Investigating shared-layer architectures in single-scale graph convolutional networks for surrogate modelling of conical diffuser flows

Designing gas turbine combustors that operate effectively with turbomachinery components over a wide range of operating conditions is a challenging task, requiring detailed 3D computational fluid dynamics (CFD) analysis, which is computationally intensive. To alleviate this, surrogate models of the different sub-components like the pre-diffuser, combustion zone, dilution zone and transition ducting can be employed. These component-level models can then be integrated into 1D design codes together with surrogate models of turbomachinery components to enable comprehensive whole-engine multi-point optimization. This paper investigates how graph and scalar boundary-condition data can be jointly processed within a single-scale spectral graph convolutional network (GCN), through a systematic study of shared-layer configurations that exchange information between the GCN and fully connected layers. Conical air pre-diffusers are used as the application testbed, with the resulting surrogate model predicting the 2D symmetry-plane velocity, temperature, and pressure distributions. The model is trained on CFD simulation data encompassing diffusers with diverse geometric parameters, including a wide range of aspect ratios, diffuser lengths, and inlet diameters, as well as various combinations of boundary conditions. As input, the surrogate model receives graphs representing the symmetry plane meshes, along with vectorized boundary condition values for each training sample. Different layer architectures are investigated to share information between the GCN and fully connected layers that process the graph and boundary condition data respectively. The best-performing surrogate model achieves high predictive accuracy, with normalized mean absolute percentage errors below 1% for pressure and temperature, and below 3.9% for velocity.