Integrated graph deep learning framework for flow field reconstruction and performance prediction of turbomachinery

The performance and reliability of turbomachinery directly affect the efficiency and safety of energy conversion systems. A dual graph neural network (DGNN) for turbomachinery flow field reconstruction and performance prediction is proposed, which utilizes flow field data at grid vertices and characterizes the neighborhood relationships of grids through the adjacency matrix. Different from previous work, this work extends deep learning methods to the reconstruction of global turbomachinery fields defined on arbitrary structured/unstructured grids. The flow field reconstruction and performance prediction of low aspect ratio rotors are used to verify the generalizability of DGNN. The proposed method can not only accurately predict performance parameters, but also achieve excellent performance in flow field reconstruction. The superiority of DGNN over the artificial neural network (ANN) in flow field reconstruction is clarified. The optimal GNN operator with the highest accuracy and lowest computation costs is obtained as SAGE. A well-trained DGNN can give the flow field distribution as well as the performance of turbines within 0.05 s. The proposed approach could be a real-time simulation and analysis approach to assist turbomachinery design and optimization. It may provide an efficient solution for establishing the digital twin system of turbomachinery in the future.