[AlexRonto]

Hello,

I am trying to set up and run a steady state simulation for a truck using Acusolve (FEM solver). I am using the k-epsilon turbulence model (Spalart-Almaras for first try) and I have some troubles regarding the near wall treatment schemes implemented.

Having tried models both with and without wall functions, I end up that my model is prohibitely huge for a No Wall Function scheme (demanding y+=1 at all cell nodes) so I choosed the Wall Function option (y+ ranges from 0 to 80). Monitoring the residuals I see that the No Wall Function model (see Pic.) has a much faster convergence and the curves are obviously smoother than the model with Wall Function option (see Pic.). The results seem are less accurate though, maybe due to high y+ values (unfortunately I cannot plot the y+ contour).

So my question is, is there any possibility that the Wall Function for the cells near to the wall could affect the residuals' behavior? Is this the reason for the oscillating behaviour for the residuals? How could I justify the slower convergence for this model?

Thank you in advance.

Alex

[sbaffini]

While there might be any kind of issue due to how wall functions are implemented in a code, wall functions can obviously have an effect on convergence, including a stall as in your case. But one should know how such wall functions are implemented to understand what the issue might be, besides knowing your specific case.

If you are allowed to by the code, you should try to plot the residual, to see where it is higher. If it is near a wall, as everything seems to suggest, then check what y+ values you have there and what's happening there in general (probably y+ is changing at each iteration).

Nonetheless, your plots do not seem fair enough, compare the two formulations for the same number of iterations.

[AlexRonto]

Originally Posted by sbaffini

While there might be any kind of issue due to how wall functions are implemented in a code, wall functions can obviously have an effect on convergence, including a stall as in your case. But one should know how such wall functions are implemented to understand what the issue might be, besides knowing your specific case.

What exactly do you mean by "implemented"? As far as I know Acusolve uses a standard log-law wall function...

Originally Posted by sbaffini

If you are allowed to by the code, you should try to plot the residual, to see where it is higher. If it is near a wall, as everything seems to suggest, then check what y+ values you have there and what's happening there in general (probably y+ is changing at each iteration).

Unfortunately I don't have the choice of post processing for the residuals (or at least I haven't find it till now). But I suppose you mean that if I find local y+ oscilaltions then the residuals' oscilaltions could be due to other reason (e.g. mesh..) instead of the near-wall-treatment model applied...

Originally Posted by sbaffini

Nonetheless, your plots do not seem fair enough, compare the two formulations for the same number of iterations.

OK, I could turn off the convergence criteria. What you mean is that if the no-wall-function model runs for the same number of iterations, then I could make a judge about the STABILITY of my solution (not the speed of convergence), right..?

Really helpful answer. Thank you!

[sbaffini]

Originally Posted by AlexRonto

What exactly do you mean by "implemented"? As far as I know Acusolve uses a standard log-law wall function...

Well, this is not what they state:

"AcuSolve utilizes a two-layer model on the dissipation rate equation. At low y+ values, the solver uses an algebraic expression to compute the dissipation rate. As the distance from the wall increases, the solver blends the value of the algebraic expression into the solution of the differential equation, then eventually transitions to using the differential equation solution fully beyond a specific y+. This two layer treatment produces a robust and stable solution from the k-epsilon models and is the default wall function for all three variants of k-epsilon."

While this does not say anything about the wall function for the velocity, it is typically the case that this is also in a form that is y+ insensitive for the two layer formulation (otherwise there is no much sense).

However, I see that you are using SA from the residuals, which only needs a WF for the velocity (the SA variable is just linear in the whole boundary layer).

The point is, an y+ insensitive formulation, if properly calibrated for a given turbulence model, can give actual y+ insensitive results. But you can't bet on the developers always doing the right thing, my experience is quite the contrary (I include myself in this).

If a standard wall-law is used for the velocity (in SA or k-epsilon), what happens is that the specific formulation has a switch based on y+ instead of a single formula valid for all y+. Thus, if the solution keep switching between the two branches from iteration to iteration, the solution won't converge.

If an y+ insensitive formulation is used, and all the remaining parameters of your simulation are correct, you shouldn't see this sort of residual stall (yet, residual histories can generally be different).

Originally Posted by AlexRonto

Unfortunately I don't have the choice of post processing for the residuals (or at least I haven't find it till now). But I suppose you mean that if I find local y+ oscilaltions then the residuals' oscilaltions could be due to other reason (e.g. mesh..) instead of the near-wall-treatment model applied...

What I meant is that if y+ oscillates from iteration to iteration, and it does so by going across limiting y+ values for the branches of your (NON y+ insensitive) wall function, than this can cause residual stall and/or oscillation.

Originally Posted by AlexRonto

OK, I could turn off the convergence criteria. What you mean is that if the no-wall-function model runs for the same number of iterations, then I could make a judge about the STABILITY of my solution (not the speed of convergence), right..?

Yes, compare the residual histories for a given, equal, number of iterations, say, 1000 (100 could be too low). If they both converge, even if with different rates, there is nothing wrong in principle.