Establishment of a generalizable model on a small-scale dataset to predict the surface pressure distribution of gas turbine blades

The main challenge of establishing a model to predict the flow fields of turbomachinery was insufficient data. This study aimed to establish a generalizable and accurate model on a small-scale dataset to cost-effectively predict the surface pressure distribution of a turbine rotor cascade with widely varying geometries and boundary conditions. To meet this purpose, a novel concept of transfer learning was introduced, which was defined as transferring knowledge from a large-scale but low-fidelity dataset to a small-scale but high-fidelity dataset. A Conditional Generative Adversarial Neural Network was designed as the pre-trained network for the transfer learning to regress the surface pressure distributions. Two models transferred from datasets with different fidelity and an independent model were established and compared in detail. The results showed that the proposed method successfully reduced the modeling cost with a low error in predicting the surface pressure distributions. The model transferred from the higher-fidelity dataset had better generalization performance, which reduced the root mean square error and modeling cost by 40.2% and 9 times, respectively. The presented method could serve as a base framework for modeling surface pressure distribution of complex objects using a small-scale dataset. (C) 2020 Elsevier Ltd. All rights reserved.

Automated summary

By introducing the concept of transfer learning, this study successfully established a model with generalizability and accuracy on a small-scale dataset to predict the surface pressure distribution of a turbine rotor cascade with varying geometries and boundary conditions cost-effectively. The model transferred from a higher-fidelity dataset showed better generalization performance, reducing the root mean square error and modeling cost significantly.