1. Lepton identification in Au+Au at 1.23 GeV/u with HADES
using multi-variate analysis
Szymon Harabasz for the HADES Collaboration
Outline:
Electromagnetic probes in heavy-ion collisions
The HADES experiment
Physics goals
Experimental challenges
Possible solutions
Summary
Write you name, not just S. Harabasz |

2. Electromagnetic probes in heavy ion collisions
g, g*: No strong final state interactions  leave reaction volume undisturbed
Reflect whole ”history” of collision:
From pre-equilibrium phase
From thermalized medium QGP and hot hadronic gas
From vector meson decays after thermal freeze-out
They probe the electromagnetic structure of dense/hot matter
fine
Invariant mass monitors directly the spectral function |

3. Observables: light vector mesons
Vacuum
Both r and g* have JP=1−
Medium
pp interactions and baryonic excitations f |

4. Meet the HADES
Beams from SIS18: pions, protons, nuclei, Ekin=1-2 GeV/u
Search for very rare probes
Di-lepton production governed by the factor a2
Branching ratio to e+e−: 7.14⨯10-5
Vector meson production sub-threshold
Need for fast detector
Need for large acceptance
Excellent particle identification
Need for excellent mass resolution
Track multiplicity as large as 300 per event ( combinatorial background)
 Efficient track reconstruction
e+e- use upper case for + and –
Text restructured |

5. HADES Curiculum Vitae
Explore strongly interacting matter under extreme conditions (heavy ion collisions)
Address various aspects of baryon-resonance physics (elementary collisions)
Stage I ( )
Light collision systems  limited granularity of time- of-flight system
15 peer reviewed papers since 2009
Stage II ( )
Heavy collision systems (Au+Au at 1.23 GEV/u April 2012)
p-induced reactions (2014)
Stage III ( …)
Excitation function for low-mass lepton pairs and (multi-)strange baryons and mesons.
Medium-heavy systems up to 8 GeV/u

6. Virtual photon emission in heavy-ion collisions
Ar+KCl compared to reference after subtraction of contributions from h
Isolation of excess by a comparison with a measured “reference” spectrum
First evidence for radiation from the “medium”!
Excess yield scales with system size like Apart1.4
½[pp+pn]=C+C x
w  e+e-
Quest for heavier systems!
HADES: Phys.Rev.C84:014902,2011

7. Au+Au at 1.23 GeV/u Beam time April/May 2012
AuAu 1.23 GeV/u
pNb 3.5 GeV
Ar+KCl
pp 3.5 GeV
CC 2 GeV/u
dp 1.25 GeV
pp 1.25 GeV CC 1 GeV/u
HADES DAQ: Versatile, FPGA board based system using dedicated add-on boards and data/trigger/slow-control transport via serial optical links (TRBnet)
557 hours beam Au on Au target
( ) x 106 ions per second
8 kHz trigger rate
200 Mbyte/s data rate
7.3 x 109 events
140 x 1012 Byte of data

8. The HADES event reconstruction
Beam
Magnet
Title: The HADES event reconstruction.
The HADES spectrometer, is in general not ok since last “S” in the name stays for the “spectrometer” |

9. The HADES event reconstructionp− p+ p
He4 t
Beam
Magnet
Beam
Magnet e− e+ p− p+ K+ p
He3
He4 t

10. The HADES event reconstruction
Beam
Magnet e− e+ p− p+ p K+
He4
He3 t

11. The HADES event reconstruction
Beam
Magnet
field wires
signal wires
readout cathodes Q1 Q2 Q3
ΔQ=Q2+Q3-Q1

12. The HADES event reconstruction
Beam
Magnet

13. Experimental challenges
C4F10 radiator volume
Photon detector
VUV mirror
MDC I and II Δθ
High contribution of hadrons
Possibility to match them with RICH ring and misidentify them as leptons
Protons and pions
Electrons out of acceptance, γ conversions, misidentified particles…
→ Need for deep understanding of combinatorial background
(not technical, not in scale)

