[zephyrus17]

Physics is physics, math is math, and these are always the same. Is there actually a difference in inherent 'accuracy' by using OpenFOAM compared to ANSYS Fluent, for example? Are they superior in different areas (turbine vs exterior flow)? Or are they just all the same but with different brands/setup methodology?

[cfdnewbie]

Interesting question, but I believe your argument in not fully valid. One could argue that physics is physics, math is math, so a Toyota Civic is the same as a Porsche 911.

In a way it is, they both do the same thing (take you from A to B), but in very different manners with different features, qualities and such.

Codes (commercial and research) across the board differ greatly in terms of two things:
quality and efficiency. I'm in no position to pass judgement on openfoam or fluent, but I'm very sure that they will differ significantly, like a Civic and a 911.

Most commercial codes are like plough horses: Sturdy, hard-working, a little bit chubby (i.e inefficient) and pretty good for a wide variety of tasks - but they do not excel in a certain area, and the result you get with a commercial code will - in general - be inferior to the one you get with a specialized code (in terms of both accuracy and efficiency).

Highly specialized codes (like research codes tend to be) are incredibly good at what they are designed to do, but they won't do much else - like a highly trained race horse.

What you are missing in your statement "math is math, physics is physics" is that CFD is casting math into applied maths, i.e. the numerics. Codes for the same physical problem (described by the same math) can differ so significantly in terms of numerics, that their results can be completely unrelated.

So please don't forget the sweat and brainpower of the numericist, who translates continunous maths into valid discrete numerics!

[zephyrus17]

That's very well pointed out. I was prompted to this issue when I was looking at the various commercial CFD packages like CFX, Fluent, OpenFOAM, etc and just wondered, when solving simple flow cases, whether any of them, having the exact same mesh/turbulence models/etc, would give the exact same results.

But you make a good point about the dedicate packages too.

[cfdnewbie]

I don't have much experience with commercial code (being a researcher), so I can only speak very generally. Even for simple test cases like flow over a flat plate (laminar even), I would expect to see different results, especially on the same grid. All the codes will use different discretization strategies (e.g. order of consistency), so the associated errors will be different as well.

You can argue that if you do a fully resolved simulation in each case with each code, you should get the same results - that's absolutely true, but all codes will need a different number of DOF to achieve this. So in the end you will be able to get the same results for simple cases, but the codes will differ in terms of efficiency and speed.

Keep also in mind that for the majority of flows, you will not be able to resolve it fully in a DNS sense, i.e. your approximation errors will play a role, and therefore the numerics will become "visible" in the solution.


So short answer: Ideally, fully resolved flow: Yes, no difference (if same eqn are solved in a consistent manner).
For all the rest (what we do most of the time): Differences due to different numerical approximations / errors.

hope this helps!
cheers!

[otd]

[QUOTE=cfdnewbie;335742]Interesting question, but I believe your argument in not fully valid. One could argue that physics is physics, math is math, so a Toyota Civic is the same as a Porsche 911.

Probably frivolous, but I own a Honda Civic, not a Toyota Corolla.

[cfdnewbie]

[QUOTE=otd;336256]

Originally Posted by cfdnewbie

Interesting question, but I believe your argument in not fully valid. One could argue that physics is physics, math is math, so a Toyota Civic is the same as a Porsche 911.

Probably frivolous, but I own a Honda Civic, not a Toyota Corolla.

LOL, sorry.... only German cars here, never anything else
Owned a Chevy once, broke down all the time!

[otd]

[QUOTE=cfdnewbie;336584]

Originally Posted by otd

LOL, sorry.... only German cars here, never anything else
Owned a Chevy once, broke down all the time!

Sorry! I got a bit smart-a**y there. The honda corolla or the toyota civic are both good 'generic models' for everyman's car.

My '53 Chevy rusted out before it broke down - but I get your point.

[cfdnewbie]

Hey otd,
no problem, you were well within your right to call me out on this

[Robert@cfd]

Very nice postings cfdnewbie.

Concerning this:

Originally Posted by cfdnewbie

So short answer: Ideally, fully resolved flow: Yes, no difference (if same eqn are solved in a consistent manner).

I thought even with "exact" resolving in time and space (DNS) and the same numerics (modelling, discretizations) results will be different if running on different machines (see butterfly effect). Even with double precision after long time simulations you will get different results.

[cfdnewbie]

Originally Posted by Robert@cfd

Very nice postings cfdnewbie.

Concerning this:

I thought even with "exact" resolving in time and space (DNS) and the same numerics (modelling, discretizations) results will be different if running on different machines (see butterfly effect). Even with double precision after long time simulations you will get different results.

Hallo Robert,
I have to admit this issue puzzles me as well, so all I can give you are some experiences / hunches on this:

1) The numerical dissipation of the scheme (unless one is using an non-dissipative one of course) will help you dissipate everything at the small scales, so especially round-off stuff that comes in at the smallest scale you can discretize

2) The nature of the flows I'm investigating (turbulence) also helps keeping round-off under control, since round-off error appear in a homogeneous isotropic way, as do the smallest scale turbulence fluctuations. In other words, the flip of a bit will look like dissipation at the smallest scale.

3) I've done a large number of DNS simulations, where the flow field at the smallest scale (influenced by round-off) is instantaneously different from a previous run, but the large scale statistics (mean values, skin friction, heat transfer) are the same. I don't want to go into to much detail, but the issue of "scale separation" helps us here in this case...

4) While I personally haven't seen a case where round-off indeed "butterflies" to the large scales, I think it might indeed be possible and can't be ruled out categorically.



Sorry for this rather loose collection of thoughts, I haven't been able to find a stringent way of looking at it for myself!

Very interesting topic, thanks a lot for your input,

Cheers and Merry Christmas to all!

[kid]

hi guys,
Good generic talks here, but let us not forget OpenFOAM is free of cost and Fluent is commercial.

Regards,
kid