Jump to content


- - - - -

Dakota Optimization after Sobol Sequence


  • Please log in to reply
12 replies to this topic

#1 Mr. Suraj Pawar

Mr. Suraj Pawar

    Advanced Member

  • Members
  • PipPipPip
  • 46 posts

Posted 25 June 2018 - 05:35 PM

I am working on the optimization of ducted propulsor for underwater vehicle. I have parametrized the geometry of the propeller and stator. Then I generated and exported number of geometries using Sobol sequence. Then I ran CFD simulation for all these geometries externally and extracted the required results. I do not have the complete optimization framework setup in CAESES. 

 

I would like to know

1. If I can use Dakota optimization toolkit for performing sensitivity analysis or response surface generation. I have an excel file which contains the design variable (generated by Sobol) and the results of CFD simulation.

 

Thank You. 


  • 0

#2 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 27 June 2018 - 05:09 AM

Hi Suraj,

 

Yes, you can use the "import result pool" functionality for this purpose import_result_pool.png . Once you have imported the data import_result_pool_dialog.png , it will be considered as a run import_result_pool_finished.png which then can be used as result pool for the response surface generation.


  • 0

#3 Mr. Suraj Pawar

Mr. Suraj Pawar

    Advanced Member

  • Members
  • PipPipPip
  • 46 posts

Posted 27 June 2018 - 08:55 PM

Thank you very much. I followed your instruction and uploaded the data for Sobol results. I tried to use Dakota sensitivity analysis. I have given two design variables as an input and one variable as an evaluation. When I run the Dakota optimization, the evaluation variable is always calculated to be zero. I do not have any constrain for Dakota. I might not be setting up the Dakota optimization correctly. Can you please look into this and give your feedback. 

I have also attached the csv file which contains the results of Sobol run. 

 

Thank You very much. 

Attached Files


  • 0

#4 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 28 June 2018 - 07:12 AM

Hi Suraj,

 

Now I understand. Optimizing on an imported data set using only the response surface is not readily available in the current release of CAESES. Sure, you can use the imported data to build up the initial response surface, to make use of it in the dakota design engine. However, the response surface algorithm requires an evaluation for the optimal design (e.g. using the CFD analysis). This evaluated design is then added to the response surface to further improve it. In your setup, there is no CFD etc linked to the objective so it simply returns zero all the time (that's the constant value of the parameter).

 

If you want to only find the best design on the response surface, we can use some other mechanisms. We have also done this but I have to double-check with colleagues whether we can hand it out.

 

Cheers

Joerg


  • 0

#5 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 28 June 2018 - 07:55 AM

Actually, when looking at your first post: Yes, if we had the complete optimization setup in CAESES, it would work without problems. The situation you have is a bit special (not super special, but something we still have to fully support).


  • 0

#6 Mr. Suraj Pawar

Mr. Suraj Pawar

    Advanced Member

  • Members
  • PipPipPip
  • 46 posts

Posted 28 June 2018 - 01:15 PM

Do I need to have the CFD setup in Dakota optimization framework even for parameter sensetivity analysis?

 

Thank You. 


  • 0

#7 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 28 June 2018 - 01:35 PM

Hi,

 

If you switch on the charts, you can also check the correlations between the different input variables and the output parameters. parameter_correlations_sensitivity_analysis.png This also helps to detect the most relevant parameters.


  • 0

#8 Mr. Alexandros Priftis

Mr. Alexandros Priftis

    Member

  • Members
  • PipPip
  • 11 posts

Posted 10 July 2018 - 03:40 PM

Hi Suraj,

 

Now I understand. Optimizing on an imported data set using only the response surface is not readily available in the current release of CAESES. Sure, you can use the imported data to build up the initial response surface, to make use of it in the dakota design engine. However, the response surface algorithm requires an evaluation for the optimal design (e.g. using the CFD analysis). This evaluated design is then added to the response surface to further improve it. In your setup, there is no CFD etc linked to the objective so it simply returns zero all the time (that's the constant value of the parameter).

 

If you want to only find the best design on the response surface, we can use some other mechanisms. We have also done this but I have to double-check with colleagues whether we can hand it out.

 

Cheers

Joerg

 

Hi Joerg,

 

sorry for getting involved in this discussion but I have a question related to what you wrote.

 

From the above description, I understand that when MOGA w/ RSM is utilised in CAESES/Dakota connection, an initial result pool is used to define the RS at first, but during the optimisation, the selected optimal designs from the Pareto front must be evaluated using the higher-fidelity method (e.g. CFD) also used to create the initial result pool. Does this happen every time a new generation is created in the optimisation process? Because if that's the case, the CAESES project used for the formal optimisation would have to be set up appropriately to be connected to CFD software and/or High Performance Computer so that the external s/w connection is run for the evaluation of the optimal designs?

 

Cheers,

 

Alex


  • 0

#9 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 11 July 2018 - 11:48 AM

Hi Alex,

 

No problem! Within each single iteration there is an internal MOGA optimization. The Pareto designs then need to be evaluated using high fidelity methods (CFD), right. And yes, this assumes a working software connection. It is like this because we decided that we want this kind of use case, it is frequently used in optimization projects.

 

Like I said before, we are also working on other use cases (e.g. without software connectors). This is not a big deal but still needs to be provided in an easy-to-use way. There are some things available already (beta).

 

Actually, you could also try out to run 1 iteration (even without having a valid software connector with CFD). Then the optimal design has invalid results, but you can look up the optimal design (design variables / results) on the response surface in the log files of the dakota run (look into the design engine folder). No ideal solution, but rather a test.

 

Does this help?


  • 0

#10 Mr. Alexandros Priftis

Mr. Alexandros Priftis

    Member

  • Members
  • PipPip
  • 11 posts

Posted 11 July 2018 - 03:25 PM

Yep, it's clearer now! :)

 

I would still be interested in keeping the response surface as is though (i.e. without re-evaluating it). Are the beta things you mentioned related to this post (https://www.caeses.c...4-easy-steps/)?

 

Thanks!


  • 0

#11 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 12 July 2018 - 08:00 AM

Beta stuff: yes, that's one of our solutions for this use case! Good research! :-)


  • 0

#12 Mr. Suraj Pawar

Mr. Suraj Pawar

    Advanced Member

  • Members
  • PipPipPip
  • 46 posts

Posted 12 July 2018 - 06:31 PM

So the RSM feature is not there in current version of CAESES? The RSM feature mentioned in this link (https://www.caeses.c...n-4-easy-steps/)

 

Thank You. 


  • 0

#13 Jörg

Jörg

    Moderator

  • Moderators
  • 503 posts
  • LocationBerlin, Germany

Posted 13 July 2018 - 10:37 AM

No it is not included, we can share something on request


  • 0