Hello! I'm interested about
 

Hello!

I'm interested about real-life problems which are solved using CFD and evolutionary algorithms - what would be examples where such methods are used in present days?

Kārlis

Hi Karlis Two of such may b
 

Hi Karlis

Two of such may be the design of the lower fuselage of the YF23 A VTOL, also called RAPTOR:

The problem was the hot exaust intake by the turbines on vertical landing, with the ensuing loss of lift efficiency, and crash down.

The other was the design of the profile of an RV shape made by the guys at Von Karmann Institute.

In the latter case the used an evolution strategy similar to simulated annealing; the design variables were the distance from the generic point of the profile to the axis of revolution, in the first case, i do not know what was the strategy involved, but i guess that a general GA would be too time consuming, since each fitness evaluation would necessarily require a fine remeshing to capture probably minute geometry changes, and a full turbulent,viscous CFD analysis;

That's allright (!?) when you are optimizing with one parameter... but when yr design space grows in dimensionality, duhhhh :-(

...better forget it... unless you can afford a CRAY, or an IBM Rs 6000.... :-)

I think that a much efficient kind of attack would be designing a set of CFD simulations usind DOE methodology, Krieging or Latin Hypercube, and then build a best fit polynomial response function and optimize it using standard calculus to find your best set of optimization parameters in design space:

Variables are continuous, optimization fuctions are smooth, design space is convex,and has real continuity

That is not a job for a GA...

... at least in my opinion... :-)

Alex

Yes, computing requirements mi
 

Yes, computing requirements might be terrible.
But does that mean, there is no way to get unexpected, but good results? Does DOE or similar allow it?

Kārlis

DOE is more user biased: funct
 

DOE is more user biased: functions are C(n) continuous, since we want "easy" funtions to handle analytically; Bouns imposed on parameters restrict the optimization to be local...

...Thats the price that we pay for computing power economy...

I guess that GA's aren't a very wise choice, because, since most of the solutions are viable, and design space topology is dense, we are talking of [n1, n2]^n here... the GA would lose much of its computing power evaluating viable but non optimal individuals...

if a fitness evaluation is a full remeshing, and an NS solution... well :-(

GA's can really find unexpectedly good results, since, the optimization criteria evolve smoothly in design space (are real valued functions of real quantities), but can be highly non monotonic... - have several local extrema -

I think that the best here would be to invert the NS solution in an analytically tractable form, explicit that solution in geometric BC parameters, and optimize them using standard calculus...

That would be the dream of those dudes at General Dynamics, or Grumman...

If it could be done, i'd probably be designing the best Hyper-mega-high-tech super fighter aircraft... :-)

...but i am talking of "art" here... the "solutions" are highly dependent on intuitive knowledge, biased by expert designers, and rely heavily on a "history" of successful designs...

... there is no Brute Force computing attack that can compensate for a lack of that knowldge, and "lock" in the optimal solution in a global scope... at least not in a reasonable amount of time...

Alex

Hi Kārlis, I don`t know if
 

Hi Kārlis,

I don`t know if Tosca Fluid does what you mean by evolutionary algorithms? http://www.fe-design.de/fileadmin/software/TOSCA_Fluid/TOSCA_Fluid-DatasheetPRIN T.pdf (english) http://www.fe-design.de/en/tosca-fluid/tosca-fluid.html (german)

Tosca Fluid was startet by Daimler

elvis

TOSCA fluid does not use GAs.
 

TOSCA fluid does not use GAs. It uses field-line integration of Lagrangian particles to identify cells through which high resident time fluid passes - i.e. recirculation zones.

You can do the same thing in OpenFOAM (for a much lower cost) by simply taking the dot product of a potential flow solution with the normal RAS solution. All elements with positive dot products can then be retained using subsetMesh to produce pretty much the same results as those presented in the TOSCA fluid pdf.

Oh and we have used GAs to do
 

Oh and we have used GAs to do multi-parameter duct optimization using a brute-force remeshing approach. Works quite well. Take a look at this paper.

http://www.openfoam-conference.com/2..._IslamMoni.pdf