Multi-fidelity network framework for field and performance parameters prediction of turbomachinery based on deep graph learning

To tackle the intrinsic conflict between prediction accuracy and data cost in surrogate models within the turbomachinery field, this work introduces a multi-fidelity (MF) framework based on deep graph learning, which enables the comprehensive utilization of data with different fidelities. This framework enables simultaneous prediction of fields and performance parameters. Taking SCO₂ turbines as the research object, this paper has verified the superiority of the model. The results demonstrate that the MF model constructed based on the SAGE operator achieves the highest prediction accuracy. The maximum errors of the pressure, temperature, and velocity field predicted by the MF model are 0.15MPa, 3K, and 2m/s respectively, corresponding to relative errors of 1.43 %, 0.41 %, and 1.03 %. And the error of each physical field is lower than that of the single-fidelity model. Most of the predicted results for performance parameters fall within the 0.5 % error band, and the MF model is more capable of achieving unbiased estimation of parameters compared with the single-fidelity model. The MF network architecture effectively addresses the challenge of high-precision prediction in small-sample scenarios. It provides technical support for optimization in industrial scenarios, helps accelerate the intelligent upgrading of the turbomachinery manufacturing industry, and can be extended to other rotating machinery fields to broaden its engineering application scope.