Efficient large-scale graph learning for predicting the 3D multi-physics flow fields of axial compressor Rotor37 with variable geometry

Accurate prediction of the multi-scale, multi-physics flow characteristics in axial compressors is essential for optimizing the aerodynamic performance . Traditional experimental and numerical simulation methods are costly and difficult to achieve agile design of modern compressors. To achieve fast yet accurate predictions of the complicated flow characteristics of axial compressor, a data-driven deep forecasting model, named GraphRotorNet , has been built via graph neural networks (GNNs) for NASA’s Rotor37 axial compressor. The model employs sub-graph clustering to efficiently process large-scale multi-physics snapshots from Rotor37’s 3D single-passage fluid domain. Simultaneously, a deep message-passing mechanism is developed to enhance the ability to capture multi-scale spatial features of compressible flow fields under varying geometric parameters . Besides, extensive numerical experiments show that the GraphRotorNet outperforms state-of-the-art competitors in accurately predicting Rotor37’s 3D multi-physics fields and performance characteristics. Finally, the model also imposes a lower GPUs burden when processing large-scale graphs for real-world engineering applications . • Efficient graph model of 3D flow fields of Rotor37 compressor by subgraph clustering. • Deep message-passing graph model for capturing the multi-scale flow characteristics . • Quantified uncertainty with variable geometrical parameters is provided. • Lower GPUs load of large-scale graphs for real-world applications.