Hi

There are many schools of thought on this and I believe that the answer to your question both is Yes and No.

It is Yes, when both your model and your test data truly are accurate and valid.

Undoubtedly, the the answer is No when either or both are inaccurate and/or in some sense diverge.

To answer why, ponder this simple example of an inverse mini problem - if I tell you that the answer is four (4) - please tell me what is my question? The crux is that the answer solution space perhaps is infinite or at least very large, as, to exemplify the answer might be 1+3, 2*2 or for that matter - the number of shoes positioned in a cupboard.

Model updating can be made in a forward or inverse manner.

The forward approach relies on first principles and can be done by hand or with an optimizer. This type of model updating, I find, is the more robust approach as it allows input data to be limited into sensible values.

The inverse approach relies on the core assumption of a very good model/test match.

For what it is worth. My experience is that inverse procedures that rely on test data work best for small scale models, in practice, often limited to models of solving, say, 2 by 2 matrix problems.

Some ramblings of mine on the subject with practical advice can be found here

https://qringtech.com/learnmore/why-simulate-measure-correlate-automate/

Presentation Analysis Technology:; A discussion on some of the fundamentals

Article FE Model Correlation & Mode Shape Updating using Qualificati...

Last but not least. A forgotten item in model correlation is that of test and model geometry matching. If there is not a good match between assumed and actual geometry several first principles become violated.

Practical problems in this realm are things like

• Most pipes are neither straight, nor round.
• Sheet metal plates are not flat and do not always arrive in a stress free initial state
• Component assemblies are not made with zero tolerance, as assumed in the geometry model. Geometry tolerance may in the assembly cause pre-stressing that affects output, e.g. through plate buckling.

All in all - there is a big world out there. It is easy to go astray and, I find, model updating is one of the easier ways to get lost.

That said, I do highly recommend working with model QA as it to my knowledge is the only path to systematic modelling assumption QA. A short cut to better and more accurate models by inverse procedure, in my experience - not so much.

Sincerely

Claes

Hi

There are many schools of thought on this and I believe that the answer to your question both is Yes and No.

It is Yes, when both your model and your test data truly are accurate and valid.

Undoubtedly, the the answer is No when either or both are inaccurate and/or in some sense diverge.

To answer why, ponder this simple example of an inverse mini problem - if I tell you that the answer is four (4) - please tell me what is my question? The crux is that the answer solution space perhaps is infinite or at least very large, as, to exemplify the answer might be 1+3, 2*2 or for that matter - the number of shoes positioned in a cupboard.

Model updating can be made in a forward or inverse manner.

The forward approach relies on first principles and can be done by hand or with an optimizer. This type of model updating, I find, is the more robust approach as it allows input data to be limited into sensible values.

The inverse approach relies on the core assumption of a very good model/test match.

For what it is worth. My experience is that inverse procedures that rely on test data work best for small scale models, in practice, often limited to models of solving, say, 2 by 2 matrix problems.

Some ramblings of mine on the subject with practical advice can be found here

https://qringtech.com/learnmore/why-simulate-measure-correlate-automate/

Presentation Analysis Technology:; A discussion on some of the fundamentals

Article FE Model Correlation & Mode Shape Updating using Qualificati...

Last but not least. A forgotten item in model correlation is that of test and model geometry matching. If there is not a good match between assumed and actual geometry several first principles become violated.

Practical problems in this realm are things like

• Most pipes are neither straight, nor round.
• Sheet metal plates are not flat and do not always arrive in a stress free initial state
• Component assemblies are not made with zero tolerance, as assumed in the geometry model. Geometry tolerance may in the assembly cause pre-stressing that affects output, e.g. through plate buckling.

All in all - there is a big world out there. It is easy to go astray and, I find, model updating is one of the easier ways to get lost.

That said, I do highly recommend working with model QA as it to my knowledge is the only path to systematic modelling assumption QA. A short cut to better and more accurate models by inverse procedure, in my experience - not so much.

Sincerely

Claes

Next to the good answers provided above, I would like to add a word of caution.

It is generally speaking very dangerous to 'calibrate' your model for purpose X, to then start using it for purpose Y. What you mostly do in practice is, given a set of measurement data, set your model parameters such that the responses match in a reasonable way. Since you can safely assume that your explicit model contains all sorts of linearisations and other sorts of approximations, the approach where you calibrate using modal data (of which the model is usually way easier to approximate) is prone to failure. As dr. Fredö mentioned, it is a big world out of there.

