[arvhult]

Hi guys! I am totally new to this forums, and am looking for some input on a matter of simulating purely buoyancy-driven flow.

I have a model of a vertical pipe with inversely stratified water, the interface being in the centre. The top water is 370 K and the bottom is 470 K. I have used NIST to produce my own polynomials for water properties. The pressure is 7 MPa, and I am absolutely certain the polynomials are valid outside of my temperature interval. The solver is buoyantPimpleFoam, so that heat transfer is also modeled.

So when relaxed, the water should mix by buoyancy forces. My 250 cm long pipe with 20 cm diameter doesn't diverge, but multiplying the length by 8 (and thus also the number of cells) does. The pressure field reaches 1001 iterations in every time step, and after about 0.9 seconds the residual is too great. Any other parameter is solved for in 1 iteration. It is apparent to me that most probably the buoyancy forces become too great, and the vast acceleration of the fluid parcels simply make the physics problematic. I myself understand the physics and fluid mechanics, and recieved help from a teacher in setting the numerics, e.g. correctors and relaxation, etc. My only set boundary conditions except zeroGradient in temperature, and no-slip in U, is 7 MPa pressure on top, and fixedFluxPressure on the walls.

I have set nOuterCorrectors to 3, and nonOrthogonalCorrectors 0 (changing the latter didn't make a difference, I found). I have a relaxation factor of 0.5 for all parameters. in controlDict i juse adjustable time steps, with a maximum Co of 0.5. With 8 CPUs, I simply decompose the pipe vertically and solve in parallel.

Other than the lengthening, the model is completely identical to the one that did not diverge. QUESTION: Is there any way to make this model work, other than refining the mesh so much that it becomes unpractical to finish the simulation within reasonable times? I am running the finer mesh now (3 million cells), but so far there are only slightly lower residuals in p_rgh.

(I am validating the model against an experiment, so I need at least 150 seconds of the simulations. With 8 CPU's at hand, that time is precious to my project.)

Thank you for any input