An investigation on implementation of generating adversarial network-based surrogate models for prediction of turbine endwall film cooling effectiveness

Abstract

Film cooling is an essential cooling technique for turbine blade, which guides coolant to cover the turbine blades surface by arranging discrete film holes on it. For the newly designed cooling units, the process of predicting film cooling effectiveness is really complex and time-consuming. For the purpose of accelerating the design procedure, the U-Net which is a very popular algorithm in the field of deep learning was improved with the basic layer structure, then it was adopted as the proxy function for the film cooling effectiveness prediction. The basic samples were generated by Latin hypercube sampling method and calculated by commercial software. The results were verified by experiment in a linear cascade. The training and testing samples were calculated at experimental conditions (inlet Mach number values 0.12). The results showed that comparing with the basic U-Net network, the novel U-Net network has a great better accuracy on the prediction of contour results. Both of the improved U-Net networks had higher prediction accuracy than original models. The correlation index reached more than 99.5% and the first order error was lower than 0.3%. The number of basic neural layers is not positively correlated with the network prediction accuracy which exists an optimal ratio. The experimental verification shows that the generating adversarial network prediction results are consistent with the experimental results. The final prediction results showed that the U-Net model can be used as a surrogate model and meets the requirements for accuracy of film cooling effectiveness prediction.