The development of spark ignition engines with increased efficiency and reduced emissions requires simulation tools capable of predicting cycle-to-cycle variations (CCV). Large-Eddy Simulations (LES) have this potential, but typically fail to reproduce cyclic variability observed in experiments. Although evidence for the correlation between early flame kernel development and CCV was provided long ago, there is still a lack of fundamental understanding of the initial combustion phase, and accurate LES models do not exist. In this work, these shortcomings have been addressed. First, multi-cycle LES with low numerical discretization error have been realized by integrating state-of-the-art methods and models into an existing Cartesian LES code. The application of the enhanced LES framework to compute CCV in a production-type engine is demonstrated. Second, flame kernel/turbulence interactions and differential diffusion effects have been analyzed based on a new engine-relevant direct numerical simulation (DNS) database. Run-to-run variations in the global heat release rate of flame kernels are shown to be caused by flame area dynamics. A novel analysis attributes this effect to flame kernel distortion by large-scale turbulent flow motion with characteristic length scales greater than the flame kernel size. Differential diffusion is shown to substantially reduce the global burning rate, thus increasing CCV tendency. The analysis shows that the large positive global mean curvature of flame kernels may detrimentally affect the local mixture state inside the reaction zone, which can initially be compensated by spark energy supply. Further, it is shown that differential diffusion effects under engine-typical Karlovitz numbers are not weakened by small-scale turbulent mixing.weiterlesen