An alternative approach could be to resort to non-deterministic model updating, as it, instead of providing a one-to-one mapping in parameter space, it tries to calibrate your model for a subset of the parameter space you are considering as a total. Maybe a recent paper of mine can guide you in this:

Article A multivariate interval approach for inverse uncertainty qua...

Although the application there is uncertainty quantification, the change to model calibration is a pure philosophical one, since the methods are exactly the same.

I hope this helps!

Best regards

Matthias

Fabian Dwenger , Claes Richard Fredö , Matthias Faes

Thank you all for your helpful answers.

Sincerely,

Mohammad

In my experience with extremely accurate optical measurements, it was again of utterly importance to have a very detailed model, stick to the real geometry and dynamic properties of the test. In this way, model updating was trying to fine tune the main parameters in restricted ranges, but there was not a speedy convergence, as the initial FE model was already adherent to the test setup and received a lot of attentions. You can check it in particular the latest

Article Full field optical measurements in experimental modal analys...

Best regards,

Alessandri

I wanted to add a thing to the FEA model specification.

The seat frame model will eventually be used in a rear impact crash simulation which takes place in a split second.

Regards,

Mohammad

Hi Mohammad

For what it is worth - not my area of expertise.

I have seen verification test rigs where crash loads are simulated as a static load.

The rationale is that the peak load comes from the peak acceleration times the load it carries. Naturally, such a static test does in no way describe events during or after the crash but it does in a simple and straightforward way simulate the peak load that can be used to qualify the equipment's function, i.e. that it is not overloaded.

/C

Alternative arrangement may be used, I think hanging model in free- free condition using thin metallic wire using appropriate support structure.

Experiential modal analysis normally gives you results in modal space while explicit FEA models your systems in PHYSICAL space. There is no direct way to connect both together.

Another challenge is that a modal analysis (analytic or experimental alike) is based on a linear system. In the impact analysis, it is likely to be nonlinear, because the structure may experience both material and geometric nonlinearity.

As the structure gets deformed, some portion get into contact with other portion. As the result, your stiffness matrix changes, even if your material is still within yield limit.

@ Zane Yang

Thanks for your contribution.

With regards to connecting modal analysis and physical model, they for sure can be compared with each other, since we are doing FE modal analysis on our FEA model as well and the results from both can be compared.

Regards

Mohammad

May not be reliable, as creating a prototype is an approximate and compromised task.

What parameters you have determined from FEA?? Because by using experimental modal analysis you can determine only FRF, FREQUENCY, DAMIPING, and MODE.

Jai Kumar Sharma

FEA determines the same parameters that you get from a typical FEA modal analysis, i.e. natural frequencies and modes. We can also get damping but we are not interested in it.

Sorry above, I was partly wrong as I thought about linear models only.

If it is worthy, please do not forget that a crash deals heavily on non-linearities, therefore the tuning must take care of it. In my previous contribution I was stating that a successful tuning is a non trivial task also in the lab for linearly behaving structures, when they are complex and their dynamics is fully crowded by strong contributions. Non-linear modal analysis might be quite tricky with experiments, but... Once you grasped the risk in your non-linear variables (as functions), then you might want to simulate in the time domain the evolution of the crash, paying attention to the energy absorption behaviour.

Traditional modal analysis and related model updating are dealing with a specific linearisation point: to use these methods on non-linear systems you might start by multiplying these linearisation points and see how the modal base is changing and affecting your variables in related tuning.

Kind regards,

Alessandro

Further words: any impulsive events, like very short transients, carry broad frequency band contents in the dynamics of a system. Therefore, the locally linearised equivalent system might describe also the approaching of resonant excitations, which means the structure exhibits very low stiffness and can collapse at the speed of sound in the material also when it can statically be loaded by the same amount of static forces without issues. Indeed static stiffness is much different from frequency dependent stiffness.

In the upcoming Article Flexible multibody system dynamics by means of a spectral ba...

Conference Paper ON THE EFFECTS OF FLEXIBILITY IN A QUICK RETURN PLANAR MECHANISM

Hope this clears you my contribution,

best regards,

Alessandro

Alessandro Zanarini

Thank you very much for the clarification and useful guides.

Best Regards