twillis449
commented
8 years ago I don't have any advice, but I'm glad to see you're working on this sort of stuff. I need a training session from you next week, :), to get up to speed on stefcal / modelling / solving / applying corrections etc for my baseline dependent averaging etc. I'm super rusty. Just got the LIONS ionosphere modeling to generate what looks like reasonable ionosphere jitter over the SKA-mid array. BTW - bring your woollies - temperatures are gonna be about 2 to 3 C here next week
IanHeywood
o-smirnov
Scenario:
Processing a VLA (B config) Measurement Set that has the selfcal hard-applied. Intention is to tidy up the junk around some off-centre sources.
Approach:
Image the MS, make a sky model, apply "dE" tags to problem sources. Use calico-stefcal with "solve for G" disabled, and "solve for dE" enabled with the source subset set to "=dE". Rationale here is that the usual
Vpq = Gp [ dEp X dEq* ] Gq*
RIME will be applied, Gs will be fixed to unity, dE will be fixed to unity for sources in the model without dE tags, and the dE solutions derived for the tagged sources will tighten up the image.
The problem:
The problem illustrated below persists for processing multiple SPWs either with independent per SPW monchromatic sky models, or with models with spectral index info derived from MT-MFS images formed from 256 MHz chunks. I've tried a few approaches in terms of time and frequency tiling, including single solutions for the whole time range (about an hour and a bit) as well as smoothing.
I've also tried it with and without multithreading.
The example here is for a single 64 MHz SPW, frequency tile = 0 and time smoothing kernel is 180 time slots (which for the 9s integration time is 27 minutes).
Anyway, here's the image for the MS made prior to it being loaded into MeqTrees:
And here's the image formed from the corrected residuals:
There's a deep negative at the position of every source in the model. Using turbo-sim.py to predict a set of model visibilities and then subtracting them without solving for anything leaves the sort of residuals you'd expect (a few faint unmodeled sources and some junk around the dE candidates), so the model is good. Dividing the model image by the residual image leaves values very close to 2.5 at the position of each source, so this might be the factor we're looking for.
Here's a plot of the solution amplitudes (from a no-smoothing run, a single solution per antenna for the entire data set):
If you take a quick by-eye average of these amplitudes and take the square of the reciprocal (divisional corrections and the product of the two antenna gains) then you get about 2.5. But: I don't get why these are being applied to everything in the model and not just the tagged source when forming the corrected model.
Using calico-wsrt-tens.py gives the behaviour I'd expect, the downside there being it's way slower.
Measurement Sets, LSMs, tdlconf.profiles available on request. Cheers.