Population solver dynamic ion range #187
Conversation
|
@lukeshingles, any ideas on why this is failing with a "terminate called after throwing an instance of 'std::bad_array_new_length'"? All I have done is add the nions_used and first_ion_used variables and set them equal to what was previously used in the get_element_nlte_dimension and get_max_nlte_dimension functions. Is there a way to run the tests locally so that I can get more detailed information to diagnose this? |
|
@fionntancallan the most common cause of that would be trying to resize the vector to a negative length. You can look in the ci.yml workflow and run those same commands locally. |
|
@lukeshingles in "solve_nlte_pops_element" when assigning memory to the rate matrices, e.g.: Why is max_nlte_dimension used for the resize? Since solve_nlte_pops_element is only called for one element at a time could you not just use nlte_dimension for the resize making max_nlte_dimension redundant? Or am I missing something? |
|
Is it because the expensive operation is when you have to resize to a matrix bigger than the memory allocated for the previous iteration of the matrix, so its generally better to allocate the max memory and then reduce what is actually utilised with the gsl_matrix_view_array? |
Yes, memory allocations and deallocations are slow and they cause memory fragmentation. It's better to allocate just once if it's possible to know ahead of time how much memory is required. I don't see why you would have to touch that code though, since you're only potentially reducing the size of the matrix, the max_nlte_dimension should not change. |
Ok thanks, the output of max nlte_dimension shouldn't change but it works by looping through all the elements and calling get_element_nlte_dimension for each element which now needs two extra arguments which can be applied if the number of ions have been reduced. This change in dimension shouldn't apply to the call of get_element_nlte_dimension from get_max_nlte_dimension provided the arguments passed are consistent with the entire ion being treated in NLTE |
|
@lukeshingles do you prefer this branch is brought up to date with the current develop branch with update with rebase or update with merge commit? |
Either one is good since it will look the same once it is merged anyway. |
…uppermost ion used is based on LTE phi factors
…whole element is treated in NLTE in loop through elements in get_max_nlte_dimension()
56dfb96 to
1edf903
Compare
…ment when nlte solver fails
This reverts commit a9cd5db.
|
@lukeshingles the new functionality seems to be working for single zone test simulations. What is the best way to proceed from here. Are there larger scale tests you run to make sure everything is behaving? The new functionality changes the results of the automated tests so its probably a good idea to do some more comprehensive testing before we would update the tests. |
Please add some description of your new options to artisoptions_doc.md. Currently, our two NLTE tests do not pass through the NLTE error condition and recovery steps. I suggest duplicating and modifying input model/time range for the nebular_1d_3dgrid test so that it gets into this kind of state and preserves the functionality of your new options. |
|
@lukeshingles I want to run some full simulations to test the new functionality in more detail and see if it changes anything in terms of the synthetic observables predicted. Are there specific runs on the ARTIS google drive which you generally rerun after an update to the code? |
I think we discussed re-running the W7 and a Shen2018 with your new options enabled to see if it affects the nebular spectra (hopefully not). Google Drive path: But you don't necessarily need to do that before we merge the PR. You can always follow up with another PR if you want to make further changes to your new experimental options. |
tests/nltephotospheric_dynamic_ion_range_1d_1dgrid_inputfiles/recombrates.txt
Outdated
Show resolved
Hide resolved
tests/nebular_dynamic_ion_range_1d_1dgrid_inputfiles/abundances.txt
Outdated
Show resolved
Hide resolved
Yes I can run those after we merge the PR,I was just wondering what the exact versions to compare to are since there are a few different versions of the W7 and Shen model in that folder. I guess its the subchdet_shen2018_1.00_5050_20211111_150_410d_ntfixnewphixsbfest_2e9pkt_virgo and w7_outercut_20211111_150_410d_ntfixnewphixsbfest_ntworkfn_2e9pkt_virgo runs I should be comparing to but wanted to double check |
Yes for the subchdet, but for w7 use w7_outercut_20210331_150_410d_2e9pkt_fixntfrac_juwels. The artis options have evolved a bit since there, but you should be able to simply take the latest artisoptions_nltenebular.h and edit the packet count. 2e9 packets is probably overkill, and 1e9 packets should look nearly identical anyway. So for 960 cores, set MPKTS to 1042000. |
e62f9f
90CE
c to
29092a0
Compare
Add functionality that allows the range of ions that are used in the NLTE solution to be dynamically changed if the NLTE solution fails due to either negative populations or large populati 8000 on inversions:
-Previously when the NLTE population was negative for a level it was directly replaced with the LTE level population. This resulted in cases where the negative excited population was replaced with a value much greater than the NLTE ground pop resulting in extremely large partition function values which in some cases then overflowed the float limit resulting in the simulation failing. Additionally, negative NLTE ground populations were just replaced with the LTE value. However, if the ground population is negative this means the entire solution for that ion is unlikely to be sensible.
-There were previously no checks on population inversions. Large population inversions can lead to partition function values which overflow the float limit causing the simulation to fail.
The functionality added in this pull request should solve issues related to the NLTE solver giving negative populations or large populations inversions for ions which have very low populations (e.g for neutral ion states when the dominant ionisation state is a highly ionised state and vice versa) by allowing the NLTE solution to be recalculated when the solver fails but with the individual ion that is causing the issues excluded from the solution for that element. Previously the NLTE solution could only be calculated including all ions included in the simulation for a given element and there was only functionality to set the populations of the entire element to LTE values of the NLTE solver failed.