Survival Normalization #2673
Survival Normalization #2673
Conversation
Description of Boolean enabling Survival Normalization. Description briefly of functionality and current limitations.
Described Survival Normalization for phase-space sources. Unfortunately there is little description or justification for this out in the general topic literature. Though I have found it to work better in multiple test cases, notably the FOM in areas with middling statistics (~10%). In term should add capability of normalization for biased general sources.
Code commenting for actual normalization mechanism
Code commenting
Survival normalization pr 3
Remove duplicate description of Surv Norm
Remove duplicate description
Survival Normalization by Default Off
Update text on Survival Normalization, default=false
Better explaining when to use survival normalization
There was a problem hiding this comment.
Thanks @yrrepy for the PR! We do need to implement general source biasing relatively soon, so I would say we might want to rework this in that direction a little bit. Namely, I don't think it would be too difficult to make this work with general sources. My recommendations from a first round of comments are below:
- Add just the initial source weight to the particle class rather than two other variables
- multiply relevant weight_survive and weight_cutoff values by initial weight
- This will have to also be updated in our weight windows implementation
I would think this would make the implementation cleaner, make it ready for when we introduce biased sources, and have minimal performance drawbacks, but curious to hear other peoples opinions. @paulromano @pshriwise
Single variable for Survival Normalization
Use single variable wgt0, and multiply cutoffs here (on the fly), rather than having to multply in simulation.cpp. and then telling the physics modules to get this or that cutoff.
Assign Particle Start Weight
Remove lingering cuttoff variables in particle_data.h
Bracket typo in physics.cpp
Not entering final if statement (russian roulette for not survival normalization)
Commits to address review
multiplication by wgt0 if Survival Normalization on
Pedantic capitalization to match physics_mg.cpp
|
Hi Ethan, Indeed, I find your suggestion to simplify the mechanics of the normalization much cleaner. Having the variable wgt0 in the particle class is arguably much more useful than normalized cutoffs in the class. I left the if statement in physicsX.cpp for the case where a source is biased (or of varying weights) but the user does not want to normalize. But perhaps there is a way to have the code check instantly there when it multiplies by wgt0?, I'm not terribly familiar with the fancier functions in C++. Re: General Sources |
Reorganize and reorder Boolean flags to account for changes that have occurred in the interim with the recent addition of weight_window_checkpoint_surface and weight_window_checkpoint_collision. Thus Resolving a conflict in the present PR.
|
What should I focus on the keep this moving forward? Some tests? Python bindings? |
|
@eepeterson @pshriwise and I talked about this PR the other day. I'll let @eepeterson summarize our thoughts as we were all on the same page. |
There was a problem hiding this comment.
@yrrepy sorry for the delays in the review. I think we can accomplish what you want with minimal changes to the existing code and no new settings or runtime parameters aside from setting the initial particle weight as previously discussed. That way we can rely on the existing test suite to make sure we haven't broken anything and since we haven't introduced any new "features" shouldn't need to write additional tests.
| @@ -53,14 +54,19 @@ extern bool surf_source_write; //!< write surface source file? | |||
| extern bool surf_mcpl_write; //!< write surface mcpl file? | |||
| extern bool surf_source_read; //!< read surface source file? | |||
| extern bool survival_biasing; //!< use survival biasing? | |||
| extern bool survival_normalization;//!< use survival normalization? | |||
There was a problem hiding this comment.
| extern bool survival_normalization;//!< use survival normalization? |
| extern bool temperature_multipole; //!< use multipole data? | ||
| extern "C" bool trigger_on; //!< tally triggers enabled? | ||
| extern bool trigger_predict; //!< predict batches for triggers? | ||
| extern bool ufs_on; //!< uniform fission site method on? | ||
| extern bool urr_ptables_on; //!< use unresolved resonance prob. tables? | ||
| extern "C" bool weight_windows_on; //!< are weight windows are enabled? | ||
| extern bool write_all_tracks; //!< write track files for every particle? | ||
| extern bool write_initial_source; //!< write out initial source file? | ||
| extern bool weight_window_checkpoint_surface; //!< enable weight window check |
There was a problem hiding this comment.
