Compressed natural gas (CNG) has been gaining attention in the automotive industry since it allows for a substantial reduction in carbon dioxide emissions. Recent advancements in direct-injection technology have contributed to an increase in volumetric efficiency and engine power, but the physical modeling of such gaseous injection with Computational Fluid Dynamics (CFD) is still challenging. This work describes the development of a CFD methodology for the simulation of CNG engines, an evaluates its predictivity by varying certain simulation parameters.
Firstly, turbulence modelling has been investigated by comparing the RNG k-ε RANS, Dynamic Smagorinsky LES and Delayed DES approaches. The DES provides the best trade-off between computational cost and accuracy during fuel injection, but the model makes abrupt transitions between its RANS and LES-like regions, leading to inaccuracies during combustion. The expensive nature of the LES and the adequate accuracy of the RANS model lead to the preference of the latter. Secondly, the turbulent Schmidt number was investigated. By lowering the number, mixing is more dominant in the simulation and the mixture formation becomes more stoichiometric. The influence on combustion is even larger because of turbulent transport across the flame, with extremely low values even leading to falsely-predicted engine knock. Thirdly, by comparing multiple consecutive engine cycles to each other, it was found that the first cycle still contains a large amount of initialization error in the turbulent field. The second and third cycles show much better agreement to each other. Finally, the amount of turbulent fluctuations at peak power and their influence on laminar flame speed partially lead to flame extinction at the start of combustion. As a consequence, the G-equation model is not justified and does indeed provide inaccurate results for the heat release during combustion. Even though the SAGE detailed chemistry solver in conjunction with the Gri-Mech 3.0 mechanism is assumed to have a larger computational cost, it was found that the opposite is in fact true, and the latter should be preferred.
The aforementioned conclusions are used to calibrate the simulation model against measurement data. The recommended settings are the RANS turbulence model, SAGE detailed chemistry and a turbulent Schmidt number of 0.6, while considering results from a second engine cycle. The Lower Heating Value (LHV) of the fuel was decreased in order to account for an unforeseen inconsistency with the measurement. This ad hoc solution showed that the predictivity of the RANS simulations can substantially be improved, and a good agreement between simulation and measurement was observed. The calibration is however not consistent with other operating conditions, as the part load analysis showed. Finally, the comparison between the peak power and part load operating point have revealed that the engine performance can potentially be improved by altering the piston bowl shape or the position of the injector and the spark plug.