[AndreP]

Hello guys.

NO, THIS IS NOT ONE MORE TOPIC ASKING FOR HOW TO REDUCE RESIDUALS! I PROMISE!

Well, of course this come after still having high residuals in a simulation, but after a long search about residuals, I couldn't find one place that could tell me how to interpret the residuals! Not even the help file (that only have mathematical interpretations... I know that's what they are, but of course there's a more "practical" way to interpret them).

I know for instance, that Tdr mostly relates to the quality of the prism layer, while the continuity refear for example more with how is our physics correct!

Is there anywhere I can search/read that has a more practical interpretation of the residuals? (Im using K-Epsilon Turbulence model by the way).

What I'm looking for, would be ideally, an explanation of what each Residual is calculating/evaluating and what does that translate into my CFD Simulation settings (as the 2 examples I gave above).

Is there such a thing?

[AndreP]

Before you all go mad at me... YES! I turned off the Normalization on all the residuals!

[me3840]

Originally Posted by AndreP

Before you all go mad at me... YES! I turned off the Normalization on all the residuals!

Why would we do that? There's nothing wrong with normalized residuals.


The residual is a measure of imbalance of a particular equation. A high momentum residual in a cell represents that cell having a large imbalance of momentum. That's very physical.

[AndreP]

Originally Posted by me3840

Why would we do that? There's nothing wrong with normalized residuals.


The residual is a measure of imbalance of a particular equation. A high momentum residual in a cell represents that cell having a large imbalance of momentum. That's very physical.

I know, but EVERY topic I've read about someone asking how to reduce residuals, that's the first answer always people give! Just wanted to give a heads up that I've done that already!

[me3840]

Reducing the residuals by unnormalizing them is just as good as arbitrarily dividing them by 1000. Make yourself feel great and divide by 100000!

It's just a normalization. It doesn't change what's actually happening with the numerics.

[AndreP]

From what I've read on this forum and help section, if the normalization is on, it'll normalize from the first 5 iterations, and so, the residuals will be very dependent of the Initial Conditions I impose. So I choose to turn that off, so the initial conditions don't influenciate the residuals value.

But that was my interpretation, it might completly be wrong!

[LuckyTran]

You only need an intuitive feeling for what it is, because you rarely need to use it for anything. In Star-CCM the residuals are quite literally residuals, if you have equationA=0 as a governing equation, you'd evaluate the transport equation and get a non-zero residual at each cell. Then take the root-mean-square over all cells. Then normalize.

What's not useful about using residuals in this way is that your transport equations are not non-dimensional. The residuals have to be interpreted w.r.t. the particular problem so that you cannot say anything about whether a 1e-06 is good enough. It is even more obvious when you realize the residuals should have units... Yet another reason why you want to normalize by something.

Note that other software often scales the residuals based on mean momentum in each cell to make the reported residual less scale dependent. These can easily be interpreted as % residual.

Originally Posted by AndreP

From what I've read on this forum and help section, if the normalization is on, it'll normalize from the first 5 iterations, and so, the residuals will be very dependent of the Initial Conditions I impose. So I choose to turn that off, so the initial conditions don't influenciate the residuals value.

But that was my interpretation, it might completly be wrong!

The point of dividing by the worst residual in the first 5 iterations is so you divide by a very bad number so that you can see the residuals decrease and feel that the solution is improving. Why 5? Well because your initial guess is normally a uniform flow with no imbalance except at boundaries, i.e. a very small residual! In Star-CCM you can see exactly what this normalization number is and see that it is quite arbitrary as mentioned earlier. You can even switch to manual normalization and enter in any number you like, such as 100000.

Originally Posted by AndreP

I know for instance, that Tdr mostly relates to the quality of the prism layer, while the continuity refear for example more with how is our physics correct!

I disgree. The continuity residual has very little physical meaning except as a sanity check because just about any decent solver should satisfy continuity unless the solution is diverging. Continuity doesn't tell you how well your momentum and energy balances look. Hence, the continuity residual is the least physically meaningful for making these judgments.

Residuals, at least the way we calculate them, don't tell you where the solution is wrong, unless you plot the field of local cell residuals. But then that would be a very different discussion. The reason you can probably guess that the prism layers are the culprit when your tdr residual is high is because all the turbulence happens in the shear layers and boundary layers. And if you are simulating wall bounded flows this correlates nicely with walls and ah-hah prism layers. Try simulating a laminar free shear layer or a mixing layer between two different fluids at different temperatures. You can even amplify the effect by intentionally coarsening the mesh and putting terrible quality cells where these interfaces occur.

In general, what you are looking for is seeing residuals decrease if your solution is still converging. That's because the rms is dominated by the largest local residual. Of course this doesn't tell you whether all the cells are improving or if the worst local residual is just moving around from cell to cell. Eventually the residual should not improve anymore as the solution converges and stabilizes (the residuals will become asymptotic). Of course you can also end up with a case where the residuals do not converge but also do not diverge and you see residuals bounce up and down forever. Then you have to go hunting for where the problem is. When you get here, you are probably looking at the local fields to find the problem and not looking at the residual report.

[AndreP]

Thanks a lot for your extensive answer. Although not what I was looking for (because what I was looking for, from what I see doesn't exist), I got the feeling that unfortunately there's not a direct association of the residuals with the physical correctness or incorrectness or our simulation!

I was hoping some residuals would refer most of the times to certain aspects of the physics or mesh setups, but after reading what you said and conjugating with the help file, I see that's not always the case!

Thank you for your time and help!