I think these are from another PR - if you rebase on develop these should go away.
src/physics.cpp
Outdated
| // if survival normalization is applicable, use normalized weight cutoff and normalized weight survive | ||
| if ((settings::source_file || settings::surf_source_read)&&(settings::survival_normalization)) { | ||
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | ||
| russian_roulette(p, settings::weight_survive*p.wgt0()); | ||
| } | ||
| } else if (p.wgt() < settings::weight_cutoff) { | ||
| russian_roulette(p, settings::weight_survive); |
There was a problem hiding this comment.
| // if survival normalization is applicable, use normalized weight cutoff and normalized weight survive | |
| if ((settings::source_file || settings::surf_source_read)&&(settings::survival_normalization)) { | |
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | |
| russian_roulette(p, settings::weight_survive*p.wgt0()); | |
| } | |
| } else if (p.wgt() < settings::weight_cutoff) { | |
| russian_roulette(p, settings::weight_survive); | |
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | |
| russian_roulette(p, settings::weight_survive*p.wgt0()); |
I think we can apply this globally for all cases rather than needing additional settings and checks.
src/physics_mg.cpp
Outdated
| // if survival normalization is applicable, use normalized weight cutoff and normalized weight survive | ||
| if ((settings::source_file || settings::surf_source_read)&&(settings::survival_normalization)) { | ||
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | ||
| russian_roulette(p, settings::weight_survive*p.wgt0()); | ||
| } | ||
| } else if (p.wgt() < settings::weight_cutoff) { | ||
| russian_roulette(p, settings::weight_survive); |
There was a problem hiding this comment.
| // if survival normalization is applicable, use normalized weight cutoff and normalized weight survive | |
| if ((settings::source_file || settings::surf_source_read)&&(settings::survival_normalization)) { | |
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | |
| russian_roulette(p, settings::weight_survive*p.wgt0()); | |
| } | |
| } else if (p.wgt() < settings::weight_cutoff) { | |
| russian_roulette(p, settings::weight_survive); | |
| if (p.wgt() < settings::weight_cutoff*p.wgt0()) { | |
| russian_roulette(p, settings::weight_survive*p.wgt0()); |
same comment here as above.
There was a problem hiding this comment.
fine with removing SSR gating, but I do think it is best to be able to turn this feature off
src/settings.cpp
Outdated
There was a problem hiding this comment.
I think we can accomplish the point of this PR with no additional settings, so my recommendation would be to revert those everywhere (settings.cpp, settings.h, etc)
| p.current_work() = index_source; | ||
|
|
||
| // set identifier for particle | ||
| p.id() = simulation::work_index[mpi::rank] + index_source; | ||
|
|
||
| // set particle history start weight | ||
| p.wgt0(p.wgt()); |
There was a problem hiding this comment.
We should set the initial particle weight in the Particle::from_source method instead of here in the same fashion that wgt and wgt_last are set.
There was a problem hiding this comment.
I tried what was mentioned, but it didn't work:
p.wgt0(p.wgt());
The legacy one does:
p.wgt0() = p.wgt();
| .. _survival_normalization: | ||
|
|
||
| Survival Normalization | ||
| ---------------- | ||
|
|
||
| This parameter can be used to make the adjustable parameters of survival biasing, | ||
| :math:`w_c` and :math:`w_s`, relative to the starting weight of a source particle. | ||
|
|
||
| It is currently only implemented for phase-space sources (MCPL file sources | ||
| and HDF5 surface sources). For these, the normalization of the parameters | ||
| is done per source particle. That is, for each history, the parameters | ||
| :math:`w_c` and :math:`w_s`, are multiplied by the start weight of the current | ||
| history. | ||
|
|
||
| This normalization is recommended for problems where numerous source particles are | ||
| initialized below the survival biasing weight cutoff :math:`w_c`; to prevent them | ||
| from being immediately rouletted. This can notably be the case for biased sources. | ||
|
|
There was a problem hiding this comment.
| .. _survival_normalization: | |
| Survival Normalization | |
| ---------------- | |
| This parameter can be used to make the adjustable parameters of survival biasing, | |
| :math:`w_c` and :math:`w_s`, relative to the starting weight of a source particle. | |
| It is currently only implemented for phase-space sources (MCPL file sources | |
| and HDF5 surface sources). For these, the normalization of the parameters | |
| is done per source particle. That is, for each history, the parameters | |
| :math:`w_c` and :math:`w_s`, are multiplied by the start weight of the current | |
| history. | |
| This normalization is recommended for problems where numerous source particles are | |
| initialized below the survival biasing weight cutoff :math:`w_c`; to prevent them | |
| from being immediately rouletted. This can notably be the case for biased sources. |
We don't need a new section here, but in the existing survival biasing section it could be worth mentioning the the weight_cutoff and weight_survive used for russian roulette are the respective fractions of the initial particle weight.
