I am trying to simulate a thermal plume case with LES, which is based on a Smagorinsky model with a constant coefficient (Cs=0.1) and an eddy-diffusivity model for the turbulent heat flux, with a constant turbulent Prantdl number (Pr_t=0.72). The important paramters are Re=1000,Fr=1,Ma=0.2, and the temperature ratio of hot plume and the neviromental air is 3. The computational box is y=6(lateral direction),z=20(streamwise direction). The corresponding DNS grid I used is ny=320, nz=800. I tried a grid system which is ny=120, nz=400 for LES. But unfortunately it crashed. So far, I have tried to do many adjustments, like adjusting the model constants, switching off the model terms in enery equation for testing, artificially setting the SGS terms to be some part (20%) of rhs terms of the original equations for testing, but all failed. The only successful LES case I got is to use the same grid as in DNS. But it seems too strigent for me to use such a high grid resolution. I am even not sure if this is a numerical or physical problem. Will the buoyant effects in thermal plume require more considerations for a routine Smagorinsky model? Do you have any suggestions for this? Thanks a lot.

Try reducing the time step by an order of magnitude. Let us know what that did.