LBNL-2020: Galactic Foreground Modeling

From CMB-S4 wiki
Jump to navigationJump to search

Connection details

Parallels 5 and 6, Session D. Wednesday April 1, 11:30 AM to 1:30 PM pacific.

https://lbnl.zoom.us/j/659092332

Monitor: Ben

Scribe: Brandon

Charge

  • Presentation / discussion of new developments in simulation techniques, codes, since the last meeting.
  • Identify the other working groups with science cases relying on Galactic foreground modeling (low-ell / BB, extragalactic foregrounds, etc)
    • What are the conclusions of the previous data challenges for these groups, in what direction does the foreground modeling require more complexity.
  • Are we prepared to provide the simulations that will be required to answer outstanding science questions for CD1.
    • What trade off studies are being looked at, will these require new foreground modeling efforts?
    • Are existing simulations sufficient for the next data challenge?

Agenda

Break

Discussion

Notes

  • Data challenge paper and next steps (Victor, 10)

Read the paper! Still under review, comments from everyone encouraged
Iterative framework--Instrument model, mapmaking, analysis, parameter sensitivity
Uses achieved performance
Suite of foreground models--see table from Ben Racine
Some decorrelation models induce bias, but generally shows up in poor model fits
Study of 10 different (buildable) instrument configurations, found #5 to be best
Baseline survey redefined, time to go through loop again (DC06, next talk)
Analyzed also different survey strategies using new hitmaps
Also looked at different foregrounds masks, biggest effects for Chile
Allowed for unmodeled foreground residuals at 1% level--should be achievable
Propagated through to sigma(r)
To come: foreground variations vs sky coverage, new foreground models (esp. With ‘natural’ decorrelation, not added by hand), realistic delensing treatment BUT can’t do foreground sims for delensing
Continue pushing through loop, getting more finalized each time

Q: How was level of decorrelation chosen in analyzed model?
A: At boundary of what was allowed by Planck analysis

  • DC06 (Caterina, 10)

DC06: 3 foreground models, 3 sky masks, 100 realizations each (at NERSC)
Models selected on ability to represent small scales
Vansyngel model based on Planck maps superimposed on MHD turbulence-based dust and synchrotron emission
Need: Small scale sims, more non-Gaussian models, dust-sync correlation / TE / E/B asymmetry

Q: In amplitude modulated case, was there smoothing?
A: Yes, 10 degree FWHM

  • PySM3 (Andrea, 10)

All PySM2 models now available in PySM3 except for small discrepancies in HD2017 model
NSIDE 4096 for LAT, 512 for SAT; 8192 templates in the works
Extragalactic foregrounds from Websky
Performance improvements over PySM2, compiled with numba, more memory efficient
Straightforward interface for implementing new models and extensive documentation--want the community to contribute more models!

  • Filament-based model (Carlos, 15)

Motivated by need to generate small scale sims
Based on a model of filaments: dust emission from ellipsoidal filaments
A few characteristic parameters, like aspect ratio, degree of alignment with B-field
Able to recover observed properties of dust polarization; power spectrum slope, BB/EE, rTE
Extension to 3D: realization of distribution of filaments in a box (400pc on a side) with observer at the center
Size distribution of filaments constrained by power spectrum slope, Larson’s Law
Note: focus is small scale structure
Misalignment angle (filament vs B-field) → rTE
Filament aspect ratio → BB/EE
So can calibrate these parameters with data
Also investigating using Planck 353 GHz dust emission as a spatial template so maps look more like Galaxy
No fundamental limitation on ell, only computing power
To come: use of this model on foregrounds for delensing

  • Frequency decorrelation (Brandon, 15)


  • Neural Net Models, Dust (Ben, 15)

Want to compress foreground observations into a model
Use NN to reproduce dust maps: Train on Planck 545 GHz maps
Only one sky, have to subsample (2000 non-independent maps)
Realizations have consistent power spectra with input data (up to ell~1800)
Constrained realizations possible
Next steps: train on MHD sims (data too resolution limited), could even add MHD info at small scales to well-measured large scale data; look into synchrotron; map-based sampling

Q: ell~2000 blurring problem, do you understand it?
A: Has to do with how the loss function is defined in auto-encoders. Could maybe get around this with GANs, but harder to train

  • Neural Net Models, Synchrotron (Giuseppe, 15)

Goal: use CNN to inpaint T and P foreground maps
Training sets: 353 GHz dust (Planck), 2.3 GHz synchrotron (S-PASS), IQU for both. 3x3 deg patches (15,000)
Pixel counts, power spectra, and Minkowski functionals very well reproduced (Planck maps)
Deep Prior method performs worse than GANs in Minkowski functional comparison
Similar for S-PASS QU maps
Available Python package: PICASSO


  • Discussion

We managed to fill available time with talks