[123Mark]

Hi all,

Is there consensus in the CFD community about a proper approach to mesh convergence for hybrid meshes (with structured regions and unstructured regions)?

For a structured mesh, the approach can be very simple, just multiplying the number of nodes in each dimension by a certain factor.
For example: For a coarser mesh, halving the number of nodes in each direction, and then working with a mesh with 1/8th of the total cells compared to the original mesh (in the 3D case).

For hybrid meshes it's less clear, as the unstructured regions behave differently to the structured regions. It's often not as simple as, (using the same example again):

1. Halving the number of nodes in structured regions
2. Increasing the stretching ratio in unstructured regions
3. Repeating step 2 until achieving the overall 1/8th decrease in number of cells

Achieving the "1/8th" goal in step 3 can result in very very large stretching ratios for unstructured regions. (Eg. greater than 3)

Is there a general consensus about what's appropriate for hybrid meshes like these?
Might anyone in the community have advice or be able to share their own methods?

Cheers!
Mark

[LuckyTran]

My advice is to just don't be lazy. It is clear what needs to be done, just do it and then be satisfied.

Too often people lounge around staring at the problem and seek second opinions. Days go by. Weeks go by. Months go by and they still haven't done anything. The meshing and simulation could have all be done in a few days.

There is a clear recipe in the cookbook, follow it. The maximum baking time to make bread is 30 minutes. But if you keep worrying about the best environmental conditions for making the bestbread, then you end up spending way more than 30 minutes, and you still have no bread.

[123Mark]

Thanks for the fast reply, Lucky! While I appreciate the encouragement and advice to stay proactive, I was hoping that my post would stimulate more specific discussions about the technical aspects of possible approaches to mesh convergence for hybrid meshes.

I'm particularly interested in hearing about any practical experiences or insights people in the CFD community may have encountered when dealing with similar challenges.

Cheers,
Mark

[LuckyTran]

Here's my practical experience. Advice is bullshit.

You have a bunch of knobs you can turn. Just turn them all. Mesh them all. Run them all. And then let the data tell you answer rather than trying to get it from a magic 8 ball.

People think that that halving the cell size and getting 1/8th volume reduction will yield a nice monotonic asymptotically converging plot it doesn't. Even if you have this plot... People think that the plot will tell them this is the mesh sizing you need to use. It doesn't. The convergence plot tells you a sort of numerical uncertainty you will have at each setting. The user still needs to decide how much error they are willing to deal with.

Next, in my 15+ years now doing CFD, I've never seen anyone half the cell size except when I did it. I've seen tons of people start a grid convergence study, change the total cell count by a laughable 10% and call it quits. The difference between 10% and 800% is pretty much 800%.

Then, a 800% change in cell count is not practical. You're never going to use a 1 million cell grid for one engineering design and then suddenly go to a 8 million cell grid. It's not even representative of the mesh sensitivity that you have when you make tweaks to your design. The halving of cell sizes is for studying the order of convergence, and most people don't give a crap about this order.


Btw, it's been way more than 30 minutes already. You could have been done with the mesh sensitivity study already if you had not waited. Rather than looking for data, just generate it yourself.

[FMDenaro]

Generally, you can think something about a non-dimensional mesh ratio between the max and min volumes. A good way is described in the Ferziger, Peric, Stree textbook.