NEURAL NETWORKS IN HIGGS PHYSICS

1
NEURAL NETWORKS IN HIGGS PHYSICS

• Silvia Tentindo-Repond,
• Pushpalatha Bhat and Harrison Prosper
• Florida State University and Fermilab D0
• ACAT - Fermilab 16 Oct 2000

2
Higgs Physics

• The most challenging task of HEP ( Tev and LHC )
  in the coming decade will be the search for
  Higgs.
• In many theories, the Higgs Boson would explain
  the still mysterious fundamental mechanism of the
  electro-weak symmetry breaking (EWSB).
• SM predicts Higgs in the mass range
  107 Gev ( - 45, 67 )
• MSSM predicts a lighter Higgs at 130Gev, that
  would be reachable at Tev
• Studied here 90 lt Mhiggs lt 130 Gev

3
Predicted Higgs Mass from SM (measured Mtop and
MW)
4
Integrated Luminosities for Higgs Discovery at
Tev vs Higgs Mass (SM Higgs)

5
Multivariate Methods vs Traditional in Higgs
Physics

• Multivariate Methods (NN) are used to maximize
  the chance to discover the Higg Boson
• To reduce the required Luminosity for equal
  signal Significance (S/sqrtB)
• To reduce the required Luminosity for making a 5
  sigma discovery

6
SM Higgs final states
use b-tagging to reduce background
use leptons to reduce QCD backg.
M.Spira
use particular lepton signatures, use angular
correlations to reduce di-boson backg.
7
Typical cross-sections ( TeV)
spb (mH100 GeV)
gg H
1.0
WH
0.3
ZH
0.18
WZ
3.2
WZ/ZH production are preferred
Wbb
11
tt
7.5
tbtqtbq
3.4
QCD
O(106)
8
Traditional Analysis vs NN

• Example
  p p ? W H ? l v b b
  signal p p ? W b
  b background
• Need to enhance signal over background
  Use global event variables (Ht, Sph,
  Apla,MissEt, etc) jet variables ( Etj,
  Etaqj,Ehad, Eem,Ntr,Etr,btag,Ht
  InvMass(jj), etc )
• Use corrections ( e.g. jet energy corrections).
  Use parametrized b tag displaced vertex,soft
  lepton - etc.)

9
Traditional analysis vs NN (cont.)

• Traditional Analysis improves S/B by imposing
  cuts to each event variable. Rarely
  optimized,unless signal and background
  distributions are well separated.
• Multivariate Analysis uses for example NN to find
  optimal cuts. optimizes separation between signal
  and background therefore maximizes the chance
  of discovery

10
Example of NN for Higgs Search
PRD62,2000

• Study the process
• p p -gt W H -gt l v b b signal
  p p -gt Z H
  -gt l l- b b
  p p -gt Z H -gt v v b b
• NN analysis of these three processes leads to
  remarkable Luminosity reduction allowing Higgs (
  90 lt MH lt 130) discovery at Tev
• NN variables used to train
  
• Etb1, Etb2, M(bb), Ht,Ete,ETAe,Etmiss, S,
  dR(b1,b2), dR(b1,e)
• NN configuration 7 input 9 hidden nodes 1
  output node

11
NN for Higgs search training variables
WH -gt ev bb
Dark Signal Light - background
12
NN for Higgs search NN Output
WH Signal - D1
WBB Bkgd D0
13
NN and Higgs Search required Luminosities
Compared required Luminosities for Higgs
Discovery NN cuts and Standard Cuts
14
NN and Higgs Search Luminosity further studies
YES
Can we do better ??

• Re-train NN
• Configuration 6-6-1
• same as previous, but no S
• Different number of epochs
• and hidden nodes .
• ---------------------------------
• Configuration 8-6-1
• - same as before, add ntrj1 and ntrj2

NO
15
NN and Higgs Search Luminosity further studies
16
NN and Higgs Search Luminosity further studies
17
Channel-Independent B tagging with NN for Higgs
Search
Heavy Flavor Tagging ( C and B jet tagging)
R. Demina Traditional Analysis makes no
distinction from b and c. NN Analysis
combines lifetime variables (track consistent
with secondary vertex, Impact Parameter ) and
kinematic variables (mass, fragmentation) This
tagging method can potentially outperform
existing Tagging algorithms .
18
Channel-independent B Tagging NN output
(bottomness)
bottom charm primary
R. Demina, march 2000
Points- single m data, black - fit.
19
Channel-Independent B Tag NN output (m
jet)
bottom
charm
primary
R. Demina march 2000
20
Channel-dependent B tagging with NN for
Background Reduction in Higgs Search

