As @ Douglas Vinson Nance said it is based on the application. However, if it is plus or minus 10% variation will be good.

The answer to your question, in some sense, depends upon the application you have in mind. If you are interested in simulating steady flow around a 3D wing, then I suspect the codes can deliver less than one percent error. In fact, CFD algorithms are now being for full aircraft design including lift and drag simulation even in the more difficult transonic flight regime. An example is the Catia software suite produced by Dassault. If your interest is in unsteady flow or aeroacoustic simulation, perhaps it's more like 5 to 10% error. The aeroacoustic simulation issues deal with mechanisms for predicted the cause of structural fatigue. Jet Screech Tone resolution stands as an example. Algorithms for solving the high angle of attack problem using the Navier-Stoke equations with various turbulence models also perform quite well. Bear in mind that experimentation also suffers in capturing the physics for some of these problems. Now some of the challenges remaining for CFD include simulation things like near wall turbulence and other subtleties like high-end LES. CFD also addresses turbulent combustion quite well.

As @ Douglas Vinson Nance said it is based on the application. However, if it is plus or minus 10% variation will be good.

Dear Majid Farahani

This is just my point of view:

Many factors can be involved in the difference between the numerical and experimental results. Cases of 2-D, unsteady, incomprehensible, etc have their own error. You should consider all the errors in your assumptions. Using scale analysis leads you to a desirable estimation of error in your simulation. Scaling is accomplished by assigning the order of magnitude values to dependent and independent variables in an equation. By estimating the order-of-magnitude for the quantities of interest, you can figure out how much the allowable error is.

Good luck.

Nice contribution Mojtaba Bezaatpour.

The nature of the application has a great impact on the variation value. I never come across a specific value. However, 10% or less would be reasonable.

In most cases, about a 5% error is acceptable in CFD calculations.

It depends on what accuracy you need. Please mention that the experimental results have also uncertainty. For instance, in turbulence convection heat transfer, the uncertainty of the correlation would be 25-30%. So, it depends on the problems and the accuracy of the solution we need.

Regards.

It is very hard to generalize a specific tolerance in this issue, for it highly depends on applications, plus numbers of variables are engaging in the simulation. Having said that, previous CFD data on a specific topic has a role to play. Moreover, something should be noted that these all must be considered after sensitivity analysis for grid size, element size, etc....

Hi Mr.Farahani

Actually it depends on your subject and the methodology that you have applied. For knowing this, you should check different validated publications who they have done your issue by your manner. In some problems the differences between numerical and experimental results is acceptable even upper 30%, because the trend of the curves in numerical work is same as experimental work. So you should check the references.....

I agree with the previous comments. I want to add that when possible, in CFD it is also good to resort to other type of error quantifiers such as the Root Mean Square Error (RMSE)

Majid Farahani

10-12 %

Round 10% it is acceptable

Dear Dr. Al-obaidi

I invite you to see the reference below how upper 40% has been acceptable, because in such problems the trend is important and the curves of experimental and numerical data has the same trend. Actually it directly depends on the problem and the manner that they have been applied.

Talukdar, P. K., Sardar, A., Kulkarni, V., & Saha, U. K. (2018). Parametric analysis of model Savonius hydrokinetic turbines through experimental and computational investigations. Energy Conversion and Management, 158, 36-49.‏

The variation between experimental and Computational Fluid Dynamics was depending on the

- application ( case study)

- applied the boundary and initial condition

- the CFD program of simulation ( the technique)

- method to solve the matrix with select the equation to solve ( explicit or explicit)

- fine of mesh grid generate for the case study

- number of iteration to arriving to converge solution

- weighting factors

We can arrived to the result the variation depending on how to build and organized your program

First you have to know the estimated errors of the simulation AND the experiment, if you are trying to validate some dependent variable, as a Nusselt number for instance. Both errors may add up.

Sometimes you are trying to validate a qualitative feature (a flow pattern for example). In these cases you need to verify that your simulation gives a physically realistic behaviour.

7%-10%

I think 10% is quite acceptable

First check the uncertainty in the primary and derived quantities of experimental values. You will have a clearly picture w.r.t it. Also in CFD carry grid sensitivity test before arriving to the final value. At last when u compare both, it should not be greater than 10%.

I hope this might help you.

Upto 10% variation is acceptable. Better you perform uncertainty analysis for experimental results for adding scientific value.

If two variants are compared using the same methodology and the same parameters, most often, it does not matter how much error there is in an individual result. The error is contained in both variants and has the same sign. The relative ratio of results remains approximately the same.

If an individual test is performed, it is necessary to distinguish the purpose of the test. It is not the same if it is fundamental scientific research or industrial research or something else.

