Multi-fidelity graph neural network for flow field data fusion of turbomachinery

Abstract

Efficient and accurate prediction of the flow field in turbomachinery is vital for tasks such as optimization and off-design modeling. Deep learning methods offer inspiring tools for flow field prediction when there is sufficient high-fidelity data for training. However, high-fidelity flow fields may be insufficient in practice due to the high computational/experimental cost. In this work, the capabilities of deep learning methods for fusing multi-fidelity flow field data are further explored. A multi-fidelity graph neural network (MFGNN) is proposed. The proposed framework contains two networks for approximating the low-fidelity flow fields and the correlations between the low-fidelity and high-fidelity flow fields, respectively. The data fusion method is validated by the off-design flow field prediction of a turbine. With limited high-fidelity data, MFGNN can accurately predict flow fields and is superior to the graph neural network that only uses high-fidelity data. The effects of low-fidelity dataset size and the extrapolation performance are also explored. With appropriate prior guidance by low-fidelity data, MFGNN can predict unknown flow fields within and beyond the range of high-fidelity training datasets. The proposed deep learning method shows the advantages of high precision and generalizability in addressing the physical field prediction problem.