[Emily1412]

According to the documentation Running SU2,"Objective functions (drag, lift, etc.) are determined running a direct flow solution in SU2_CFD and gradients are obtained using the adjoint solution. For each iteration in the design process, the mesh is deformed using SU2_MDC, and the analysis is repeated until a local optimum is reached."
I want to know that when will a local optimum be reached? What the exact meaning of it? It seems that it calls the SU2_CFD many times. How many DESIGNS/DSN_00x will be performed? When will the program terminate?
Hope for your reply!!

[copeland]

Hello Emily,

Your question is a good one -- and, in general, not one that has a single answer. The optimization in SU2 uses gradient based optimization techniques to evolve the initial design to the optimal design in the manner that's described in your quote from the website. A local optimum is determined using the KKT conditions (run a google search for this) -- but in principle it's a zero gradient and only small changes in the objective function value from the previous design. The tolerances for these conditions are set in the shape_optimization.py script. Feel free to open this python script and adjust them as you see fit.

The optimizer will call SU2 as many times as it needs until these KKT conditions are met. In general, there's no way to know how many times are required -- it depends on how well-behaved your problem is and how large your design space is.


Hope that helps,
Sean

[beatlejuice]

Hi,

our cluster has a 24 hour limitation for jobs. The usual way to deal with that is to write a script that stops the simulation before the 24 h end and qsub it again (restart from previous state).

Is there a proper and clean way to stop a shape optimization (shape_optimization.py) and to restart it?

For parallel_computation.py or continous_adjoint.py i make use of a runtime.dat to stop the simulation and do the restart in the usual way with the solution...dat files. This works fine. But unfortunately, this approach does not work for shape optimization procedure.

If anyone knows a way to handle this, i would appreciate it, if you could share this information =)

Thanks!

[hlk]

Originally Posted by beatlejuice

Hi,

our cluster has a 24 hour limitation for jobs. The usual way to deal with that is to write a script that stops the simulation before the 24 h end and qsub it again (restart from previous state).

Is there a proper and clean way to stop a shape optimization (shape_optimization.py) and to restart it?

For parallel_computation.py or continous_adjoint.py i make use of a runtime.dat to stop the simulation and do the restart in the usual way with the solution...dat files. This works fine. But unfortunately, this approach does not work for shape optimization procedure.

If anyone knows a way to handle this, i would appreciate it, if you could share this information =)

Thanks!

shape optimization can be restarted by appending ' -r project.pkl' to the command.
The project.pkl file should be created in the run directory during the shape optimization run.

[beatlejuice]

Thank you hlk,

restarting works like a charm ...

But how to stop the shape_optimization in a proper way? (at any state)

Thx in advance!

Sry for double post ...

[beatlejuice]

Thank you hlk,

restarting works like a charm ...

But how to stop the shape_optimization in a proper way? (at any state)

Thx in advance!

[hlk]

Originally Posted by beatlejuice

Thank you hlk,

restarting works like a charm ...

But how to stop the shape_optimization in a proper way? (at any state)

Thx in advance!

you can set a maximum number of iterations, or increase the tolerance such that the optimizer will exit sooner. Otherwise, you do need to Ctrl-C or otherwise force it to stop at the command line. SU2 does not have a built-in abort utility, so this won't produce as nice of output, it will just stop the current simulation.

Please see the config template provided with the code for those options, and you may also find the shape optimization tutorials helpful:
https://github.com/su2code/SU2/wiki/...nsonic-Airfoil
and
https://github.com/su2code/SU2/wiki/...f-a-Fixed-Wing

[beatlejuice]

I have one question left =)

Is there a way to restart the shape_optimization from a certain DSN (not the latest one)? I had a look into the project.pkl with the idea to modify it to achieve this behavior.... but the project.pkl file is to cryptic for me ...

Is there somebody who knows how to do this?

[hlk]

Originally Posted by beatlejuice

I have one question left =)

Is there a way to restart the shape_optimization from a certain DSN (not the latest one)? I had a look into the project.pkl with the idea to modify it to achieve this behavior.... but the project.pkl file is to cryptic for me ...

Is there somebody who knows how to do this?

The easiest way to do what you are talking about is to start a new optimization problem, using the desired design as the initial point. You will find the new mesh and solution files in the DSN subdirectory, you can copy these to a new directory and after some filename changes start the optimization just as you would otherwise.

[beatlejuice]

Thank you for your help, hlk!

[WillemFloris]

Hi there!

In the .cfg script there is an option "OPT_ACCURACY = XX", when the difference in objective function between each function call is less then value "XX" then the optimization will stop as well.

I am interested in an option which is similar to "OPT_ACCURACY". However I would like to apply this value to "NIT" instead of applying it to the FC. Does anyone knows how to do this?

Kind regards,

Floris van der Schuur