I am highly sceptical of the idea that 10% or any other figures should be specified in this issue, for the cases are different intrinsically. So imagine if one can compare the simulation of a sphere in the water with a complex problems including FSI or multi fluids! Besides, how can one put the results of different methods such as turbulent models LES against K-epsilon for a specific simulation in a same benchmark?! Do you expect obtaining 10 percent tolerance for LES and K-epsilon for a same case, so what would be the point of developement of advance models then?! In my view, neither tolerance could be defined, nor try should be conducted intentionally toward experimental results, since both are fabricated. Instead, I recommend studying varoius models in your area ,and just try to run your model based upon normal practice.

Greetings

A grid independent study should be carried out with different turbulence models. This eliminates any grid effects on your CFD results. Also, by doing this, you are in a better position to understand which turbulence model is best suited for your application. Note that the percentage difference between CFD and experimental values should be low but there is no specific range one can say as it depends on your application.

On a side note, when performing grid independent study, you notice that 2 turbulence models gives close results to the experimental value, then you need to theoretically understand the turbulence models to find out which would be more suitable for your application.

To have a theoretical understanding of turbulence models, there is a youtube channel "Fluid Mechanics 101". Here, you can understand in detail of each turbulence model. I recommend everyone who wish to learn on turbulence models to refer to this channel.

(1)Spalart Allmaras model

https://www.youtube.com/watch?v=Xivc0EIGFQw

(2)K-Epsilon model

https://www.youtube.com/watch?v=fOB91zQ7HJU

(3)K-Omega model

https://www.youtube.com/watch?v=myv-ityFnS4

(4)The Transition SST model

https://www.youtube.com/watch?v=5htknS9uVEk

It should be borne in mind which simplifications and assumptions are applied in defining the physical model. The actual prototype (physical model) always deviates to some extent from the physical model.

As far as CFD is concerned, an appropriate turbulence model should definitely be chosen. And the independence of the mesh needs to be checked. Tolerances for iterations, etc. are also important.

It is also important to be aware that certain errors occur during the laboratory measurement (depending on the accuracy of the instruments, the measurement procedure, etc.). Care should be taken that the measurements are repeatable and reliable.

So the whole issue is complex.

However, there are certain guidelines for specific areas of the issue. Some are prescribed in certain standards (e.g. ASME).

The experimental measurements’ accuracy originated from two major sources: human error and equipment error. The equipment error (systematic error) is measurable, and its sources can be found. However, it is very difficult to find the source of human error, which is also known as random error. To prevent human error in the experiments, the measurements were repeated and their average was considered the final amount. Also, the numerical model owns some systematic errors.

It depends on the type of configuration (aircraft, missile, launch vehicle, etc) and Mach number. In general, the acceptable deviations are 2-5% at subsonic, supersonic and hypersonic Mach numbers and about 10% at transonic, These are mainly for steady flows. For unsteady flows, the comparisons can be anywhere from bad to terrible.

Suryanarayana Gargeshwari Is there any Paper that we can refer to justify the variation ??

There are several papers in which such variations are observed / reported. But there are no general statements made anywhere. As already posted, these variations are for steady flows. For unsteady flows, comparisons are generally not so good. It can be terrible for unsteady transonic flows, depending on the CFD methods used. Time-dependent DNS has shown good match for flow over laminar supercritical airfoils (Browse in google scholar forT K Sengupta, Computers and Fluids).

we can go up to 10 % for mean values, 12% for fluctuating values

Does anyone know a book or paper that recommends a value?

It depends on the application.

Reza Khazaee in multiphase flow

Validation of CFD w.r.t. experimental data, as eluded, depended on the application.

This question has no definitive answer. It could be 5, 10 or even 20%, depending on the application.

The question is what for the model is thought:

if you make a model and you validate it with experimental data most probably you intend to see how parameters variations could influence the result. So what is important is which is the uncertainty you can accept for the results with modified parameters. Will it be only a research of trends or will it be a research with values ?

This choice will determine the degree of precision you need for your model.

It is an error to fix a uncertainty value a priori.

Both experimental and CFD simulation are prone to error and uncertainty. Even the so called high-fidelity model just try to mimic the real physical problem.

For instance, the validation uncertainty is defined as the combination of the uncertainties in the experimental data and the portion of the uncertainties in the CFD prediction that can be estimated (Uncertainties and CFD code validation, 1997).

Nevertheless, some guides, not ‘’standards’’ for CFD results Verification & Validations (V&V) can be found in (Guide for the Verification and Validation of Computational Fluid Dynamics Simulations, AIAA G-077-1998).

Hope this will be useful.

It is definitely up to the problem. For example, in a thermal comfort study, the temperatures may vary around 20-40 C. And 0.5 C may change the sense of comfort. Thus 0.5 C is very important in such a study.

But if you're studying on a steel melting process around 1500 C, then 0.5 C is a very negligible temperature difference.

Either way,

It depends on different parameters and application.

The question has no definitive answer. It could be 5, 10 or even 20%, depending on the application.

For most applications, about 5% is acceptable.