I have read manual, and it says that k epsilon does not give better results for seperated flow. It underpredicts attchement point. Can anybody pls tell me that what may reason. i.w. which terms are preventing k epsilon to predict these thing? Thanks all.

For the backward facing step, there is mainly two reasons : 1) anysotropie is not predicted (otherwise we are talking about RANS model) 2) the spatial position of the dissipation distribution is wrongly predict in the shear layer, since production of dissipation is comming from the shear stress and not from the Kolmogorov cascade. So, the turbulent viscosity is not correctly predict, and the velocity in the recirculation region is not well describe.

That's just my opinion after few years working on multiscales turbulence model.

"production of dissipation is comming from the shear stress and not from the Kolmogorov cascade"

It is old for me now but I remember that the k-e model was calibrated on shear layer experiments (or jets). Am I becoming bedridden?

Yes, you are wright, the value of c_eps1 could be calibrated to give the correct value of uv/2k in the Boussinesq approximation, in the case of homogeneous shear stress. But in this case the velocity profil is forced, not resolved.

Maybe my formulation were a little bit confusing. Talking about backward facing step, the production terms of k and Epsilon take their maximum on the path line just behind the step, in the region were the stress is maximum, and mostly at the same position. Which is not confirmed by LES results.

Using a multiple scales model, you can change this behavior and get a result with a better approxiation of the recirculation region - but you won't have solve the anysotropie pronblem.

A more easy way than multiscale model, is to introduce a sort of turbulence's history variable. I had work on variation of such a model, developped by Lumley in the early 90's, something like a k-Epsilon-S model.

The very simple answer is that RANS-based models cannot capture the large-scale structures that exist in separated shear flows. The reattachment length is governed by the evolution of these quasi-two-dimensional coherent structures, and the lack of these structures in k-e model solutions leads to very poor results.

See the Flow over a Backward-Facing Step (Spanwise Vorticity) animations on http://www.stanford.edu/group/ctr/gallery.html for a very clear visualisation of what I'm talking about.