CFD Online Discussion Forums

CFD Online Discussion Forums (http://www.cfd-online.com/Forums/)
-   CFX (http://www.cfd-online.com/Forums/cfx/)
-   -   First order simulation better than second order (http://www.cfd-online.com/Forums/cfx/101334-first-order-simulation-better-than-second-order.html)

Torque_Converter May 1, 2012 13:42

First order simulation better than second order
 
I am finding that the first order solution is much closer to test than second order. I am using the same convergence criteria (max res 5e-4), same boundary conditions, same fluid, etc. Just changing the convergence order. I start with first order/upwind, then from those results go high resolution on both turbulence and pressure, momentum. First order is within 8% of test torque, and second order is 115% over pedicted torque. Mesh changes definitely affect the first order result, but not second order.

This is all a little concerning, as I don't want to just use first order with a mesh that happens to match test as this serves little purpose in development when without test data.

ghorrocks May 1, 2012 20:11

The apparent accuracy of the first order solution is probably just luck. You need to refine the mesh such that you get a mesh independant solution - you will probably find this occur for the second order system with a coarser mesh than the first order one.

Your comment that the first order system is more sensitive to mesh changes supports this.

Torque_Converter May 1, 2012 23:16

Thanks, getting lucky seems right since it obviously not consistent.
I have a question thou about the coarseness of the mesh thou. The mesh is pretty fine right now, with inflation layer first layer of .003 mm, 15 layers, and body mesh of .4 mm with the domain being approximately 80x80 mm at its largest dimensions. This produces a mesh of about 2.5 million. This probably overdoes it. However, how can I know when the solution is mesh independent. Let's say 500,000 cells gives a second order solution that is correct. But then when i jump to 1.5 million cells it goes back to being over 100% overpredicted. That seems mesh dependent.

ghorrocks May 2, 2012 02:05

When a mesh becomes too fine the following problems occur:
1) numerical round off means spurious errors start creeping in. This can be helped by using double precision numeric.
2) Sometimes the reduced dissipation of a finer mesh is not fully offset by the turbulence model. This means that psuedo-turbulent fluctuations start occuring and affects the result. You can normally see this in the the results file as the residuals do not achieve tolerance (for steady state runs) or you can see transient structures in the flow which (for transient runs).
3) For transient simulations a finer mesh will probably need a smaller timestep to keep the Courant number about the same. You might need to do a new time step sensitivity check for the finer mesh.
4) Likewise, the CFX convergence tolerance is pretty good at being independant of mesh size, but sometimes it is not perfect. You might need to do a new convergence tolerance sensitivity check for the finer mesh.

This is why sensitivity checks are slow. Once you have checked convergence tolerance, time step size and mesh you have probably found a new time step size, tolerance and mesh is required and you have to do the exercise again. It is an iterative process as you converge on the best mesh, convergence tolerance and time step size.

Torque_Converter May 2, 2012 13:15

Thank you for all your help. Starting with a coarser mesh and looking at first and second order I see that second order is consistently about 100% high, while first order is all over the place. This shows me that as you said first order is just luck if you happen to stop at some mesh count, while second order is mesh independent past a certain point.

I think perhaps the 100% overprediction is not a facet of first being somehow better than second order (which I did not feel good about before!) nor is it a facet of mesh, but somehow the physics isn't right. Perhaps there is cavitation present, or some other factor that reduces the torque production.

ghorrocks May 2, 2012 19:12

Quote:

somehow the physics isn't right
I agree. Something important has been missed.

Good work for actually looking into this and figuring this out - many people don't bother and they fool themselves into thinking their model is accurate.

Torque_Converter May 2, 2012 22:00

Since you have been so helpful, for myself and on the site in general I wanted to let you know any interesting find on this in case you were curious.

The parasolid I have has sharp perpendicular wall intersections and sharp blade geometry. I then got a look at the actual device and noticed everything was rounded off to 3mm fillets and has drafts. I asked around and found this was necessary since it is sand cast and needs these features to be manufacterable. Tomorrow my new direction will be to study the effects of these relatively large variations from "ideal" sharp geometry. I imagine it will be significant considering it is in high pressure and key momentum change regions.

Thanks again

immortality January 8, 2013 05:07

hi glenn
I have a 3mm duct with 6cm in length.i have set courant number to .1 for convergency but results are not as expected.how should i do?my grids are 20*500.should i make them finer?how much?and CO lower?
Thanks.


All times are GMT -4. The time now is 21:47.