14. Why not use combined information from many observables?

15. Why not use combined information from many observables?

16. Toolkit for MultiVariate Analysis (TMVA)
Currently implemented classifiers
Rectangular cut optimisation
Projective and multidimensional likelihood estimator
k-Nearest Neighbor algorithm
Fisher and H-Matrix discriminants
Function discriminant
Artificial neural networks (3 multilayer perceptron implementations)
Boosted/bagged decision trees
RuleFit
Support Vector Machine
Advantages:
Not need to iteratively optimize each single cut
Finer, multidimensional decision boundary
Particle hardly accepted by a number of hard cuts may be removed by MVA
Particle removed by one hard cut but well fulfilling conditions of other cuts may be accepted by MVA
ROOT: is the analysis framework used by most (HEP)-physicists
Provide common analysis (ROOT scripts) and application framework

17. Nonlinear Analysis: Artificial Neural Networks
Achieve nonlinear classifier response by “activating” output nodes using nonlinear weights 1 i
. . . N
1 input layer
k hidden layers
1 ouput layer j M1 Mk
2 output classes (signal and background)
Nvar discriminating input variables
(“Activation” function)
with:
Feed-forward Multilayer Perceptron
Weight adjustment using analytical back-propagation

18. Use of neural network to identify leptons
Discriminating input variables:
RICH ring parameters:
Number of fired pads
Avg. charge per pad
Ring geometry parameters
Velocity β
Energy loss in MDC and TOF
Difference of charge in PreShower’s chambers
Momentum
Signal
Background
Num. of fired RICH pads
Avg. charge per RICH pad β
background
signal
TRAINING SAMPLE
MDC dE/dx
Shower ΔQ
Momentum |

19. Identification conditions
Neural network response function
Output: "probability", that the given particle belongs to signal (is a lepton)
RPC region, θ < 45O
TOF region,
θ > 45O e− e+ e− e+
Leptons p− p+ p− p+ p ?
Identification conditions
Hadrons |

20. Signal-to-background estimates using RICH rotation technique
Characterizing "true" (signal) and "random" (background) track-RICH ring matches 5 1 5 1 4 2 4 2 3 3
MDC
RICH
Rotate RICH software-wise by 60O
Correlate tracks with rings
Get only random matches

21. Signal-to-background estimates using RICH rotation technique
Characterizing "true" (signal) and "random" (background) track-RICH ring matches
Understanding of background with simulation
No ID cuts
electrons
pions
protons
muons
SUM
background from RICH rotation
The same analysis as for experimental data
UrQMD GeV/u Max. Impact parameter 9 fm The same analysis chain as for experimental data
Background
Signal |

22. Trade-off: purity vs. efficiency
Signal-to-background estimates using RICH rotation technique
Characterizing "true" (signal) and "random" (background) track-RICH ring matches
No ID cuts
MLP > 0.6
More ID cuts
You can add to the transparency:
Trade of: purity vs. efficiency
Background reduced by 93%
Signal reduced by 21%
Trade-off: purity vs. efficiency |

23. Comparison to hard cuts Purity and momentum distributions
Multi-variate
Hard cuts
purity
MLP – write better multi-variate
Integrated purity: 94.8 %, 84.8 %
STAR min. bias ~ 92 % PHENIX 90 %
[Trento 2013] |

24. Summary:
Di-leptons are excellent tool to investigate hot/dense nuclear matter formed in heavy ion collisions
Their measurement is challenging and requires effective detector system and sophisticated analysis methods
Using neural network algorithm one extracts from Au+Au HADES data a very pure sample of electrons
Next steps:
Combine identified leptons into pairs
Identify sources of combinatorial background and subtract it
You can say here like : “Fun will really start now” |

25. The HADES collaboration|

26. Thank you for your attention

27. Backup slides

28. GiBUU and HSD
[Janus Weil and Ulrich Mosel  2013 J. Phys.: Conf. Ser ]
[E. L. Bratkovskaya et al. Phys. Rev. C 87, (2013)]

29. Dileptons from heavy ion collisions
 contribution subtracted
[Janus Weil and Ulrich Mosel  2013 J. Phys.: Conf. Ser ]
[PRL 98 (2007) , PLB 690 (2010) 118]
[E. L. Bratkovskaya et al. Phys. Rev. C 87, (2013)]

30. Di-leptons and stages of a heavy-ion collision

31. Negatively-charged particles Positively-charged particles
Negatively-charged particles
0 100
Positively-charged particles |

32. Purity – definition Signal (hatched grey area) from not rotated RICH
after all PID cuts and
RICH matching QA cut minus background)
Background
(solid red area)
from rotated RICH
after all PID cuts
and RICH matching QA cut)

33. Identification conditions
Lepton identification results
RPC region, θ < 45O
TOF region, θ > 45O
Identification conditions