Hi everybody.

I'm doing an incompressible unsteady calculation for a flow field around a 2D hydrofoil. I'm using the segregated solver that is shipped with Fluent5. This is (to the best of my knowledge) an implicit solving scheme using the pressure correction technique.

My problem is convergence difficulties for each time-step. When I plot some integrated value vs. iterations (not time-step), say pressure on the suction side of the foil, the value doesn't seem to converge properly. The residuals look fine, however. They jolt up for each new time step, and then drop down continously until a new time step is started.

I've tried starting the transient calculations from a manually initalized solution, and from a calculated steady-state solution. The convergence problem occurs for both cases.

My time step is aprox. 0.001 L/U.

Does anybody out there have any general information on implicit transient calculations? What number of iterations pr. time step is a "normal" value? Is it OK just using the residuals as a convergene criteria for each time step, and just neglecting the absolute value not converging? (This doesn't seem likely to me though). What inital values are normally used?

I should also mention that the calculated time-varying values for lift coefficient doesn't agree very well with experimental results.

Thanks in advance.

G. Berntsen

It's natural and shouldn't surprise you at all that your residuals jump at the first iteration of every new time step. That's the nature of transient calculations. And if the flow in question is transient, isn't it natural that the solution or any "derived' quantities including integrated pressure continue to change wiht time ?

Thanks for your answer..

I think, however, that you misunderstood my problem slightly.

The problem is that the 'derived' quantity doesn't seem to stabilize until the maximum number of iterations pr. timestep is reached, and the solver continues to the next timestep. (I plotted the pressure on the suction side vs. each iteration - not timestep.) This leads me to think that the flowfield presented as the solution to that timestep is somewhat 'arbritary', or at least not time accurate. Am I wrong here?

I should mention that i turned off the convergence check in the residuals monitor, to let the solver run all iterations pr. time step.

The natural solution to this would of course be to increase the max. number of iterations pr. timestep. I have tried this within what I feel is a reasonable number and the derived value still doesn't stablilize properly.

What is a reasonable number for the maximum number of iterations? (The default value is 20)

Of course, it is both expected and wanted that the flow field changes with time.

Thank you for helping out.

Hello Kim,

How many orders of reduction in the residuals would you recommend before concluding that a problem converged at a particular time step? What are the other things would you suggest that we check?

Regarding the timestep, I need your advice. Since the mesh for a CFD analysis is normally fine, choosing an optimal time step is necessary. What are the things you suggest to arriive at this step?

Now, to the accuracy in a transient solution using RANS equations? I personally am very uncomfortable following this approach. Unfortunately, I am forced to do it because of memory limitations. I wish I had enough memory to use LES, which I reckon to be conceptually correct.

I really appreciate your feedbacks,

Thomas

(1). I have not use the code for transient flow calculations. So, my comments are related to the transient flow calculation in general. (2). The flow over an airfoil is not a simple problem, especially when you are dealing with the lift and the drag. (3). My suggestion is that, you can find a transient flow case with analytical solution first as the test case. For exmple, I have seen people using an oscillating flat plate problem to validate their code. In this way, you can select the proper time step size and validate the code to see how many inner iterations are required to reach convergence within one time step. (4). It is possible that the convergence of the inner loop depends on several parameters, such as the mesh size, mesh distribution, the turbulence model used (if appropriate), how the equations are solved, the relaxation parameters for each equations (this is especially true for the two-equation models, and the low Reynolds number versions). (5). So, as you can see that , it takes years of experience to get the feeling about these factors. (6). The fact is that the piano alone is not going to make a person to become a professional musician. I think, a commercial CFD code is just like an expensive piano.( one would hope so.) To make an instant music, one needs to have a different type of piano, like a player-piano which can play the music by itself. (7). I think, it is the users responsibility to find out whether the piano is a real piano, a player-piano, or Charlie Brown's piano. (8). So, my suggestion is: start with something simple. Somewhere along the line, one should be able to define the range of application of the code he is trying to use. (ideally, one would like to have the information available before he buys the piano or the car. to have the code on the computer system is just like to have a piano in the student's union building. It is there, but ...)