On the application of Large-Eddy simulations in engine-related problems

In internal combustion engines the combustion process and the pollutants formation are strongly influenced by the fuel-air mixing process. The modeling of the mixing and the underlying turbulent flow field is classically tackled using the Reynolds Averaged Navier Stokes (RANS) modeling method. With the increase of computational power and the development of sophisticated numerical methods the Large Eddy Simulation (LES) method becomes within reach. In LES the turbulent flow is locally filtered in space, rather than fully averaged, as in RANS. This thesis reports on a study where the LES technique is applied to model flow and combustion problems related to engines. Globally, three subjects have been described: the turbulent flow in an engine-like geometry, the turbulent mixing of a gas jet systemand the application of flamelet-basedmethods to LES of two turbulent diffusion flames. Because of our goal to study engine-related flow problems, two relatively practical flow solvers have been selected for the simulations. This choice was motivated by their ability to cope with complex geometries as encountered in realistic, engine-like geometries. A series of simulations of the complex turbulent, swirling and tumbling flow in an engine cylinder, that is induced by the inlet manifold, has been performed with two different LES codes. Additionally one Unsteady RANS simulation has been performed. The flow field statistics from the Large-Eddy simulations deviated substantially between one case and the next. Only global flow features could be captured appropriately. This is due to the impact of the under-resolved shear layer and the dissipative numerical scheme. Their effects have been examined on a square duct flow simulation. An additional sensitivity that was observed concerned the definition of the inflow conditions. Any uncertainty in the mass flow rates at the two runners, that are connected to the cylinder head, greatly influences the remaining flow patterns. To circumvent this problem, a larger part of the upstream flow geometry was included into the computational domain. Nevertheless, the Large-Eddy simulations do give an indication of the unsteady, turbulent processes that take place in an engine, whereas in the URANS simulations all mean flow structures are very weak and the turbulence intensities are predicted relatively low in the complete domain. The turbulent mixing process in gaseous jets has been studied for three different fuel-to-air density ratios. This mimicked the injection of (heavy) fuel into a pressurized chamber. It is shown that the three jets follow well the similarity theory that 152 Abstract was developed for turbulent gas jets. A virtual Schlieren postprocessingmethod has been developed in order to analyze the results similarly as can be done experimentally. By defining the penetration depth based on this method, problems as typically in Schlieren experiments, related to the definition of the cutoff signal intensity have been studied. Additionally it was shown that gaseous jet models can be used to simulate liquid fuel jets, especially at larger penetration depths. This is because the penetration rate from liquid sprays is governed by the entrainment rate, which is similar as for gaseous jets. However, it remains questionable if gas jet models can in all cases replace the model for fuel sprays. The cone angle for gas jets can deviate strongly from those observed in spray experiments. Only when corrected for this effect, the penetration behavior was similar. Two turbulent diffusion flames have been investigated with a focus on the modeling of finite rate chemistry effects. Concerning the first flame, the well known Sandia flame D, two methods are compared to each other for the modeling of the main combustion products and heat release. These methods are described by the classical flamelet method where the non-premixed chemistry is parameterized using a mixture fraction and the scalar dissipation rate, and a relatively new method, where a progress variable is used in non-premixed combustion problems. In the progress variable method two different databases have been compared: one based on non-premixed flamelets and one based premixed flamelets. It is found that the mixture fraction field in the Large-Eddy simulation of Sandia flame D is best predicted by both the classical flamelet method and the progress variable method that is based on premixed chemistry. In these cases the flame solution was mostly located close to its equilibrium value. However, when correcting for the prediction of the mixture fraction in the spatial coordinates, it is shown that the progress variable method based on non-premixed chemistry is better, compared to experiments. Especially at locations where a flame solution near chemical equilibrium is not adequate this model is more appropriate. Additionally a sooting turbulent benzene diffusion flame has been investigated. Therefore a steady laminar flamelet library has been applied which is based on a very detailed reaction mechanism for premixed benzene flames. In the Large-Eddy simulations the total PAH/soot mass and mole fractions have been computed explicitly, while the source terms for these variables are based on a classical flamelet parametrization. The regions of PAH/soot formation have been identified, showing distributed parcels where PAH/soot formation takes place. The results show a growth of PAH/soot volume fraction up to levels of about 4 ppm. The average particle size increases steadily in this flame, up to about 30 nm