Data-driven RANS turbulence modeling for compressor stall

Abstract

Compressor stall in an aero-engine can result in partial or total loss of power for the engine, leading to engine failure or shutdown. Hence, it is a critical parameter during the design stages of an aero-engine. The conventional method to predict the compressor stall margin at the design phase is to perform a Reynolds-Averaged Navier-Stokes (RANS) simulation with a turbulence model. However, this method often underpredicts the stall margin due to the deficiency of the turbulence model in predicting 3D separated flows. Recent research has demonstrated the capabilities of a data-driven turbulence modeling approach in 2D canonical flows. In this thesis, a data-driven turbulence model for better prediction of compressor stall margin is pursued.

The ingredients for a data-driven turbulence model include (1) data, (2) flow physics, and (3) a data-driven framework. To address each of these issues for turbomachinery applications, this thesis is organized accordingly, with each one of the three major parts of the thesis corresponding to one issue.

In the first part of the thesis, an upgraded delayed detached eddy simulation (DDES) method is devised for compressor tip leakage flows, which provides a cost-effective means to generate a turbomachinery turbulence database. The method is validated and verified in a backward-facing step and a low-speed axial compressor rotor. Afterward, the method is applied to generate a turbulence database of compressor tip leakage flows. By analyzing the mean flows, the turbulence statistics, and the modal behavior, the understanding of turbulence modeling for compressor tip leakage flows is advanced. The proposed upgraded DDES method is promising for generating a larger turbomachinery turbulence database.

In the second part of the thesis, comprehensive uncertainty quantification and evaluation studies have been performed on the model coefficients and the model forms of the Spalart-Allmaras turbulence model. Results indicate that a data-driven correction at the transport equation level is most promising for compressor stall prediction. However, established empirical corrections fail to capture the flow features of 2D transonic flows and 3D separated flows. To tackle these limitations, two novel flow features namely baroclinicity and vortical pressure gradient are proposed, respectively. Modified turbulence model forms (i.e., SA-B and SA-PG$_\omega$) are proposed with respect to each of these new flow features, and the newly emerged model coefficients are calibrated with relevant flow data. The calibrated models demonstrate great improvement in prediction accuracy in similar but unseen flow cases. In particular, the SA-PG$_\omega$ model demonstrates adequate accuracy in predicting the stall boundary of a range of compressors. This modified model is the most useful contribution of the thesis.

In the last part of the thesis, a data-driven turbulence modeling framework is established and demonstrated in 2D transonic flows. Methods that improve the explainability of the trained machine learning model are applied, and the causal link between the input flow features and the output eddy viscosity difference of the trained machine learning model is explained. It is found that the trained machine learning model re-discovered (1) the scaling between the eddy viscosity and its source term and (2) the effect of shear and rotation on the eddy viscosity source term, which are well-known for human turbulence model developers. The presented method is transferable to a general data-driven turbulence model, and it can help industrial users build trust in the data-driven turbulence model.

The methods and knowledge obtained from the thesis provide engineering solutions for compressor stall prediction.