the varying degrees of impact of different physical factors (turbulent diffusion, thermophoretic force, Brownian force, and Saffman lift) on the film cooling effectiveness. For test samples outside the initial training domain under extreme blowing ratio conditions, the proposed generalization-enhancement strategy can significantly reduce the mean absolute error by 75%, dropping from 0.048 in the baseline model to 0.012. The experimental results demonstrate that the proposed framework can provide an accurate, efficient, and highly robust prediction method for mist-assisted film cooling distribution modeling.