LARGE-EDDY SIMULATION OF SCALAR DISPERSION FROM A POINT SOURCE OVER A WAVY WALL David A. Philips Department of Mechancial Engineering Stanford University Stanford, CA 94305 dphilips@stanford.edu Riccardo Rossi CTFD-Lab Seconda Facolt` a di Ingegneria di Forl`ı Universit ` a degli Studi di Bologna Via Fontanelle 40, 47121 Forl`ı, Italy Gianluca Iaccarino Department of Mechancial Engineering Stanford University Stanford, CA 94305 ABSTRACT A Large-Eddy Simulation (LES) of scalar dispersion over a wavy wall has been performed and is compared with Direct Numerical Simulation (DNS) data for the same case. LES statistics show excellent agreement with DNS data at a much lower computational cost. The LES mesh cell count has been reduced by a factor of eight and the timestep doubled as compared to the DNS calculation. Mean flow and Reynolds stress data is presented along with mean scalar and scalar flux statistics. Barycentric maps are used to visualize the state of turbulence in the flow field captured by LES while scat- ter plots of mean scalar concentration give an overall picture of predictive performance. For both of the prior analysis tech- niques, RANS results are included in the discussion to place the LES results in appropriate context in terms of computa- tional expense and accuracy. INTRODUCTION From tracking the spread of the release of a hazardous material to predicting the impact on air quality due to emis- sions, there are many important applications for scalar disper- sion modeling. The high population densities found in urban areas make them important regions in which dispersion must be understood (Britter and Hanna, 2003). The complex ge- ometries found in urban areas pose an additional modeling challenge. With the advent of automated unstructured mesh- ing techniques, Computational Fluid Dynamics (CFD) meth- ods are suited for scales at which individual buildings must be resolved. Comparisons of Reynolds-Average Navier Stokes (RANS) and Large-Eddy Simulation (LES) approaches to dis- persion modeling are still common in the literature for both real urban areas (Gousseau et al., 2011) and idealized geome- tries (Salim et al., 2011) as the best method is still up for de- bate (Fernando et al., 2010). This LES study builds upon prior investigations of dis- persion (Rossi et al., 2010; Rossi, 2010; Philips et al., 2010) Figure 1: Computational domain in which various RANS methodologies were explored. RANS simulations in complex geometries were found to yield rea- sonable results when the turbulence models effectively cap- tured anisotropies in the Reynolds stresses. Additionally, a scalar flux model that takes advantage of this anisotropy infor- mation is required. The Standard Gradient Diffusion Hypoth- esis (SGDH) uses a scalar value for the diffusivity that does not always capture the correct behavior of the turbulent scalar flux. Tensorial formulations of the diffusivity are better suited to incorporate flow anisotropy information in the scalar field prediction. In the current work, the anisotropy in the turbu- lent flow predicted by LES will be examined. If the modeled subgrid stresses and fluxes are truly isotropic, then large scale anisotropy should be captured without the need for the type of models employed in the RANS approach. NUMERICAL EXPERIMENT An LES of fully developed turbulent flow through a wavy channel is performed at a Reynolds number of approximately 6, 800 based on the mean channel height, 2h, and the bulk ve- locity. A passive scalar with a Schmidt number, Sc = 1, is released from the wave crest and allowed to develop down- stream. The accuracy of this LES study will be assessed via comparison with DNS data presented and analyzed in 1