Quantification of spatial and temporal concentration profiles of vapor clouds resulting from accidental loss of containment of toxic and/or flammable substances is of great importance as correct prediction of spatial and temporal profiles can not only help in designing mitigation/prevention equipment such as gas detection alarms and shutdown procedures but also help decide on modifications that may help prevent any escalation of the event. The most commonly used models - SLAB (Ermak, 1990), HEGADAS (Colenbrander, 1980), DEGADIS (Spicer & Havens, 1989), HGSYSTEM (Witlox & McFarlane, 1994), PHAST (DNV, 2007), ALOHA (EPA & NOAA, 2007), SCIPUFF (Sykes, Parker, Henn, & Chowdhury, 2007), TRACE (SAFER Systems, 2009), etc. - for simulation of dense gas dispersion consider the dispersion over a flat featureless plain and are unable to consider the effect of presence of obstacles in the path of dispersing medium. In this context, computational fluid dynamics (CFD) has been recognized as a potent tool for realistic estimation of consequence of accidental loss of containment because of its ability to take into account the effect of complex terrain and obstacles present in the path of dispersing fluid. The key to a successful application of CFD in dispersion simulation lies in the accuracy with which the effect of turbulence generated due to the presence of obstacles is assessed. Hence a correct choice of the most appropriate turbulence model is crucial to a successful implementation of CFD in the modeling and simulation of dispersion of toxic and/or flammable substances. In this paper an attempt has been made to employ CFD in the assessment of heavy gas dispersion in presence of obstacles. For this purpose several turbulence models were studied for simulating the experiments conducted earlier by Health and Safety Executive, (HSE) U.K. at Thorney Island, USA (Lees, 2005). From the various experiments done at that time, the findings of Trial 26 have been used by us to see which turbulence model enables the best fit of the CFD simulation with the actual findings. It is found that the realizable k-epsilon model was the most apt and enabled the closest prediction of the actual findings in terms of spatial and temporal concentration profiles. It was also able to capture the phenomenon of gravity slumping associated with dense gas dispersion. (c) 2011 Elsevier Ltd. All rights reserved.