code_saturne User's Forum

Hello. I made calculation starting from "good" initial approximation and it was successful with real viscosity and SOLU discretisation with blending factor 0.75. Thanks for your help!

Here is an algorithm I used to solve this problem.

Stage 1
Fluid viscosity: 500 x Real (suitable for air).
Target CFL: 1.0.
Discretisation scheme: Upwind.
Limits/relaxation: none.
Flux reconstruction: all variables.
Iterations: ~100.

Stage 2
Fluid viscosity: Real.
Target CFL: 1.0 (same as before).
Discretisation scheme: Upwind.
Limits: Velocity components => +/-70 m/s, Pressure => +/-1500 Pa.
Relaxation==relaxv: Velocity => 0.1, Pressure => 0.1, Turbulence: 0.5
Flux reconstruction: all but turbulence.
Iterations: ~100.

Stage 3
Discretisation scheme: SOLU with blending factor 0.75.
Other: the same as before.

I think this algorithm may be simplified, maybe fitting in just 2 stages with SOLU scheme on second stage. Particular pressure and velocity limits can be adjusted for specific case, in my case there is an air-cooled heat exchanger with low pressure drop and low velocities. But it's rare that non-fully-coupled solver will reach convergence in just one stage even with k-epsilon model (in Fluent, we also use first order scheme for the first stage, although there is always relaxation for pressure, velocity and turbulence). So, with RSM, we need 2-stage approach to converge in non-fully-coupled solver. By the way, CFX with it's coupled solver diverges easily with RSM (the same is for Fluent), it's not a simple model to reach convergence with.

It would be nice if options for all variable limits and relaxation will appear in Saturne GUI. I used cs_user_parameters.c and cs_user_extra_operations.c to implement limiting and relaxation, but it would be simpler to adjust GUI settings.

Another thing I encountered with the fan model. It's not connected with common calculation stability. Fans need some relaxation too. I run another geometry now with the same other parameters. Calculation is stable even with SOLU (blending 0.75) from the beginning, that was never seen before (without relaxation and limiting). But fan pressures started to diverge when they was approaching their expected values. It's because with real fan curves I will not obtain fixed mass flow that I need. So I make fan curve hard (stiff) to force correct flows. It usually works but in this case fan pressures diverged incridibly. Example: 315 Pa, -879 Pa, 3042 Pa, -1.5E4 Pa, 1.7E4 Pa. Please, introduce a relaxation coefficient for fan pressure in fan model. Also, fixed fan flow would be useful (selection between fan curve or fixed mass flow - both options are needed in engineering practice).

Update.
I needed to calculate the same setup with different geometry on RSM. I used simplified stabilizing algorithm and all things was good. No any convergence problems appeared despite of RSM that is known to be very unstable. Method I used is as follows.

Stage 1
Discretization: Upwind for all variables.
Relaxation: relaxv=0.1 for velocity and pressure, relaxv=0.5 for turbulence.
Limitation: Velocity components: -70...70 m/s, Pressure: -1500...1500 Pa (it's for particular setup).
Flux reconstruction: All but turbulence.
Gradient reconstruction: Non-orthogonal faces threshold.
Time stepping: Steady, local time step.
Target CFL: 1.0 (ref step: 10^-5 s, minimal factor 0.0001, maximal factor 1000).
Fan curves: hard, for "fixed flow", no correction needed.
Air thermophysics: real values, no correction needed.
Iterations: 460 (I think, 100...300 is enough).

Stage 2
The same but SOLU for Velocity with blending factor blencv=0.8.
Iterations: 150.

I attached case XML (please don't forget to introduce relaxation and limitation via user functions).

Resulting average fan head is only 2.6% higher than with k-epsilon.
Please note that if your fan head (pressure difference) needed is low while fan curve is too hard (you intended in fixed flow to determine fan head needed) the solution will diverge. As I can see, there are no means in Saturne to manage fan model in this aspect (given head, flow or curve; relaxation factor) so make curve more soft at least at first stage.

If anybody has problems with convergence you may use this approach. On k-epsilon and SST you don't need 2 stages. Just start with SOLU + blend=0.8 for velocity. With RSM 2 stages may be needed, although, maybe just one stage will also work.