Represent fvSource by tensor basis

[zhangzhen117]

Hi,
I want to achieve the following:
fvSource = div(g_1T_1+g_2T_2+...), where g_i is a scalar field used as the coefficient, and T_i is a tensor field constructed by tensors S and R used as the tensor basis.
I have two candidates.

• One is to modify the regression model, where I set g_i as the output of the NN, then access S and R and calculate T_i, calculate fvSource, and return it as a vector field.
• The other is to use the regression model to calculate scalar fields g_i, then, in the DASolver, calculate T_i, and update fvSource.

My questions are:

1. I believe the first way is consistent with the adjoint method, because all calculations are in the regression model. Is it correct?
2. If I want to keep the regression model unchanged and use the second way, I think the forward simulation is fine. However, I am not sure whether the adjoint functionality works correctly or not. Do I need to modify anywhere else to enable a correct adjoint simulation?
   Thank you!

[friedenhe]

If I understand you correctly, I believe the best way to implement this is to keep the regression model and DASolver unchanged, but add a child class to DAFvSource. In this child class, you need to implement how to use g_i and T_i to compute fvSource (check other child classes for example). Then, in the optimization, you can use this newly added childClass in runScript.py->daOptions->fvSource, and directly use g_i as the design variable (field inversion; set this g_i in inputInfo). Alternatively, you can call the regressionModel in runScript.py->daOptions->regressionModel to compute g_i and then use the regressionModel's parameters as the design variables.

The adjoint should be automatically done, even if your change is outside of the regression model. DAFoam uses automatic differentiation to facilitate the adjoint computation, so as long as your forward simulation works, the adjoint is automatically done. One thing you need to do is to create g_i as a member variable in the DAFvSource child class and make sure to register it into the mesh database, such that the regressionModel can access g_i from mesh.thisDb() later.

[HerryJhonson]

self_fvsource.zip @HerryJhonson Hi, I attached my code here for your reference.

Thank you so much for the code, I detailed contrast the DAFvSourceNonlinearStress, my code with you on the results are consistent.
I'll try to directly compile your code to run it. The encoding method is to incorporate the one in your attachment. Copy the DAFvSourceTBNN.C and DAFvSourceTBNN.H files to src/adjoint/DAFvSource/, add them to src/adjoint/Make/files, and then run the following code in sequence

• source $HOME/dafoam/loadDAFoam.sh
• export DAFOAM_NO_WARNINGS=1
• ./Allmake
• source $HOME/dafoam/OpenFOAM/OpenFOAM-v1812-ADR/etc/bashrc
• ./Allmake
• source $HOME/dafoam/loadDAFoam.sh

Then add it in runscript

"fvSource"
{
"TBNN"
{
"type": "TBNN",
},
}

Then add

"inputInfo": {
"g2_TBNN": {
"type": "field",
"fieldName": "g2",
"fieldType": "scalar",
"distributed": False,
"components": ["solver", "function"],
},

The operation result will display the following error:

--> FOAM FATAL ERROR:

request for volScalarField g2 from objectRegistry region0 failed
available objects of type volScalarField are

9
(
nut
yWall
nuTilda
nu
alphaPorosity
pRes
p
nuTildaRes
betaFINuTilda
)

I would like to confirm if you also declared g2 in the DASpalartAllmaras file? Or is there something wrong with the way I compiled and ran it above?

[zhangzhen117]

It seems correct. I did not declare g2 anywhere else.

[HerryJhonson]

Thank you and the developer's reply, I compiled from source and change the same file, the corresponding DAFvSourceNonlinearStress compilation is implemented, when I call in the runscript nonlinearStress can generate correct I need variables. The same operation doesn't work in the docker version of the software, which is the reason why I haven't succeeded before.

But then I made an appointment with other questions:

1. In your DAFvSourceTBNN, g2 is defined as NO_READ. Does this mean that the value of g2 is not updated by reading the file during the optimization process? Then, how is this parameter field updated and the corresponding TauRe generated? I tried to set the parameters to MUST_READ, but that doesn't work, the initial value I wrote in the document will not be read. It seems that we shouldn't read files in DAFvSource?

2. It seems that multiple FvSources are not supported in runscript. I want to set a constant gradP and a nonlinearStress, and write it in the following form:

"fvSource": {
  "nonlinearStress": {
    "type": "nonlinearStress",
  },
  "gradP": {
    "type": "uniformPressureGradient",
    "value": dp0,
    "direction": [1.0, 0.0, 0.0]
  },
},

But in fact, only the first fvSource will take effect. Whether it takes effect after testing has nothing to do with the type of fvSource, but only with the sequence.
May I ask if you have ever encountered such a problem?

[zhangzhen117]

Hi, glad to hear it works.
1 NO_READ/MUST_READ are ways that OpenFoam uses to initialize fields. If you want to update any fields, like g2, you should set it as the design variable and adjoint input. Then DAFoam will update it for you.
2 I am not sure. You can check the code to see whether the second fvSource is supposed to be read or not.

[zhangzhen117]

#392 may be related to your second question.

[friedenhe]

@HerryJhonson No, you. can't add two fvSources in DAFoam.