Hi, I am starting LES calculation and would like to clear some points. 1)How to decide about the grid and size of the filter.? 2)How much time I should run my LES calculations. 3)How to specify the intial condition. Is it important to resolve the wall boundary or it is enough to have wall functions. pp

1) As a preliminary remark, unless you plan to use a dynamic model, you most probably don't need a filter, the filtering process being implicit through the grid (Shannon theorem).

The correct grid resolution near a solid wall can be a posteriori deduced from turbulent minimal criteria coming from channel flow computations : x+~10 to 80 (streamwise), y+~1 (normal to the wall), z+~ 2 to 30 (transverse). Away from the wall, you can look at the minimal lengthscale of the physical phenomenon you want to resolve (if you know it). With a typical second order scheme, you need more than 12 points per lengthscale to fully resolve the wave. With higher order or dispersion optimized schemes, the number of points is reduced.

To analyze the grid resolution, one rough and simple way is to (time) average the subgrid viscosity and to look at its relative magnitude compared to the physical viscosity. It should be O(10) or so, not O(100) or higher like in RANS computations. Otherwise, the classical way is to compute the spatial Turbulent Kinetic Energy spectrum and check for the cutoff wavenumber lying in the inertial range. The problem is that it is difficult to compute this spectrum if there is no homogeneic direction. So you can use a similar approach but with spatially localized TKE time spectra (i.e. Fourier transforms of velocity time series for some selected grid points)

2) First you need to go beyond the transient process. Look at the time some of integral value (drag, ...) and wait until the value does not significantly evolve anymore. Then acumulate statistics until first and second order moments (and higher if you're interested with) are converged.

3) A realistic initial condition should be a way to reduce the transient time. So use an (approached) analytical solution of your problem as initial condition. If you don't have analytical solution, a computational one (from potential to RANS or even DNS with lower Reynolds number) can be retained. However note that, from my experience, a very clever initial condition only marginally reduces the transient time compared to a rougher one.

3') It depends on your problem. For a canonical boundary layer or a separated one under geometrical singularity (corner), a wall function could be a possibility though it is not well regarded in the LES community. In all other cases, it should be avoided.