A single AVL Metamodel in case of multiple runs

[Divyam Verma]

To the SUAVE authors and users,

I have noticed that the utilizing AVL in our simulations is expensive as it trains and generates a metamodel to estimate CL, CDi (induced drag) and span eff. factor.

The above process becomes extremely expensive if I run, say 100, simulations (varying spans and root chord) in one go. That means that the AVL training will be done 100 times.

I wanted to know if there is a possibility to generate a single metamodel for AVL which will serve as a the metamodel for all the subsequent 100 runs? I was thinking of generating a metamodel based on aoa, mach, span and root chord and interfacing the metamodel generated with SUAVE.

What are your thoughts on this? Do you think its possible?

-Thanks

Divyam

[Emilio]

Hi,

Sure it's all possible. You have two options, one easier than the other.

You could add the variables of span and chord to the existing training. This is a one-off hack that is probably not reusable for other design problems. Then you have one metamodel.

The other is to build on top of SUAVE and put another metamodel on top. You can be stingy on the existing training to be a tighter range of points (and use only 1 config!). With that, you can then employ a sampling strategy, like a Latin Hypercube sample to build the outer metamodel.

Just thoughts.

Happy hacking!

-Emilio

[Divyam Verma]

Hi Emilio,

I am thinking of going with the second option of creating my own metamodel for AVL prior to finalizing the mission. Could you please give me an idea as to where in the AVL_Invicid.py code I have to make the changes to populate the CD and CL coming from AVL?

Thanks a ton!

-Divyam

[Emilio]

Hi,

Just to be sure, are you actually using SUAVE or just trying to extract AVL info to build the meta model? Just depends on where you want to hack in.

Also, are you sure you want to use root chord and span with CL and CD as the output?

-Emilio

[Divyam Verma]

Hi,

I am extracting AVL to build the metamodel (with span, root chord, aoa as a design variable) before the finalize() statement. And then, that metamodel will return a function (say fcap) for CL and CD estimation in the SUAVE's simulation runs.

I am doing this, so that the the training is only done once in the  beginning of an iterative loop.

Hope you understood my methodology.

-Divyam

[Emilio]

Hmm, I was wondering because every time the span and root chord change the reference area would change. Thus making CL and CD comparisons an apples to oranges situation. Just be sure you have some pieces of code to keep things consistent.

-Emilio