Is LES of reacting flows predictive? Part 1: Impact of numerics

Combustors in modern aerospace propulsion systems are highly optimized devices. Further enhancements to their performance will require the existence of novel, highly predictive modeling capabilities. LES simulation has often been proposed as a leading candidate in this vein. Its predictive ability, however, has remained a challenge. In this paper a methodical study is performed to assess the impact of numerical architecture on the predictive ability of LES. The bluff body stabilized flame experiment of Volvo Flygmotor AB is simulated as a canonical representation of many realistic combustion systems. Four codes are employed, namely two research codes, the open source code OpenFOAM and the commercially available code Fluent. Physical models and computational grids are kept identical. Reacting and non-reacting solutions are obtained. Results indicate that all codes effectively reproduce the non reacting flow in terms of large scale unsteady vortex shedding behavior, mean profiles and second order turbulence statistics. In contrast, none of the codes is truly predictive in the reacting case. While all solutions approximate the experimentally observed behavior, significant qualitative and quantitative discrepancies remain. These range from a lack of prediction of large scale physics such as shedding mode behavior and recirculation zone features in some cases, to inaccuracies in the prediction of the mean velocity and temperature profiles and related second order turbulence statistics. These discrepancies are significant enough to raise questions as to whether the true physical manifestation of the flow is being reproduced. More importantly, however, these discrepancies are frequently not consistent between the various codes used. While the consistent discrepancies can be attributed to the limitations of the common physical models employed, the inconsistent ones point to a significant impact of the numerical architecture on the solution. It is recognized that the detailed root cause driving this behavior will require further study which is likely to yield necessary enhancements in the way LES is employed both in terms of numerical enhancements as well as the definition of new numerical metrics for application. The need for higher fidelity, more discriminating experiments for LES validation that include detailed dynamic information is also pointed out.