A Machine Learning Approach for the Prediction of Time-Averaged Unsteady Flows in Turbomachinery

Abstract Recent advances in deep learning have led to its increased application in the field of fluid dynamics. By using a data-driven approach instead of a more conventional numerical approach, it is possible to reduce the computational cost of fluid simulations significantly. Especially unsteady computational fluid dynamics (CFD) are known to require a considerable amount of time and resources. Hence, it is common practice in turbomachinery to model the flow as steady by averaging the flow between the rotor and stator in a so-called mixing plane. While this approach is numerically efficient, the full interactions between the rotor and the stator can no longer be predicted accurately due to the averaged upstream flow field. Contrary to this, the time-average of an actual unsteady flow does not contain such a modeling error while still resulting in a flow field that is decoupled from its temporal fluctuations. In this work, we introduce a graph neural network (GNN), which predicts the time-averaged flow field of a rotor stage in an axial compressor based on its steady solution. Because GNNs are able to operate directly on the high-fidelity CFD mesh, we are able to retain the spatial resolution necessary to depict more complex flow behaviour. Consequently, the fidelity of our predictions can compete with conventional high-accuracy flow simulations. Our model is capable of predicting the velocity, pressure, density, and temperature field of a single rotor stage from a 4½-stage axial compressor test case and shows significant improvements compared to the steady state solution, while also being substantially faster than conventional unsteady CFD simulations.