Predicting Time-Averaged Unsteady Flows in Turbomachinery via Graph Neural Networks

Abstract Recent advancements in deep learning have led to its increased application in the field of fluid dynamics, offering a promising alternative to conventional numerical approaches. By using a data-driven methodology, deep learning can significantly reduce the computational cost of flow simulations. Despite continuously increasing computational resources, obtaining unsteady results using computational fluid dynamics (CFD) remains a time-consuming and expensive task for complex flows. In turbomachinery design, the flow is hence often modeled as steady by circumferentially averaging the flow between blade rows in a so-called mixing plane. While numerically efficient, the full interactions between the rotor and stator rows are no longer accurately predicted. Time-averaging an unsteady simulation, in contrast, preserves the rotor–stator interactions while still depicting the mean flow behavior. In this work, we present a graph neural network (GNN), which predicts the time-averaged unsteady Reynolds-averaged Navier–Stokes (URANS) flow field in a 41/2-stage axial compressor based on its steady-state Reynolds-averaged Navier–Stokes (RANS) solution. Given that GNNs are able to operate directly on common numerical meshes, the model retains the spatial resolution necessary to accurately depict the complex flow behavior. The fidelity of the predicted flow field is therefore comparable with conventional flow simulations. Our model is able to predict the velocity, pressure, density, and temperature field as well as the turbulent kinetic energy and logarithm of the turbulence dissipation rate across all nine compressor rows. The resulting flow fields are, on average, about 79% more accurate than the initial RANS solution at virtually no additional computational cost. The dataset utilized for training this machine learning model is publicly available for further research and validation.