GPU-accelerated high-order scale-resolving simulations using the flux reconstruction approach

Abstract

High-order methods in Computational Fluid Dynamics (CFD) offer a potential route towards the resolution of hitherto intractable fluid-dynamics problems in industry. The Flux Reconstruction (FR) approach provides a unifying framework for a number of popular high-order methods such as the Discontinuous Galerkin (DG). Its suitability for use on unstructured grids along with its ability to facilitate massively parallelised implementation on architectures such as GPUs provide a means to tackle computationally challenging flows around complex geometries. Such a flow can be found in the rod-aerofoil tandem configuration: Complex, unsteady flow structures generated by and interacting with more than a single solid body are central to a number of applications in the aerospace industry. The current thesis attempts to demonstrate the suitability of the FR approach in successfully simulating the flow around a rod-aerofoil configuration. The in-house CFD solver employed in the research is presented and the FR implementation analysed. Computational grid resolution issues arising from the rod-aerofoil problem are studied and a novel strategy for the stabilisation of the computation is implemented in the form of local entropy stability. The results obtained are analysed and conclusions are drawn on the utility of the FR approach in the absence of a sub-grid scale model (Implicit LES - under-resolved DNS). The present work confirms the utility of local entropy stability for the stabilisation of the rod-aerofoil simulation of aerofoil-chord based Reynolds number of Re=480, 000. It will also demonstrate that the under-resolved DNS setup that resulted in a computational cost of approximately six hours for a single flow pass over the aerofoil chord on 200 Nvidia P100 GPUs resulted in moderate success for a significant portion of the flow dynamics, which not adequately predicted when compared with experiment. The latter led to a series of useful conclusions. The core of the conclusions involved the apparent over-prediction of time-averaged velocity and momentum deficits across wakes and as well as over-prediction of turbulent intensities. An identification of the problematic areas is therefore given and potential alleviation techniques outlined.