A physics-guided global-local deep learning method for rapid reconstruction of film-cooled turbine blade temperature fields

Film cooling is a critical technology for protecting gas turbine blades from extreme thermal loads. However, the design of complex cooling configurations is severely constrained by the prohibitive computational cost of high-fidelity computational fluid dynamics simulations. Existing data-driven methods generally suffer from geometric rigidity, intrinsically coupling local fluid dynamics with global blade geometry, thereby requiring massive fully-cooled 3D datasets and costly retraining. To overcome these fundamental bottlenecks, this study proposes a physics-guided global-local deep learning framework for the rapid reconstruction of full-blade film cooling temperature fields. The fundamental novelty lies in effectively decoupling local flow physics from global geometric morphology. First, an adapted generative adversarial network is trained exclusively on computationally efficient canonical flat-plate data to capture high-fidelity local flow features. Subsequently, a deterministic global superposition reconstruction method is developed. It mathematically unwraps the complex 3D blade into a 2D plane, maps the local predictions using dimensionless coordinates, and thermodynamically couples them with the uncooled blade's baseline temperature. Validation against full-scale 3D simulations demonstrates that the local generative network achieves a high coefficient of determination of 0.981. For full-blade reconstruction across various multi-hole layouts, the absolute prediction error across the majority of the blade surface remains within 0.12. Remarkably, the proposed method accelerates the complete global thermal field evaluation by three orders of magnitude. This physics-driven decoupling framework successfully preserves sharp thermal gradients, providing a reliable and highly efficient surrogate tool for the rapid parametric design optimization of next-generation turbine blades.