G'day guys,

I'm currently conducting a mesh dependence test for a natural convection problem (steady state). I am using 3 models with approximate mesh sizes of ;60x60, 120x120 and 240x240. The first two models converged nicely, and they both had similar looking residual plots. Ie initially some waves etc, but then they reached a point and continued to fall down in a relativeley straight line.

My problem is with the 3rd model (240x240) I am having extreme troubles getting the continuity residual down. Generally speaking if I make the under-relaxation factors small, the continuity residual will fall as well, however it will then level out quickly and over time will not change. So basically I can make the under-relaxation factors tiny O(10^-6) to obtain an acceptle continuity residual. I am also monitoring the Nu number and in a similar manner, i the residual is dropped, the Nu changes accordingly.

I am curious if anyone has any ideas why the first two models would converge nicely and the third model, with a finer mesh simply refuses to converge. The first two models are ok compared with experiments and other literature, however the third model (240x240) will not converge even with the small residuals discussed in the previous paragraph.

The only reason I can put this down to is the mesh itself. The model is bassically a square, and I'm using a structred mesh of rectangular elements with a small grading ratio. Due to the way the model is meshed, as the mesh density increases, the aspect ratio of some elements on some of the walls also increases O(100~200) for 240x240.

Even so I'm still asking if anyone has any other ideas that may help me on my quest. Any ideas are welcome

Cheers

I have an approximate answer to offer although I did not face such a problem. I will tell what I know about refinement and residuals. In some cases, over-refinement in mesh can lead to problems because of the reason you stated at the end of your post. too disparate dimensions in rectangular meshes result in mass imbalances in two directions (x and y). Secondly, it is not advisable to reduce URFs too much. Any way, URFs artificially dampen changes in variables and too much dampening can lead to unrealistic results. So, I suggest that you should try using a square mesh and do not use very small URFs.

I hope this will give some directions.

Swarup.

i think you change the factors and method of your multigrid model. my email is waitting for your good news.