Transferring data from StatisticsReporter from one sub app to another

[amasri-22]

Hello,

Description of Problem

I am currently running a MOOSE Parallel Subset Simulation simulation which has two sub apps (an MC step and a basic diffusion simulation). The purpose of this simulation is to essentially optimize two quantities of interest.

The quantity of interest is the average concentration over the entire domain.
The boundary conditions of the diffusion problem are controlled by a normal distribution, whose mean values are unknown.
To calibrate these mean values correctly, a Monte-Carlo simulation is run on the physics simulation, whereby the boundary conditions are randomly sampled, applied to the physics problem, and then the final average concentration is obtained. After all MC steps are performed, we take the mean of the final average concentration and compare this value to a benchmark. Depending on how far away this value is from the benchmark, the mean values of the boundary condition distributions will be modified to try and minimize this discrepancy. This is where the parallel subset simulation enters the picture -- and also where the trouble begins.

Issue

The parallel subset simulation is trying to maximize the objective of interest -- averaged across the entire distribution. This objective is calculated in a post processor in the physics simulation, the objective is then average in the MC file using StatisticsReporter. Yet, when I try to transfer the data from the statistics reporter in the PSS simulation file, I always end up with the problem that the data cannot be found. I end up with the following error:

The following occurred in the Transfer 'reporter_transfer2' of type SamplerReporterTransfer. Reporter with the name "stats/storage_data:objective_MEAN" within app "sub0" was not found.

Questions

As a result I have the following questions:

1. Is this problem even possible with MultiApps and the modules of interest -- can they be run in serial, or must they be in parallel?
2. Is it possible transfer values from StatisticsReporter into a different subapp? If so, how?

Thank you!

physics.txt
master_PSS_optimizer.txt
Monte_Carlo_driver.txt
MC_driver_out.json

[GiudGiud]

Hello

Serial is fine, you can set the number of processors (either for the full sim or for a multiapp) to 1 to work in serial.
Parallel should work for PSS

Is it possible transfer values from StatisticsReporter into a different subapp? If so, how?

yes. The SamplerTransfer needs the correct name for the from_reporter though. The from_reporter is in the from_multiapp.
Use

[Debug]
  show_reporters = true
[]

to get the correct name.

There is an example (not merged) of using PSS here if you want
https://github.com/idaholab/virtual_test_bed/pull/669/files#diff-73e5f22d0f0c6d408a86dc77916d6a4e3688a9d6890437ef285683e28c639b1a
see Baseline_Input.i for the multiapp and pss.i for the parent input

[amasri-22]

Thank you for the speedy response!

While the advice you gave was certainly useful (thank you!), it seems to me that I am still having trouble finding the StatisticsReporter path that I can use for transfer. When utilizing your debug block, I only get the following reporters outputted

Reporter "constant_final/reporter_transfer:converged":
    Type:
      std::vector<bool>
    Producer:
      StochasticReporter "constant_final"
    Context type:
      StochasticReporterContext<bool>
    Consumer(s):
      None
  
  Reporter "constant_final/reporter_transfer:objective:value":
    Type:
      std::vector<double>
    Producer:
      StochasticReporter "constant_final"
    Context type:
      StochasticReporterContext<double>
    Consumer(s):
      None
  
  VectorPostprocessor "storage/data:objective":
    Type:
      std::vector<double>
    Producer:
      StochasticResults "storage"
    Context type:
      VectorPostprocessorContext<std::vector<double>>
    Consumer(s):
      None

This seems like it does not contain the information I am after. I am specifically looking for the mean value which is calculated using StatisticsReporter.
What I would like to do is at the end of the physics simulation, I take the mean value of the objective function at the final timestep, using StatisticsReporter. This would then be transferred to the PSS simulation set. When looking at the output from the debug block I added, for some reason I do not see the statistics reporter listed -- which is odd because the statistics are calculated in the resulting JSON.

Additionally, to see if there were any potential work arounds, in the PSS file itself, I decided to just transfer the raw objective value data into the PSS simulation (from the MC reporter). When doing this, and then running the simulation, for some reason I get the following error after the first PSS round begins to run (i.e., MC step solved all of the physics the specified number of times).

`*** ERROR ***
The following Reporter(s) were not declared:

Reporter "constant/reporter_transfer2:data:objective":
Type:
std::vector
Producer:
None
Consumer(s):
AdaptiveMonteCarloDecision "adaptive_MC" (mode: DISTRIBUTED)
ParallelSubsetSimulation "sample2" (mode: UNSET)
`

I feel as though this error may be related to my misunderstanding of the reporter system. It should be noted though that this was obtained through playing around with the execution scheduling in various blocks. Without doing so, I would get this error immediately.

Nonetheless, thank you for showing me the debug block, and thank you for that additional PSS example. Those are both very useful for me and have already helped me!

[GiudGiud]

The declaration in the statisticsReporter is happening in initialize() which may be a little later than in most reporters, possibly too late for Debug/show_reporters.

You can use the ReporterDebugOutput instead with the execute_on parameter set to both initial, timestep_begin to see the names
https://mooseframework.inl.gov/source/outputs/ReporterDebugOutput.html

Otherwsie you can look at the code:
It sets the names as

  for (const auto & item : _compute_stats)
  {
    const std::string s_name =
        r_name.getObjectName() + "_" + r_name.getValueName() + "_" + item.name();

so ends with "mean" not MEAN. And see the separators as "_"

[amasri-22]

This pretty much solved my problem -- thank you very much!

I should note that when I put in the correct path, I got the following error

Transfers/reporter_transfer2/from_reporter: Reporter value stats/storage_data:objective_MEAN is of unsupported type std::pair<double, std::vector<double>>. Contact MOOSE developers on how to transfer this type of reporter value.
This was resolved by modifying the StatisticsReporter to just output the calculation as just a vector, and not a pair additionally containing the statistics type. This is probably not the ideal fix, but, so far, it seems to be working.

[GiudGiud]

oh we should make an issue for this. Should be implemented
Can you please share the updated input?

[amasri-22]

Here are the files of interest (all master application, sub application, sub sub application, along with the modified StatisticsReporter class. I want to point out that this modification was made this morning, and so I am not sure if the change is correct. I have yet to do check it by hand. Also the PSS file isn't working as intended, but at least some sort of value seems to be transferred, and the MC file's JSON output shows that the mean is being stored as only a value and not a pair.

StatisticsReporter_C.txt
StatisticsReporter_h.txt
Monte_Carlo_driver.txt
physics.txt
master_PSS_optimizer.txt