Verifying Engineering Analysis 6

[leju]

Does anybody know a method for verifying foreign complex engineering analysis like FEA-strength computation of a car, airplane etc.? Is there an established procedure to check foreign computational model, loads and results or is this up to the checking engineer how to do this?

Thanks leju

 

[corus]

A company that performs a FE analysis for design purposes is responsible and liable for that design and it is up to that company to perform the necessary QA checks on that analysis. It's not uncommon however for the company that receives that design to do its own verification of the design using a separate FE analysis, particularly when its known that that design company has inadequate resources.

corus

 

[GregLocock]

Typically we'd run element quality checks, look at what boundary conditions have been used, and try and figure out why they've used rigid links and other such abominations.

If we don't like what we see, they get asked to try again.

So far as correlating with real world data goes, that depends on the contract. Sometimes we do it, sometimes they do it. Usually we measure the real data, they fit to it.

Cheers

Greg Locock

 

[GBor]

Before handing over any results, make sure you perform a mesh sensitivity analysis. This is often overlooked, unfortunately, but take your original mesh and double it (many packages have a method of doubling the density either physically or by adding nodes at the midpoint of edges within your mesh). Make sure your results don't change by greater than some reasonable factor (10%). Also check to see if, within a given element, you have no more than about 10% of your stress scale (for example, max. stress of 10ksi, you should have no greater than a 1ksi difference between two adjacent nodes). If you are OK with these and have access to a second FEA package, this is an excellent suggestion. Also, verify the application of your FEA package against simple hand calculation for which you have a solution (dig out your old engineering books and try to duplicate some of the problems). Then, for a simple enough problem, you may also be able to run some simple tests in your office that you can compare. If the software seems accurate for these cases, the same calculation proceedure will take place for larger, more complicated situations (i.e. a car, or a plane). It gives you greater confidence that you have a good solution. Ultimately, someone probably should do a full-scale test, but if that is impractical, these things usually give me some pretty good confidence and, in the past, some pretty good success.

 

[ProbaSci]

It's common to divide errors into two broad areas, which I will refer to here as Modeling errors and Numerical errors( Does anyone has better descriptive terminology ? ). The two areas aren't really exclusive, but it's convenient to think of them as such.

Modeling errors refer to how you choose to simplify/represent the real world problem without regard to the computer/numerical issues. For example, taking a mult axial, time varying load and representing it with a single peak force. Or assuming that a press fit assembly is represented with "welded" or "ridgid" contact without accounting for the press-fit stresses. Proper selection of material properties is also included in this area. I will also add that you should have a good plan of how you will use the results before you get them. Will you be comparing to a previous design or to an engineering failure criteria such as yield stress or von Mises (I think that not having a plan for the results is one of the most common mistakes made in using FEA)? "Modeling" errors are where an experienced and knowledgeable client may be better suited than a consultant analyst at judging the fitness or acceptability of the assumptions. Your design team and analayst should discuss the modeling representation used and their appropriatness. Ideally, the analyst could show that assumptions (such as ignoring small off axis loadings) make very small contributions to the result.

Numberical errors refer to problems with descritizing the continuous model for computer analysis and numerical issues that may arise (especially in nonlinear analysis). For example, the number and type of elements used to represent the real world device, or the type of solver used in a nonlinear analysis as well as the time interval in transient analysis. These errors are more difficult for those not familiar with FE to check. As mentioned previously, a mesh density study will help show sufficient resolution. One should check for the continutity or smoothness of the depenedant variable (eg are the stresses/displacements "smooth"). Beware of results plots that are averaged because averaging can hide meshing problems as well as make judgements of stress continuity more difficult. Exagerated plots of the deformations are a good tool for this. Also, statistics of the element quality can be helpful, though poor element shape does not automatically mean the result is erroneous. As mentioned in modeling error section, in an ideal situation, your analyst could show the effect of descritization or averaging by demonstrating the effects of each parameter to the overall result (eg., that the difference in stress between 1000 and 10000 elements is negligable).

Unfortunately, its not so easy to look backwards at a result and know if errors are modeling or numerical related. Sometimes, the two errors can acutally cancel each other. Doing some manual calculations (combined with experience) is one of the best ways to validate a model and you should always include this. Trying picking some "easy" locations (the geometry is simple, or the loading components are simple) and see how close the FEA matches what you get. You can also try bounding the calculations to get a reasonable range that you expect the stresses should fall within. Real world testing is the gold stanadard, but that can have a whole other set of errors that sometimes make it less desireable than a hand check.

A final word of wisdom: I read from another FEA user (I think on Eng-Tips, I apologize to the individual for not providing credit) that his goal was to provide such detailed error checking that for someone to prove him wrong would be too expensive/timely to even try. This maybe the best advice I've seen for FEA.

http://www.probasci.com

-
Implantable FEA for medical device manufacturers

 

[CESSNA1]

LEJU: Most FEA software usually calcualtes a "residual" which should be very low. it is however, not always a reliable indicator. The best way is to go to the handbooks and verify the solution using classical engineering techniques. FEA developer's like to say you can throw away the handbooks, but that is bunk. If anything buy more. FEA programs will look for cerrtain things such as restraints and connectivity and other things, but I have seen FEA programs that had errors execute completely and give reasonable looking solutions that were wrong. Start with a small model with few elemets so that you can work up to a detailed model and see the solution converge. Then check it. Remember Boeing built and intentionally broke a 777 wing to verify the FEA analysis. FEA analysis, like other things requires engineering judgement.

Good Luck
Dave

 

[BATman2]

Leju,

In our consultancy, we do a combination of everything noted by the others in the posts above. In addition, we require the electronic files from the models to be submitted for our review. We then check everything to confrim, at least, there's nothing being "hidden". Sub-modeling of complex, high stress areas is also done sometimes to improve accuracy both for modeling and numerical accuracy (good description Probasic)
Manual calculations certianly help judge what's going on. Typically, we find the modeling errors Probasic speaks of most often in the boundary conditions, loads and supports.

In my opinion physical testing is the best check. In our line of work we require tests of structures large and small.

Batman2

 

[MToft]

Hi,
The NAFEMS Analysis Management Working Group has recently been looking at the following ASME definitions (re ProbaSci's comments on Modelling/Numerical checking):

Verification
=========

The process of determining that a model implementation accurately represents the developer's conceptual description of the model and the solution to the model.

Validation
========

The process of determining the degree to which a model is an accurate representation of the real world from the perspective of the intended use of the model.

The more important thing than mesh accuracy can be the source/conservatism/physical sense of the loading regimes, constraint locations etc - ideally you should get a report justifying these choices, demonstrating reasonable conservatism etc; if not in writing then a verbal "grilling" should give you some confidence!
The more important the application, the more difficult testing can be (eg nuclear installations)...

 

[badlam]

Hi,
Prior to obtaining test data, there is a good example of verification procedure for parts of a large model when it is developed by outside agencies. It's basically NASA Ames formalization of much of what is mentioned in the previous responses.

http://techreports.larc.nasa.gov/ltrs/PDF/NASA-95-cr4675.pdf

 

[BATman2]

Badlam,

EXCELLENT LINK. Thanks very much!!!

Batman2