[Zhen ZHANG]

Hi all,
I want to define a new design variable, which is the turbulence Prandtl number. There are several questions:
1, Is it possible to define such a DV, which is not related to the geometry? Maybe it is not needed to call the SU2_DOT(_AD) to get the sensitivity.
2, What is the basic process to define a DV and use the discrete adjoint solver to get the sensitivity? Are there any codes to learn from?
3, Is it possible to extend the DV to be the field distribution of the turbulence Pr number? For example, the point-wise distribution among all mesh points.

Thank you!
zhen

[pcg]

Hi,
1 - It is possible, we already do it for some, like AoA.
2 - For Prandtl number you would need some modifications like these https://github.com/su2code/SU2/issues/679
3 - Everything is possible but it would require significant refactoring to make the entire code aware of a locally defined Pr number.

[Zhen ZHANG]

Hi,
Thank you for the reply!
I am working on it.

Zhen

[Zhen ZHANG]

Hi,
I am following the modifications in https://github.com/su2code/SU2/issues/679 to define the turbulence Prandtl number as a design variable.
I have defined a variable member named Prandtl_Turb in CFVMFlowSolverBase class, in CNumerics class, and in CDiscAdjSolver class. I registered the CDiscAdjSolver.Prandtl_Turb in CDiscAdjSolver::RegisterVariables. The next step is to set the Prandtl_Turb in direct-solver and numerics classes. I wrote all the inline virtual Get and Set functions. Here are my questions:

1. In CDiscAdjSolver::RegisterVariables, I set the Prandtl_Turb to the direct solver just as the temperature and the pressure. Is it correct?
2. To set the Prandtl_Turb in numerics class, how to determine which solver to get and which numerics to set?
3. When setting the Prandtl_Turb, I used the Get function that returns by value and the Set function that receives by value. Is that correct? Does it maintain the recording process?

If my understanding is not correct, please also point it out.
Thanks in advance!
zhen

[pcg]

Hi,

Yes temperature and pressure are good examples.
Set it in every solver and every numerics that is not "nullptr".
Passing by value is not a problem, but you need to make sure that the target values (in solver and numerics) are not overwritten after you set them equal to the variable you registered.

[Zhen ZHANG]

Hi Pedro,

I modified the code as what was done in issue679 to define the turbulence Prandtl number as a design variable. I defined a data member in the discrete adjoint solver, registered it, and set it to numerics and the fluid model. However, after returning a small value, the solver returns zero in the iteration.

Could you please have a quick look at the code and give me some advice? Thank you a lot!

https://github.com/zhangzhen117/mySU...e801dfd6884a99

The SU2_CFD_AD returns

+--------------------------------------+
|  Inner_Iter|  rms[A_Rho]|    Sens_Prt|
+--------------------------------------+
|           0|  -11.792377| -2.1225e-19|
|           1|  -11.706523|  0.0000e+00|
|           2|  -11.623393|  0.0000e+00|
|           3|  -11.531036|  0.0000e+00|
|           4|  -11.440695|  0.0000e+00|
|           5|  -11.356140|  0.0000e+00|
|           6|  -11.278596|  0.0000e+00|
|           7|  -11.208591|  0.0000e+00|
|           8|  -11.146480|  0.0000e+00|
|           9|  -11.092506|  0.0000e+00|

[pcg]

What you have done is correct.
Make sure that you are reading the derivative value in the same order as you register the variable.
This is a pitfall we fixed recently.
The other common cause for 0 derivatives is that something is not captured in the SetDependencies, for example maybe something needs to be recomputed to capture the influence of the Prandtl number on other quantities.

We have developer meetings every Wednesday at 3pm UK https://meet.jit.si/SU2_DevMeeting
It is easier to talk about code problems there.

[Zhen ZHANG]

Hi,
Thank you for your reply! I will check the points you mentioned.

[Zhen ZHANG]

Hi,
I have another question when I check the code. What is numerics_simd, and what is the difference between numerics and numeircs_simd?
Thanks!

[pcg]

Oh right I forgot about those.
They are special versions of some numerical schemes that use vectorization (to be faster, and save some memory in the adjoint).
Currently they are available for the compressible solver, always used for JST and used for Roe if USE_VECTORIZATION=YES.
The solver has a member variable called "edgeNumerics" and you will have to set the Pr number in there as well.

[Zhen ZHANG]

Hi,
I set the Prt to the edgeNumerics, and it returns the right order of magnitude (~e3).

+--------------------------------------+
|  Inner_Iter|  rms[A_Rho]|    Sens_Prt|
+--------------------------------------+
|           0|  -11.046790| -3.8419e+03|
|           1|  -11.009328|  0.0000e+00|
|           2|  -10.980025|  0.0000e+00|
|           3|  -10.958730|  0.0000e+00|
|           4|  -10.945269|  0.0000e+00|
|           5|  -10.939468|  0.0000e+00|
|           6|  -10.941165|  0.0000e+00|
|           7|  -10.950207|  0.0000e+00|
|           8|  -10.966425|  0.0000e+00|
|           9|  -10.989582|  0.0000e+00|

However, it gets zero after the first step. I will go on checking it, and do you have any advice? I think the recording step has finished when the iteration starts, but why the derivative goes zero after one step of iteration?
Thanks!