Large eddy simulation of separating flows from curved surfaces
The capabilities and limitations of LES in predicting separation from curved surfaces at high Reynolds number are at the centre of this Thesis. Issues of particular interest are mesh resolution, subgrid-scale modelling and near-wall approximations aiming to reduce the computational cost. Two cases are examined: a flow separating in a channel with streamwise periodic constrictions (hills), and the flow around a single-element, high-lift aerofoil at a Reynolds number of 2.1 . 106. Prior to these studies, fully-developed channel-flow simulations are considered. These show substantial differences among subgrid-scale models in terms of the subgrid-scale viscosity magnitude and its wall-asymptotic variation. Modelling and numerical errors appear to counteract each other, thus reducing the total error. Wall functions axe shown to be a cost-effective approach, providing a reasonably accurate approximation in near-equilibrium conditions. Adequate resolution remains critical, however, in achieving successful simulations. In the hill flow, separation occurs downstream of the hill crest, reattachment takes place about half-way between two consecutive hills and partial recovery occurs prior to a re-acceleration on the following hill. A highly-resolved simulation, performed to produce -benchmark data, permits an extensive study of the flow properties. Coarser mesh simulations are then compared with the former. These highlight the influence of the streamwise discretisation around the separation point and the role played by the implementation details of the wall treatments, while the subgrid-scale models influence is less significant. The aerofoil, which features transition and separation, is extremely challenging and at the edge of current LES capabilities. None of the simulations reproduce 2 the experimental data well. Indications on the sensitivity to various parameters, including the numerical scheme, the mesh resolution and the spanwise extent, are extracted, however. The studies indicate the need for a structured mesh of about 80 million nodes to achieve the required accuracy. For the present study, this was unaffordable.