Physics Guided Graph Neural Networks (PG-GNNs): Numerical Investigation of High-Pressure Turbine Flow

Abstract This work presents the first attempt to use Graph Neural Networks (GNNs) for high-pressure turbine vane flow, while enforcing physics conservation, which we called Physics Guided Graph Neural Networks (PG-GNNs). Compared to conventional neural networks, GNNs are more effective in handling irregular graphical data, such as in turbomachinery, where flow fields are often presented as data points on a mesh (graphical data). While GNNs are shown effective in many engineering areas, in turbomachinery there is a need for enhanced accuracy, and this requires additional physical constraints. For these reasons PG-GNNs are proposed, by enforcing Euler governing equations into the training process. The performance of PG-GNNs is then evaluated by solving both forward and inverse problems. In the forward problem, the PG-GNNs can accurately capture the turbine flow field based on some labelled data randomly distributed inside the turbine passage. As for the inverse problem where the boundary conditions are incomplete, PG-GNNs show obvious advantages over the traditional computational fluid dynamics (CFD) approaches in reconstructing the whole flow field by only utilizing partial flow information from either the velocity components or the near-wall pressure. The proposed approach provides a promising solution for prediction and analysis in advanced turbomachinery design and optimization.