|
Hi, I'm happy to remove the Boolean conditions that were limiting it to only phase-spaces, but I do think it would be best to keep the survival_normalization Boolean allowing it to be switched on and off (a-la MCNP where wc1,wc2 can be <0 or >0 determining whether to norm, but here with the Boolean) |
|
Bump on the above, and comments on the approach to take for a Boolean switch |
Co-authored-by: Ethan Peterson <[email protected]>
no boolean gating off survival normalization for SSR only Co-authored-by: Ethan Peterson <[email protected]>
no SSR gating
remover "relevant for only SSR"
|
Closed in favor of #3070 which is synced with latest OpenMC version |
Please close #2650 as I will be attempting to shepherd this through.
Apologies for the long read ahead.
Description
This PR creates the functionality of Survival Normalization, perhaps better described as Survival Biasing Source Normalization. It is meant to complement and enhance OpenMC Survival Biasing in certain situations. This functionality is used to make the weight parameters (weight and weight_avg (known in the C++ code as weight_cutoff and weight_survive) relative to the weight of the source.
Situations where this should be of benefit is when numerous important particles are present on the source at weights below that of the weight_cutoff, normalization prevents them from being rapidly rouletted. This should notably be the case for biased general sources (which OpenMC currently lacks natively).
It is currently only applicable to phase-space sources (.mcpl and .h5). Should/could later be expanded and enabled for general (biased) sources.
Background
In comparisons between MCNP and OpenMC with survival biasing and with phase-spaces I kept seeing significant differences in uncertainties. This was partly complicated by the tracklength heating being disabled in np. Eventually I tracked it down to how the two are respectively handling survival biasing.
Unfortunately my analysis erred and this feature will not improve my work, I do believe however that it is still a useful feature and will aid some. (In my case FOMs decrease by ~2%).
For phase-space sources, the weight parameters are made relative to the starting weight of each initialized history. That is, unique weight parameters per particle history.
This follows the practice in the grand-daddy of Monte Carlos
https://apps.dtic.mil/sti/tr/pdf/ADA231425.pdf
doc page 64/65, pdf page 70/71
code block 375
Here is the best reference and investigation into weight cutoffs (survival biasing parameters) that I have come across.
https://www.osti.gov/servlets/purl/5823239
doc page 87, pdf page 96
paragraphs + Table 2.2
Example
An example application of Survival Biasing with Survival Normalization can be found here:
https://github.com/yrrepy/OMC-SurvNorm_Example
The .xlsx in that repository describes the problem and contains the results.
Case 1 is an SSW/SSR where the user changed the weight parameters between the two calculations. Survival normalization improves the FOM by a minimum of +12% in the outer shell, the most difficult region to score in this case. FOMs elsewhere decrease.
Case 2&3 is a constructed MCPL phase-space source masquerading as a biased isotropic point source.
Survival normalization almost across the board significantly improves FOMs
Case 2&3 Xb use the Survival Normalization feature. Case 2&3 Xc uses a manual multiplication of the weight parameters by the minimum weight of the biased source. The in-code per history Survival Normalization produces better FOMs than with the singular adjustment.
It could well be that the current phase-space condition in the code could be removed and allow Survival Normalization to work for all sources (under the hypothesis that someone produced a biased library source). But whether the per history FOM improvement continued in general I am not sure and could well be not the case.
In term one these examples could be the basis for additional test(s).
Code Changes
We added a few bits of functionality into the settings.cpp, particle_data.h, simulation.cpp and connecting physics files. The code takes the xml Boolean and enables the functionality of survival normalization. The values inside the survival normalization are then stored inside new object variables inside particle class. These values are then used inside all the physics files for calculations.
Settings.cpp
Added survival_normalization and source_file Booleans
Added source_file boolean inside the source xml toggle
Added new xml toggle boolean to enable and disable survival_normalization
Particle_data.h
Added private variables inside the particle class
Added new object variables wgt_cutoff and wgt_survive
Added setters and getters for both new object variables
Simulation.cpp
Toggle to adjust weight cutoff and weight survive by multiplying the current weight
Added the survival normalization functionality inside the given if statements scenario from settings.cpp
Physics.cpp/physics_common.cpp/physics_mg.cpp
Add the object variable into the appropriate weight_cutoff and weight_survive given the if statement scenario
Fixes
This PR seeks to address in part:
https://openmc.discourse.group/t/tally-uncertainty-openmc-vs-mcnp/1872/11
Checklist
To-Do