• In this study
• Signal 1000 W H ? e v b b
  Background 1000 W bb
  
  --------------------------------------
  
• Parton level Monte Carlo PYTHIA
• ( later on
  CompHEP ) Parton
  fragmentation PYTHIA
  Approximate response of Detector (
  D0/CDF) SHW program - includes simulation of
  trigger, tracking, cal cluster, reconstruction
  and b tagging . J.Conway

21
Channel-dependent B tagging with NN (cont.)

• Cuts for base sample Pte gt 15 Gev/c
  ETAe lt 2, Met gt 20 Gev,
  Etjet gt10Gev, Njetgt2, ETAjetlt2
• Select jet variables that are connected to b tag
  of jet
• Selected Etjet, Ntr jet, Width jet
• Train NN with a signal sample WH ? e v b b
  
• NN configuration 3 - 5 1
  3 input
  nodes Etjet, Ntr jet, Width jet
  5 hidden nodes
  
  1 output Channel-Dependent B tag
• Set NN function ( D 1 for B jet, D0 for non B
  jet)

22
Channel-dependent B tagging with NN (cont.)

• QUESTION Does this channel-dependent b-tagging
  push to lower values the background (
  Wbb Massjj distribution ?)

23
Channel-dependent B tagging Jet variables for NN
training
24
Channel-dependent B tagging Jet variables for NN
training
25
NN HB Tag output for B-flavor/no-B-flavor jets (
j1)
26
NN HB Tag output for B-flavor/no-B-flavor jets (
j2)
27
NN HB Tag output for Wbb
28
NN HB Tag output for WH (100)
29
NN HB Tag cut0.4 WH(100)
30
Channel-dependent B tagging separation
signal/background
31
Improving Mass Resolution with NN in Higgs Search

• M(jj) has proven to be a critical variable to
  discriminate signal from background in Higgs
  physics, for any channel analysis
• The assumed mass resolution in the recent RunII
  Susy/Higgs Workshop is 10.
• Methods and algorithms have still to be worked
  out to reach such resolution

32
Mass Resolution Parton and Particle jets- Final
State Radiation contributions
33
Mass Resolution Parton and Particle jets- Final
State Radiation contributions
34
Mass Resolution Detector jets - Final State
Radiation contributions
Signal W H (M_H 100 gev )
Background W b b
35
Improve Mass resolution with NN in Higgs Search (
cont.)

• Possible strategies
• Study correlations of jet properties and Inv Mass
  distribution.
• Make a correction function to improve Pt and
  Energy Resolution of jets and recalculate Inv.
  Mass of jets with the corrected values of Pt and
  E

36
Improve Mass resolution with NN in Higgs Search (
cont.)

• Study correlations among Jet variables and Massjj
  Jet Variables Nj, Et, Phi, ETA,
  d(e,j), Eem, Ehad, Etr, Ntr, Wid,
  
• plus Btag, d(b,j) , d(j,j), Mjj , Mbjj .
• No clear evidence of correlation.
  
• Apply corrections to Pt and E that could improve
  the Mjj resolution.

37
Corrections to Mass Resolution I

• Train NN to correct Mjj by giving Mjj and Ht
  and forcing the output to be
  the true Higgs mass, for several values of Higgs
  masses
  
• NN configuration 2-6-1
  2 input nodes ( Mjj, Ht
  )
• 6 hidden nodes,
  1 output node ( MH) for
  several MH
  300 epochs
  500 examples for each Higgs Mass
• MH 100, 105,110,115,120,125,130,135,140

38
Improving the Higgs Mass Resolution
Use mjj and HT (? Etjets ) to train NNs to
predict the Higgs boson mass
13.8
12.2
13.1
11..3
13
11
39
Corrections to Mass Resolution II

• Train NN to correct Pt and E of jet, by giving Pt
  distributions at parton level.
  Generate a corrected Pt
  function Ptc(Et, Eta) to apply to Mjj .
• NN configuration 2-9-1
  2 input nodes , 9 hidden
  nodes, 1 output node ( Mjj ) 5000 examples
• ..

40
Summary

• NN used to maximize Discovery Potential
• B Tagging and good Mass ( Mjj) Resolution
• NN for B Tagging is very promising
  ( could Channel-Dependent B Tagging be used
  for reduction of Background ? )
• Plan to continue systematic studies of the methods