Details
-
Type:
Story
-
Status: Done
-
Resolution: Done
-
Fix Version/s: None
-
Component/s: ip_diffim, pipe_tasks
-
Labels:
-
Story Points:6
-
Epic Link:
-
Sprint:AP F18-5
-
Team:Alert Production
Description
The prototype DcrCoadd generation code used the same filter transmission curve as the input simulations, and was able to recover the input spectra of individual simulated sources accurately. The new stackified DcrCoadds use a simple top hat transmission curve (until DM-13668 is completed), and are not able to recover the input simulated source spectra as well. This ticket is to investigate the effect of the filter transmission on the accuracy of the recovered source spectra.
Attachments
Issue Links
- blocks
-
- Done
Activity
| Field | Original Value | New Value |
|---|---|---|
| Epic Link |
|
| Risk Score | 0 |
| Sprint | AP F18-4 [ 749 ] | AP F18-5 [ 750 ] |
| Story Points | 4 | 6 |
| Status | To Do [ 10001 ] | In Progress [ 3 ] |
| Reviewers | Meredith Rawls [ mrawls ] | |
| Status | In Progress [ 3 ] | In Review [ 10004 ] |
This mostly looks OK, but I left some comments on GitHub I'd like to see addressed before I mark it Reviewed.
You shared a plot on Slack that related to this and illustrated how nicely the new filter midpoint worked, didn't you? I'd love to see before/after comparison plots on this ticket for posterity!
| Attachment | spectrum_N180_G5296.png [ 34904 ] | |
| Attachment | spectrum_N188_G5235.png [ 34905 ] | |
| Attachment | spectrum_N192_K4838.png [ 34906 ] | |
| Attachment | spectrum_N196_K4494.png [ 34907 ] | |
| Attachment | spectrum_N199_F6478.png [ 34908 ] | |
| Attachment | spectrum_N200_G5957.png [ 34909 ] | |
| Attachment | spectrum_N204_K4087.png [ 34910 ] |
I've attached several plots of example measured spectra with the different options added in this ticket turned on or off. The largest change was due to multiplying the measured spectrum by the LSST filter throughput in the final analysis. In the plots, this is seen by comparing "corrected base" to the original "uncorrected base." The filter throughput correction is included in all of the other measured values.
Two of the three modified options are plotted:
- "gain": which involves slowly increasing the gain as the number of iterations increases. This results in slightly larger changes between iterations (after the first few) and allows the model to converge faster.
- "split": which splits each subfilter in half during forward modeling and averages the result of DCR shifting between the two halves. This partially accounts for non-negligible DCR across a subfilter, without increasing the number of free parameters.
The best results are obtained by using both options together.
The "useModelWeights" option is not plotted, because it's results are indistinguishable from the base plots for three subfilters. It's benefit comes for higher numbers of subfilters where the model previously had trouble converging due to feedback between noisy pixels. With this new option turned on, the model converges smoothly with five subfilters.
Looks great! Please just fix the one mismatched reference I found (see GitHub). Thanks for all the plots.
| Status | In Review [ 10004 ] | Reviewed [ 10101 ] |
It looks like the "stackified" code is working as well or better than the prototype implementation now.
| Resolution | Done [ 10000 ] | |
| Status | Reviewed [ 10101 ] | Done [ 10002 ] |
Note that this includes a pull request in
{pipe_tasks}, which will probably not get a link from JIRA.
This includes three slightly unrelated commits that together improve the accuracy of building a DcrModel. This allows the model to smoothly converge even with greater numbers of subfilters, which mitigates the problem of residual DCR across a subfilter for high airmass observations.