42
Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Embed Size (px)

Citation preview

Page 1: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Ice model update

Dmitry Chirkin, UW MadisonIceCube Collaboration meeting,Calibration session, March 2014

Page 2: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

1. use oversize factor of 5 rather than 16

2. fit to the azimuthal pattern of emitted light

3. updated dust+EDML merged log and best tilt maps

4. fit to 7 strings of data (up from 1)

5. fit includes DOMs on the same string (except immediate neighbors)

6. no regularization

7. equal weight of contribution to llh sum from all flashers

8. Use 10x simulation statistics (run on gpu nodes of npx4)

Updates since SPICE Lea

Page 3: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Running on npx4 clusterRun on a cluster of 16 computers with 96 GPUs

•Condor loads jobs as nodes become available•Each job fetches new run parameters and runs for 10+ hours

lockfile is not reliable on nfs, decision on which result to keep is made early, to avoid bias (if job runs of several nodes)

•Master script runs on npx4: creates parameter sets collects results monitors node status (kills stalled jobs)

Page 4: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

New initial approximation (Munich)

2d tilt maps were tried,but extrapolation behaviorwas too enthusiastic.

1d tilt map remains our best choice, but tilt points updatedfrom 2, 10, 52, 21, 66, 50to 14, 2, 52, 86, 66, 50

Constructed at the location of string 86(changed from 0,0) from the average dust log.

EDML log was matched to the averagedust log map and used for extrapolation.

Updated with new logs and age vs. depth parameterization at hole 86, scaled to SPICE Lea.

Page 5: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Dust logger vs. EDML log: new

Former:linear (vs. depth) correlation between the two logs

New:hyperbolic correlation between two log(logs)

Page 6: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

New first guess

Allow slope other than 1

Two independent fits for scattering and absorption

Page 7: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Correlation to SPICE model

OLD NEW

Page 8: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Dust logger/EDML log matchingupdated dust+EDML merged log and best tilt mapsimproved dust log vs. EDML correlation (hyperbolic log-log)correlation with SPICE model: 23% 14%, thus better extrapolation

OLD NEW

Page 9: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Unfolding of flasher LEDs

• Simulate LED light (2d gaussian) every 5 degrees in azimuthal direction from 0 to 355 degrees with a specified total number of photons

• Create a [azimuth x charge_in_DOM] matrix, and unfold to charge_in_DOM in data

• The unfolded pattern is re-simulated and llh is calculated

Page 10: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

SPICEFlasher LED unfolding:

fit to the azimuthal pattern of emitted lightincluding up/down-scattered componentswith optimized/fitted LED angular emission profile

Page 11: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Unfolding/Likelihood improvementsInvestigated multiple unfolding/t0 calculation strategies:

•Unfolding using total charge per DOM first, then fit for t0•For each t0 unfold time-binned data

large fluctuations are possible fluctuations non-existent if using chi2 at this step

Perform 2-step unfolding• Integrated charge• Time-binned data

add back zero components with small weight

Perform 5 likelihood maximizations in sequence, including:• NMML• PCG (using gsl, with variable substitution, both FR and PR)• Preconditioned BFGS2 and SD At each step result accepted only if llh improves

Page 12: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Improved likelihood minimizer

Page 13: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

SPICE

Improved minimization strategy

old initial approximation llh=600 new llh=554 (new correlation)

compare to SPICE Lea llh=570 llh=548 (with improved llh algorithm)

new best: ~ 510 (sum for 410 flasher configurations)

model error: AHA 42% / WHAM! 32% Mie 25% / Lea 18% <15%

smaller DOM efficiency (nominal?) is favored to be confirmed

Page 14: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Stabilizing likelihood maximization

Compare: before and after (max-min); before and after (rms)5 deviations in llh compared at each varied ice layer

Before/after: zero unfolded components kept/reset after initial nnls approximation step

Page 15: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

New llh with SPICE Lea

Page 16: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: old llh

Py=2.70

W=0.50

Page 17: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: new llh

Llh=435.996

Py=2.78

W=0.56

Page 18: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: old llh

Page 19: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: new llh

Page 20: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: old llh

Page 21: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

LEA: new llh

Page 22: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Model error: new vs. lea

new lea

Page 23: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Model error: new vs. lea

new lea

Page 24: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Linearity data vs. simulation

new lea

Page 25: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Fixing linearity; further improvements

Try removing the DOM flasher-receiver pairs on the same string issue remains

Possible difference between nominal and DC DOMs? same trend present in both

Statistics of the simulation matters? no, same effect for 1x and 10x

Found sub-optimal digitization of charge in data: fractional charges are rounded off to the nearest integer after initial binning but before optimized re-binning.

change the order of rounding? or avoid rounding altogether … (-) also add charge sampling from SPE for simulation … (-) modify likelihood (to conv. of Poisson and SPE)?

Likelihood width: 15% back to 20% … (-)

Page 26: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Summary and outlook• Fit now runs on gpu nodes of the npx4 cluster

allows for more precise simulation (up to ~ 10x statistics)

• Much improved initial approximation, based on the dust and EDML logs much better extrapolation outside the detector volume

• Re-worked calculation of llh and maximization algorithmBetter likelihood valuesReduced model errors even for existing modelsMuch more robust maximization (eliminated spurious outliers)

• Unfortunately a linearity issue surfaced, trying to understand it now Investigated a number of detector effects, which have little effect

Fix the next model and make it available to simulation production Map the anisotropy (magnitude and direction) everywhere in the detector Write the ice update paper with emphasis on anisotropy

Page 27: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Extra slides

Page 28: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

1+6-string flasher data

Page 29: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Beam geometry optimization

Page 30: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Reduced photon yield

Page 31: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: old

Py=2.11

W=0.38

Page 32: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new -1

Llh=444.008

Py=2.99

W=0.80

Page 33: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 0

Llh=420.315

Py=2.49

W=0.50

Page 34: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 1

Llh=416.894

Py=2.38

W=0.47

Page 35: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: old

Page 36: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new -1

Page 37: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 0

Page 38: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 1

Page 39: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: old

Page 40: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new -1

Page 41: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 0

Page 42: Ice model update Dmitry Chirkin, UW Madison IceCube Collaboration meeting, Calibration session, March 2014

Next: new 1