Data-driven analysis of high fidelity simulation of multi-stage compressor

Abstract

The flow dynamics within the multi-stage compressors are complex and not well understood. To study them we employed high-fidelity computational fluid dynamics simulation. In our simulation we mimic the multi-stage environment using an idealised repeating-stage compressor model. This allowed us to generate large amounts of time-averaged and unsteady data. However, the analysis of the data poses a significant challenge due to the highly unsteady nature of the flow which results from the presence of periodic wakes from the upstream blade-rows and freestream turbulence.Standard approached based on phase averaging provide only a partial statistical representation of the flow complexity and confound dynamics that are not strictly periodic or happen rarely. In addition, appropriate estimation of phase-averaged quantities in such unsteady flowfield requires many samples and results in large computational cost. To overcome this problem and gain further insight into flow physics a data-driven approach based on Proper Orthogonal Decomposition (POD) is developed. The method is parallised and compatible with high performance computing (HPC) systems to allow for processing of massive amounts of flowfield data. The focus of this paper is therefore two-fold: firstly, the implementation of an algorithm for performing a parallel POD on HPC is discussed; secondly, the results of the modal decomposition are discussed and compared with phase-averaged results for a repeating-stage NACA65 compressor configuration. We show that the number of snapshots required to provide a converged modal representation of the flowfield dynamics is lower than the number of snapshots commonly used for phase-averaging. Further, we compare the wake periodic flowfield obtained from the POD procedure with the phase-averaged flowfield and demonstrate the ability of POD procedure to isolate dynamics such as Kevin-Helmholtz vortices or lambda-breakdown events. The results highlight the capability of data-driven approaches in efficiently identifying the main dynamics present in complex and unsteady environments such as those found in turbomachinery flows.