Comment by Ned Molter on JIRA:

Adding some known issues here that should be looked into. Strikethrough means completed; see also the more recent comment about items requiring input from instrument teams.

• -Current implementation of multiprocessing does not work.-  works now but see other comment
• -Multiprocessing may be more effective looping over segmentation map sources, instead of individual pixels within each segmentation map source.-  resolution: did not end up making sense to move this
• -Each segmentation map source is currently modeled twice, once to create simulated grism map and again to subtract each source from that map to create the contamination fields. These are not very large and could easily be stored in memory to avoid re-computing.- fixed.
• -It may be possible to ignore sources near the edge of the direct image before modeling them, instead of modeling them and then throwing out results that lie outside the FOV.  This would modestly improve runtime.- probably more annoying than it's worth; should focus elsewhere to speed up runtime
• -It should be possible to run the step using only the brightest N sources in the segmentation map.  This would make testing the step easier, because extract_2d selects for the brightest N sources.-  done. Many of the fainter sources identified in the direct image lie far below the noise level in the grism map, so this could even be investigated as a method to reduce runtime, depending on how big a difference including these faint contamination sources makes for the science case of interest.

Comment by Ned Molter on JIRA:

I am just now submitting a draft PR for this.  There are still plenty of outstanding issues to resolve, but most of these would benefit from input from other people at this stage.

Changes:

• Made multiprocessing work
• Removed necessity to simulate Order 1 of all sources twice, instead storing these in memory.  This decreases runtime.
• Added a new optional output file, _simulated_slits.fits, that stores a MultiSlitModel containing contamination-free modeled slits for all requested spectral orders and sources.  I found this very useful for examining the step's outputs, but a decision should be made about whether this functionality is desired in the final version of the step.
• Added brightest_n parameter to specify how many sources to simulate.  This facilitates testing by decreasing runtime.
• Added unit testing suite with coverage for most helper functions.
• Did some refactoring to enhance readability and ease of unit testing.

Outstanding issues:

• Output needs to be examined by the instrument teams for additional outstanding issues.
• Zeroth order contamination for NIRISS needs to be calibrated.
• Decide whether the new output file is actually desired.
• Loading sources from cache used to be supported, but I have not turned this mode on, nor tested it.  A decision should be made about whether this functionality is desired.  If so, some updates are required to support this, and if not, all related functions need to be removed.
• Using multiple direct images in different filters to give basic spectral information to model spectra is supported in principle, but currently untested. If this functionality is desired, a test dataset where this mode is being used would be extremely helpful.
• One regression test each should be written for NIRCam and NIRSS, ideally using a dataset and result that has been vetted by the instrument teams.

Potential longer-term improvements that I don't think should go in this PR:

• It may be possible to ignore sources near the edge of the direct image before modeling them, instead of modeling them and then throwing out results that lie outside the FOV.  This would modestly improve runtime.
• Support for inputting assumed/previously-known source spectra to improve grism model.

I am curious to hear what else y'all can think of that should still be done.

Comment by Jo Taylor on JIRA:

Thanks for the progress Ned! We have a NIRISS WFSS meeting in a couple of weeks and will discuss the changes here at that time.