Optimization of Savonius Turbines after improving the k-ε Turbulence Model in modeFRONTIER

The accurate prediction of turbulent flows is a fundamental issue to improve existing devices and develop new configurations. A detailed level of prediction can be obtained with Direct Numerical Simulation (DNS), but limitations in computer power restrict its application to simple configurations and low Reynolds number. Large Eddy Simulation (LES) is now feasible for many situations but it is still an expensive solution for industrial applications. Therefore, numerical simulations based on Reynolds Averaged Navier-Stokes (RANS) are still widely used today for engineering problems. In the RANS models, closure constants are introduced in order to replace higher order correlations originated during the process of averaging Navier-Stokes equation. These constants are usually determined semi-empirically based on simple flows. Nevertheless, these models are applied in quite different and complex configurations. For a particular flow, it is likely that the prediction can be improved with the adjustment of the constants and therefore a large span of values can be found in literature with values calibrated based on the experience of the user, theoretical considerations and single objective numerical optimization. The determination of the model constants for engineering turbulence models is indeed a difficult task. The values are often considered as some ad-hoc values. Changing one parameter in order to observe consequences concerning, for instance, the time-averaged turbulent velocity distribution or the shear stress distribution, is easy. But the simultaneous modification of several parameters of a turbulence model in order to increase accuracy rapidly becomes a formidable issue. If all the model parameters were changed in small steps, then the number of possible combinations would yield an enormous – and probably unnecessary – computational effort to explore the whole domain. In that case, numerical optimization techniques may help to speed-up the search procedure to find the best possible combination of the model constants with a minimum computational load, since optimization is much more efficient than a simple trial and error manual procedure