[nickninevah]

Hello All,

Can anyone here think of a good reason to not do a mesh independence study on a new model?

I have just started working for a company that does some CFD and quite a bit of FEA work. At least for the FEA, they do not conduct Mesh Independence Studies for their analyses. They have found a few industry guidelines that recommend a mesh density to use when investigating stress concentrations. But this guide was never intended as a general rule for the whole mesh.

I was taught that one should ALWAYS do a mesh independence study on any new model. Whether used for CFD or FEA, it is still a discretized domain that approximate the derivatives and as such, the level of error must be quantified and studied. Is there any good reason that one would not conduct mesh independence studies?

Thanks.

[harishg]

Industrial requirements are different from academic interest. They do not want to sit for days to analyze the problem and perform a grid independence study to select the optimum grid.

The only good reason for most is saving time. Another reason might be that many people may not understand its importance.

[Jade M]

Certainly mesh independence studies are very important ... you are absolutely correct.

It is true about time in industry versus academia. Also some industry folks simply may not have the education and/or background to understand and it may be difficult to educate them. I am kind of facing the same issue in a small group of analysts who are relatively new to analysis. In some cases, the engineers have not even quantified the error which I believe is highly dangerous especially if there is no testing. One engineer does not apply any refinements to larger problems which she thinks her computer cannot handle. Rather than letting the workstation chug away overnight or meanwhile making the argument for more resources, she simply runs with the default mesh. Another engineer feels that he has a sense for a mesh that is going to generate results that are accurate to within 5%. However, I reviewed some of this work and the error was far greater than 5%. Even with experience, I think that at least some refinement must be conducted to show grid independence. The physics and geometry of any given problem may be just different enough that even an experienced person may not know for sure. There could be all kinds of reasons why people do not understand this -- It's hard to know what people think and why. If you find out, I'd love to know! To be honest, I do not see the point of conducting the analysis if a person does not know whether the solution is accurate. I would recommend pushing for the studies especially if you find that the company is using results with unknown error. In any case, I say stick to your guns and do what you think is right rather than letting a place decrease your standards.

I have been steadfast in conducting these studies despite opposing advice. When I can, I refine until the error is less than 5%. When I am not given that option, then I simply carefully document the results with at least one refinement so that I quantify and error, i.e., present the results for the default mesh and the results for a refined mesh, provide the number of nodes and elements for each mesh, and state that the maximum value of the variable changed by Z percent. I have reports documenting 50% error in the solution so clearly I have not indicated that the answer is definitive.

Hopefully at some point, your company will see that you are right and will recognize the value and contribution of your work. If not, then you may find another work place more suitable.

Good luck!

[CapSizer]

It depends on where the computational analysis fits into the greater scheme of the project. It may be crucial to do a full mesh independence study, for example if it is the primary or only analysis or measurement being done. In fact, it would be more correct to say that you need comprehensive quality control for that, and mesh independence is just one part of it. On the other hand, in the early concept stages of a design, when you know that there will be much detailed testing and simulation to follow, there is a far greater need for turnover than accuracy. Mesh consistency then becomes a lot more important than mesh independence. There is a definite trade-off between quality control and throughput, and it is incorrect to insist that a certain level of quality control is appropriate for all work.

[Jade M]

Very interesting point, CapSizer! What do you mean by mesh consistency?

At my company, nobody has explained this, which is very interesting. I do not quite know the story in my workplace. It is understood that 5% error is acceptable although we do not use good practices to ensure this. Conceptual design might be conducted in another department.

Still, if it is in the early concept stages of design, wouldn't you want to know what kind of error is in the solution )even if this just means one refinement to quantify the error)? I'd be curous about the recommended error for conceptual design. I suppose this is probably hard to quantify across the board.

In my (limited) industry experience, the two companies at which I have worked have not used CFD during conceptual design stages. So I am curious as to when it is beneficial to use CFD. It's interesting to know how things work.

I'd be curious about your further thoughts. I'm fairly new to industry and am interested to understand how things work.

[CapSizer]

Well, the question to start with is the issue of the 5% error. If you already know what the absolute answer is (or how else would you know that the error is less than 5%?), why are you bothering to do CFD?


The mesh consistency comes in when you are trying to compare candidate designs in order to make a selection. In this case, you are only interested in relative answers, and you can often get a very good (and fast!) relative answer by comparing two coarse models, provided that they are equivalently coarse. For example, you may know that you need 100 chordwise cells on a wing to get a mesh independent answer, but you may be able to quantify the difference between two candidate wings by only using say 40 chordwise cells .... But then you need to use 40 for both, not 100 for one and 40 for the other.


In conceptual design (and even more detailed design), it is typical to use fast design level analysis methods (for example classic theory, calibrated with some empirical data). Very often you have only a vague idea of how accurate the method is, partly because it won't be equally accurate for all designs or operating conditions. But that doesn't stop you from using it, because design is an iterative process of escalating accuracy and certainty. So you start with a "rough" method, and refine progressively with more detailed methods and testing. You need a quick and dirty method to start with, in order to investigate enough design options. Now there is nothing on earth preventing you from using CFD in exactly the same way, starting off with quick low-fidelity models, and then moving to higher accuracy when needed. This works because fully qualified CFD is very slow, and very expensive, so you only use it where it is necessary.

[Jade M]

Wow, thanks so much CapSizer! This is very helpful and I appreciate your thoughtful response. I tried to add to your reputation again, but it won't let me until I add to others! Take good care and thanks again!