[unred486]

Hey guys,

I am thinking of building a CFD solver that solves steady state 3D-conduction problem with a source. I have a fairly complicated geometry(all cartesian) that is usually solved with a commercial solver(flotherm or icepak). It is still just a box with a bunch of cartesian grids. No curvature. But with a pretty big aspect ratio(~10^3 or 4). But I want to write my own solver just for the hack of it.

Seems fairly simple as I have built a CFD solver myself before for my research. But I have always used a FDM and commercial solvers usually use FEM or FVM. The geometry is entirely box-like so a box like Cartesian grid is good enough. But I am not sure if FVM or FEM implementation is absolutely needed. I would like to avoid building FVM or FEM solver if possible since I don't have a direct experience in building a code based on these concepts.

Will FDM be enough for my problem? does anyone have any guidance?

[FMDenaro]

Originally Posted by unred486

Hey guys,

I am thinking of building a CFD solver that solves steady state 3D-conduction problem with a source. I have a fairly complicated geometry(all cartesian) that is usually solved with a commercial solver(flotherm or icepak). It is still just a box with a bunch of cartesian grids. No curvature. But with a pretty big aspect ratio(~10^3 or 4). But I want to write my own solver just for the hack of it.

Seems fairly simple as I have built a CFD solver myself before for my research. But I have always used a FDM and commercial solvers usually use FEM or FVM. The geometry is entirely box-like so a box like Cartesian grid is good enough. But I am not sure if FVM or FEM implementation is absolutely needed. I would like to avoid building FVM or FEM solver if possible since I don't have a direct experience in building a code based on these concepts.

Will FDM be enough for my problem? does anyone have any guidance?

The main difference is that FVM are conservative by definition. And is also feasible on complex grids.

[bigfootedrockmidget]

If you are into FEM, you can have a look at FEM in 50 lines of matlab:


https://link.springer.com/article/10...:1019155918070
https://www.math.hu-berlin.de/~cc/cc...software.shtml


There might be somebody out there who has done the same with FVM...
A FEM solver does not have to be large and complicated if you make use of the existing capabilities of your language of choice.

But if your mesh is structured Cartesian, then maybe a simple finite difference method will be sufficient.

[sbaffini]

Few random suggestions here:

1) Properly define the scope of this experiment as having a productive code for your scenario is more difficult than you might expect. If you haven't done this before, learning must be, at least, 50% of the reason to do it.

2) If learning is the main purpose, go for it with FDM. But if it is no more than 50% of the reasons, my suggestion is to stick to recipes that are proven to work well. Maybe just try to replicate the commercial solver you were using before.

3) I'm not even sure I need to write this but, if productivity is needed, you need to consider a lot more aspects: parallel, IO, monitors, actual robustness and accuracy, etc.

4) Cartesian geometries do not automatically translate into a simple treatment. Even the simplest approach (approximating the true geometry with the smallest cartesian block fully embedding it) requires a proper handling of the grid, which at bare minimum requires tagging cells that are out of the domain (to be skipped), if not a multiblock approach and/or immersed boundaries (which are a thing in their own)

5) To answer your specific questions, there is a reason why FV has gained the largest share of commercial codes. You can do it differently, but a lot of expertise with FD is simply missing (i.e., some things routinely done with FV have never been done with FD).

So, again, do it, do it, do it... but don't think you're gonna have a commercial grade product in 6-12 months (especially if you haven't made one before) only because you have cartesian grid and geometries.

EDIT: Ok, I might have overemphasized, it's still just conduction what we're talking about, but the general principles I mentioned are still valid