[aero_head]

Originally Posted by arjun

Yes pretty much. But thanks to your comment now I have again started to think of it for particle size distribution. I spent 2 months with population balance and I could not find reliable method to offer so I put it aside until I could come with with reliable method. May be there might be help in neural nets here. My problem is reliable moment inversion methodology. So far I do not like what I have.

Yes, the way the ANSYS CFD package solves particle tracking is simply... just not correct. Our lab developed our own in-house code to tackle this. So I just thought the recruitment of neural nets for this problem would be intriguing to say the least. It seems like these neural nets could be implemented in (nearly) every field!

I now see why my supervisor spent like the majority of his academic life studying particles and their motions/distributions. Much like turbulence, there is still much to discover in this field.

[arjun]

Originally Posted by FMDenaro

Have you ever tried to generate a bad mesh for solving flows in a simple geometry test-case? Like backward facing step flows or pipe/channel flows?

Yes. Many times.

[arjun]

Originally Posted by aero_head

Yes, the way the ANSYS CFD package solves particle tracking is simply... just not correct. .

Yes, indeed the case. It seems using 3 to 4 moments is not enough.

BTW I want to bring attention to one very interesting paper

https://www.researchgate.net/publica...eakage_kernels


Their approach is is quite interesting, but still marred with problems with moment inversion. Nonetheless it is quite easy to implement but capable.

[aero_head]

Thank you for sharing that paper, Arjun.

From skimming through the article, the math does not look too complex, and the approach was indeed presented insightfully.

The moment inversion has always been the problem child, I suppose. I am hoping that, like turbulence, the solution lies with the development of more powerful computers. Today I suppose there are just too many unknowns that we must just assume, but hopefully more powerful computers (and maybe even neural networks) could change more unknowns into certainties.

I'll have to read that paper, and possibly some of the references they provide, more in-depth. Thanks again though, it was a good jump-off point for me.

[JBeilke]

Originally Posted by arjun

Yes. Many times.

This reminds me of a conversation with Prof. Peric many years ago, when polyhedra were just invented.
He told me that they (at cd-adapco) now develop and test all models with polyhedra, because errors in the modeling become visible in the first place.

[arjun]

I am attaching some results to demonstrate that making solver stable does not mean that solver shall become less accurate.

I chose two points (mesh1 and mesh2) to give an idea of what type of mesh we are dealing with. These are not the worst cells. I just picked two areas.

The plot is for CD coefficient for DriveAir benchmark case that is availble online. The calculation was run with starccm by a japanese company and I am supposed to compare the wildkatze results on the mesh provided with same turbulence model (k Omega in this case).

As we can see third order solver can predict the drag very good.

PS: I think 11 million is too much mesh but this is what is provided for this so.

PS2: The jump at around 1200 iterations is due to the fact that I realized that solver was running with first order upwind for convection term (everything else was third order). It was switched to third order upwind there.