CFX overwhelming Fluent in mass convergence of boundary layer separation case
 

I use both Fluent and CFX to optimize airfoils. I tried both of them. A lot features of Fluent make me like it very much. But one drawback of Fluent totally denies the selection of Fluent for my case.

Fluent doesn't converge as well as CFX for flow with boundary separation, especially the mass flow continuity convergence. Mass flow continuity convergence is essential for airfoil CFD (I guess that’s the most important convergence criteria for most of the other cases as well). If we look at the CFX simulations convergence histories, their mass convergence is always better than their momentum convergence, which is just opposite to the same case run by Fluent. If there's boundary layer separation, Fluent will have much much larger difficulty than CFX at mass convergence. I guess that's the drawback of partially explicit algorithm of Fluent compared with CFX's fully implicit algorithm.

Any comment on this though? Does anyone have similar experience?

Any comment or experience?

First, can you define mass continuity convergence? And are residuals involved in this definition?

Just look at Fluent Theory Guide 1.2.1. ‘The Mass Conservation Equation’.

The first convergence criteria in Fluent is typically the ‘continuity’. The continuity equation is the mass conservation equation.

Residual is defined in Fluent User Guide 28.15.1.

Yes there are mass conservation equations and residuals but convergence criteria are another story.

I am asking how you are defining convergence, which is not a Fluent definition. Reduction of residuals cannot be used as a definition of convergence (only relative convergence) and this is especially true for the continuity residual. Convergence should be defined in terms of actual solution values. To demonstrate:

If you initialize the flow with the perfect solution, your continuity residual will be stuck at 1. You may observe that continuity is not decreasing and conclude that your solution won't convergence even though the solution is already exact. The reason for this is because of the way the continuity residual is defined, the local residual is normalized by the worst residual encountered in the first five iterations.

If the worst residual is very bad in those first five iterations, then you are dividing by a very big number and your continuity tends to be smaller. On the other hand, if the worst residual is very good, then your continuity residual tends to be near unity. Depending on the solver settings, you can get wildly different results for the first 5 iterations. Using the numerical value for the continuity residual puts the superior numerical solver at a disadvantage in the comparison.

Yes. Both Fluent and CFX are defaulted to give the relative residual as you describe above. But I won’t think Fluent can give initial condition much better than CFX as both of them suggest the residuals value shown in them shall reach better than 10^-4 and they calculate or read the initial condition in almost the same way.

So there’s no clue to tell that Fluent’s 10^-4 residual converges better than CFX’s 10^-4.

I gave the example merely to point out that continuity residuals is a terrible way to define convergence. You also have to consider the differences in the solver, not just the initial conditions. There are a fair number of options under the hood that are not the same for CFX and Fluent (especially in the AMG setup). Some of the default settings for Fluent are fairly aggressive compared to the defaults in CFX or Star-CCM (especially the AMG schemes and parameters) whereas others are fairly conservative. These differences can appear in after a large number of iterations.

If you want to compare convergence, residuals is not the way to do it. Compare actual solution values, i.e. do some benchmarking.

Okay, I understand. You don’t think this drawback is caused by the difference between fully implicit algorithm and explicit algorithm. You think that’s because the default parameters in Fluent are too aggressive and I shall not use the default parameters to solve the BL separation case.

I will try to reduce those algorithm parameters in Fluent to make a try.

Regarding the convergence, not only Fluent has continuity residual fluctuates at relatively high value, but also the output parameters, lift and drag, fluctuate as well. So I think the convergence shall be really not as good as in CFX.