I think we cannot really compare different techniques based on absolute accuracy here.

Experimental results are always preferred since those are real situations.

The accuracy of analytical techniques depends mainly on how accurately we model the system physically and mathematically. Thus, we can only define how accurate is our model and not how accurate is the analytical method.

However, in comparison of numerical methods, exact analytical techniques will be more accurate for a same model because of intrinsic approximations in numerical techniques. But in many complex situations, exact analytical solutions are not available or are difficult to obtain. Thus in some cases we also try approximate analytical methods which provide better understanding of system than numerical techniques.

In really complex problems however, numerical techniques are more preferred because of their simplicity.

I think we cannot really compare different techniques based on absolute accuracy here.

Experimental results are always preferred since those are real situations.

The accuracy of analytical techniques depends mainly on how accurately we model the system physically and mathematically. Thus, we can only define how accurate is our model and not how accurate is the analytical method.

However, in comparison of numerical methods, exact analytical techniques will be more accurate for a same model because of intrinsic approximations in numerical techniques. But in many complex situations, exact analytical solutions are not available or are difficult to obtain. Thus in some cases we also try approximate analytical methods which provide better understanding of system than numerical techniques.

In really complex problems however, numerical techniques are more preferred because of their simplicity.

Hi

Take a look at here.

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

Simulation could only capture a small part of the problem but, in the end, simulation proved to be more accurate than test. That said, this is the only time I have come across this situation.

Here are some examples from past work - take a look at the slide where the same FE model is run through two different software.

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

Some of my ramblings on simulation, test, model QA and automation is found here. http://qringtech.com/learnmore/why-simulate-measure-correlate-automate/

/Claes

Pranav and Fredo mentioned the most important. Anyway allow me to add some words. Essentially, let here to mention the concepts of idealization model, verification of results and validation of results, even if some discussion is still running on their definitions.

Remember that, it is about words / concepts and accurate answer means that answer should be according to exactly to "truth" or to a reference / standard and it is a good idea to underline with respect to what it is (or when it is inaccurate).

In *verification of numerical solutions you may say that they are accurate if they are equal (enough "equal") to the analytical solution given by a theory (or its mathematical / physical model). Anyway, depending on the condition scale of the model formulas, sligth deviations in inputs may result in considerable deviation in results.

In this context an analytical result from a model (*idealization) is free from error when no simplifications (or mistakes including deviations in inputs) were done in its deduction. Numerical solutions in some cases can be equal to the analytical solution, but in general they involve round-off errors depending on the numbering representation choice, series truncation errors, and others from the approximations considered.

In *validations of numerical or analytical solutions, you may say that they are accurate if they are "equal" to the experimental solution given by experimental results (let say a mathematical / physical model). Anyway the validation, to be considered "complete", may require an envelope of scenarios .

The reasons for the inaccuracy may be found in the model assumptions, input data deviations, computation assumptions, procedure assumptions and deviations, etc. There are plenty of books / texts on each subject, see Fredo suggestion, and I hope someone with more knowledge in these fields can help to improve these comments.

Dear All

Manufacturing tolerance and boundary conditions matter.

To exemplify,

• There was a test plate milled from a solid piece of Persepex. This 'product' was flat and rectangular and there was little variation from one item to the next.
• Consequently, the FE model was just as good or, possibly, better than the test. Test/FE MAC numbers could be 0.996 and similar.
• Buy an off-the-shelf Perspex plate of identical dimensions and the situation will differ very much as it will cuve, warp and may contain prestress.
• Emboss a metal plate and measure it free free. There will be very small variation from one item to the next.
• A FE model that relies on CAD, i.e. uniform stffness will be poor while one that accounts for the embossing process will do the job nicely.
• Mount this very same plate into a built up structure with assembly tolerances and the plate may buckle. You will now have lots of variation that is very hard to predict.
• A free rotor is easy to predict. The same part assembled in its bearing resting on a foundation is another matter.
• A pipe is modelled as straight and perfectly round. A mass produced pipe has a deformed cross section and banana shape along its length direction.
• Pipe supports are anything but classic boundary conditions

Last, then you have the documentation level to account for. When you find As Built stamped on the drawing - expect the worst.

/C

It totally depends on you. Some problems require complex and expensive gadgets and apparatus to measure a variable, sometimes these facilities and tools are outdated or not well calibrated or adjusted or even not available. So the results will have errors. sometimes the operators and technicians made mistakes. On the other hand numerical modelings is challenging as you actually have to simplify some complicated concepts. Although you have to be an expert in modeling to optimize these simplifications. Commercial packages are like functions, garbage in;garbage out. You have to judge the accuracy and the credibility of results. Thus, everything can contain error, you have to verify your experiment and results based on credible resources.

i think this depends on the problem itself, what do need exactly.