Add OpenAeroStruct example

[FrancoisGallard]

Add an example that exposes the OpenAeroStruct https://github.com/mdolab/openaerostruct OpenMDAO group as Philote discipline.
The example could expose:

• One group that solves the MDA and computes the coupled derivatives
• Two groups, one for the aerodynamics, one for the structure, and the partial derivatives wrt coupling variables, states, and the design variables.

This would allow to compare the monolithic MDA, with the splitted disciplines, in another MDO framework (I will implement it with GEMSEO and publish the result), and compare the MDAs results and total derivatives.

[chrislupp]

Sure, that's no problem. If you have seen the philote examples in https://github.com/MDO-Standards/philote-examples, does that fit what you had in mind or something different.

That's only aerodynamics, but I could easily expand the example for this repo to be an aeroelastic one.

[FrancoisGallard]

Thanks,

Yes I've experimented with that example but I'm not experienced enough with OpenAeroStruct or OpenMDAO to do the required modifications myself.

[chrislupp]

sure, no problem. I can do that wrapping and will see if I can take a shot at it tonight.

[chrislupp]

I took a quick shot at the implementation. I put it under the feature/aerostruct branch of the examples repository: https://github.com/MDO-Standards/philote-examples/tree/feature/aerostruct

There are two sets of examples: one monolithic MDA and one split by disciplines. Below is a quick comparison between the monolithic and split results for two different mesh sizes.

	Monolithic	Split	Overhead
Small mesh (5x2)	0.253s, 11 iters	0.290s, 10 iters	15%
Large mesh (21x7)	0.336s, 10 iters	0.381s, 9 iters	13%

Is this in line with what you had in mind?

[chrislupp]

Here is a comparison to the python in memory OAS solution:

Case	Native (s)	Philote (s)	Overhead
VLM small	0.121	0.124	3%
VLM large	0.099	0.155	57%
Aerostruct small	0.100	0.211	112%
Aerostruct large	0.181	0.232	28%
Split small	0.146	0.282	94%
Split large	0.175	0.405	131%

I'm pretty sure that this could be optimized further, as this was a one-shot benchmark without tuning any performance-sensitive settings (like chunk size, etc.). But obviously, some overhead is expected. This also tracks pretty closely with some of our past findings, where changing the computational boundary can result in a pretty significant change in performance.

[FrancoisGallard]

I took a quick shot at the implementation. I put it under the feature/aerostruct branch of the examples repository: https://github.com/MDO-Standards/philote-examples/tree/feature/aerostruct

There are two sets of examples: one monolithic MDA and one split by disciplines. Below is a quick comparison between the monolithic and split results for two different mesh sizes.

Monolithic Split Overhead
Small mesh (5x2) 0.253s, 11 iters 0.290s, 10 iters 15%
Large mesh (21x7) 0.336s, 10 iters 0.381s, 9 iters 13%
Is this in line with what you had in mind?

Very interresting results ! Thanks a lot.

I tried to access the link https://github.com/MDO-Standards/philote-examples/tree/feature/aerostruct but github tells me
404 - page not found
Cannot find a valid ref in
feature/aerostruct

Best