A deep learning framework for turbine film cooling prediction: Parameter-based geometric representation and multi-fidelity transfer learning

Accurate and efficient prediction of turbine film cooling performance is essential for advanced gas turbine design but remains computationally expensive due to the complex interactions between the mainstream flow and multiple coolant jets. In this study, a deep-learning-based framework is developed for the rapid prediction of dimensionless wall-temperature fields in turbine film cooling. A parameter-list-based mapping strategy is proposed, in which each film cooling hole is treated as a fundamental unit. A multi-branch, multi-scale residual self-attention network is constructed to capture the multi-source and high-gradient characteristics of film cooling flow and heat transfer. The proposed approach is systematically compared with conventional image-based mapping methods using consistent model architectures. The results demonstrate that the parameter-based representation more effectively captures geometric features under conditions involving strong surface curvature and multiscale hole configurations, thereby achieving higher prediction accuracy and improved generalization. Furthermore, the influence of prior information is investigated. Incorporating the uncooled wall-temperature field improves the performance of the image-based models but has limited or adverse effects on the parameter-based models because of dimensional inconsistency. To address the trade-off between data fidelity and computational cost, a multi-fidelity modeling framework based on transfer learning is developed. By transferring knowledge from low-fidelity fluid-only simulations to high-fidelity conjugate heat transfer cases, the resulting model achieves a test-set relative error of 0.96%, a PSNR of 38.71 dB, and an SSIM of 0.988, while requiring only one-quarter of the training data and one-tenth of the training iterations. Overall, this study demonstrates that appropriate geometric representation and multi-fidelity learning can significantly improve prediction efficiency and robustness, providing a data-efficient framework for high-fidelity turbine film cooling analysis.