Application of deep-learning method in the conjugate heat transfer optimization of full-coverage film cooling on turbine vanes

Cooling design optimization with complex nonuniform heat loads is a typical challenge in the develop-ment of new generation gas turbines. Combined inlet hot streaks and swirls caused by lean-burn com-bustor force film cooling to attenuate uneven heat loads on the blade surface, especially in the spanwise direction. Besides, the complex coupling relationship among hundreds of design variables of film cooling arrangement hinders the development of the optimal design. The deep learning model shows a strong fitting ability when dealing with high-dimensional nonlinear problems, which could fit the mapping re-lationship between design variables and the temperature field. In this paper, a turbine cooling design optimization methodology based on conjugate heat transfer (CHT) simulation and conditional generative adversarial network (cGAN) is developed, and the film cooling design of the 1st stage turbine vanes is optimized through the multi-objective genetic algorithm (MOGA). An initial sample containing 96 cases is constructed by CHT computational fluid dynamics (CFD) simulation with inlet hot streaks and swirls. Based on the initial sample, a cGAN model that predicts the temperature distribution of the vane surface is trained and tested. The film hole arrangement of the vane surface is described by a 276-bit binary optimization variable. The 5% maximum temperature predicted by cGAN and the regressed coolant mass flow is used as the dual optimization objectives. The optimal film hole arrangements in rows and the scattered film hole arrangements found by MOGA are compared, which shows that the optimal scattered arrangements perform better due to well adaptability to the nonuniform thermal load in the spanwise direction. The impact of sample size and sample selection on the performance of the cGAN model is dis-cussed. The retrained cGAN model indicates that a proper abundance of specific samples can improve the prediction of complex coupling phenomena such as backflow.(c) 2022 Elsevier Ltd. All rights reserved.