Download Reconstruction and calibration strategies for the LHCb

Reconstruction and calibration strategies
for the LHCb RICH detector
R. Forty(1), C. Jones(2), C. Lazzeroni(2), R. Muresan(3), M. Patel (1), G.Wilkinson(3)
(1)CERN, (2) University of Cambridge, (3) University of Oxford
LHCb particle identification:
• The ability to distinguish between pions and kaons
over a large range of momenta is essential to extract
CP violation results in LHCb.
• Three radiators are used in the two RICH detectors:
silica aerogel and C4F10 gas to identify particles in
the range 2-50 GeV; CF4 for particles up to ~100 GeV.
• Rings are visualized by reflecting the cones of
Cherenkov light onto arrays of hybrid photon
detectors (HPDs).
LHCb ring pattern recognition algorithm requirements:
• To reconstruct distorted, incomplete
rings with
variable radius and a variable number of finite
resolution hits.
• To be robust in the presence of background.
• To process a large number of images, (~ 500 hits in
RICH1, 300 in RICH2) fast enough not to reach the
CPU limits for event reconstruction.
LHCb Detector
Flat mirrors
Spherical Mirrors
Support Structure
particles with tracks
background
RICH1
7.2 m
Ring not predicted
from tracking
information
Central Tube
Photon Funnel + Shielding
RICH1
Event display with the photodetector planes of RICH 1
drawn side by side (scale in cm) and the predicted
Cherenkov rings superimposed (likelihood method).
RICH2
Cherenkov angle reconstruction online:
• Offline strategy of calculating Cherenkov angle for each photon from RICH hit,
assumed track and knowledge of the detector optics, too slow online.
• Instead calculate approximate angles directly on the photodetector, applying local
coordinate “stretch” to correct optical distortions.
Online PID:
• “Local”: peak search in Cherenkov angle
distribution to look for the maximum
signal/background. Apply a momentum
cut to determine if the track came from
heavy/light particle.
• “Global”: similar to the offline likelihood
method
but
faster
since
full
reconstruction is not required.
Radius of the rings corresponding to the
most significant peak in signal/background
distribution (local PID).
Hough transform:
Reconstruct a given family of shapes from discrete
data points, assuming all the members of the family
can be described by the same kind of equation. To
find the best fitting members of the family of shapes
the image space (data points) is mapped back to
parameter space. Work on going to resolve the near
degenerate solutions.
Preliminary
hits, Hough centres,
track impact points
True p in blue
True K in pink
Markov rings
PID efficiency
Offline
Rich PID Callibration:
• D*  D0 (pK ) p events are selected in the high level trigger and then offline, using
kinematic information alone.
• A very high statistics sample ( 108 events per year ) of pions and kaons available for
evaluating PID information directly from data.
Kaon selection efficiency versus
momentum , as evaluated with D*
method
D0mass (GeV)
Bachelor pT (MeV)
cm
Metropolis- Hastings Markov chains:
Sample possible ring distributions according to
how likely they would appear to have been given
the observed data points. The best proposed
distribution is kept. Preliminary results are
encouraging, work on going to assess
the
performance of the method
D0peak
min bias
pure
signal
PID online: physics performances:
The time to perform the reconstruction is
~3 ms/event. This allows the use of RICH
PID in the trigger resulting in substantial
gains in efficiency with comparable
background rate. Two examples are
BsDsh and Bsff where 25% gain is
obtained.
PT distribution of the hadron, h, in B->Dh
candidates in the trigger, for both signal and
minimum bias (MB) events, before and after
applying the particle identification. The
background falls by an order of magnitude
while the signal is unaffected.
RICH2
RICH2
*authors are grateful to A. Buckley for
providing the picture and C. Lester for help in
implementing the method
Online
RICH pattern recognition:
• Likelihood: baseline offline method that has been
shown to perform very well. Does not reconstruct
rings but compares expectation based on track
information to observed data. Its limitation comes
from the associated track requirement and the CPU
time necessary: ~100 ms per event, to be compared
with the order 10 ms available to run the entire high
level trigger.
• Trackless ring reconstruction: Hough transform and
Markov chain methods, applied offline, perform ring
finding using the information from the photodetectors
alone and hence provide robustness against
background which does not have associated track
information.
• Online reconstruction: similar principle to offline
likelihood but simplified for increased speed whilst
maintaining acceptable performance, suitable for use
in the high level trigger.
21 D0s from 13M minimum bias events,
~ 0.8 seconds of data taking at
L= 2 x 1032 cm-2s-1. 500 M such D0 will be
selected by the high level trigger in one
year of running.
Initial sample
Bd->D*p decays
(GeV)