Title: Three first steps towards SUSY discovery with Neural nets
1Three (first) steps towards SUSY discoverywith
Neural nets
V.Zhukov University Karlsruhe CMS SUSY
group Prof. W.de Boer, A.Cakir, D. Daeuwel,
M.Niegel, S.Reisser, D.Stricker, E.Ziebarth
1. Find observables and create a Selector 2.
Validate the Simulation Model 3. Suppress
backgrounds and identify the signal
2Introduction
In this talk
SUSY model mSUGRA (mo, m1/2, tanb, A, m) used
as a reference, other extensions (MSSM, split
SUSY, NMSSM, GMSB) are similar
Simulations MC ISASUGRAPYTHIA,
PROSPINO(Kfactor) CMS detector official CMS
framework FAMOS_1_4_0 verified with ORCA_8_3_1
Data samples Signal mSUGRA m0-m1/2 inclusive
scan (10000ev per point 100x25 GeV) Bkg most
important SUSY bkg ttbar(2l), Zjets(2l),
Wjets(ln), QCD(1fb-1 )
Analysis based on NeuroBayes phi-t neural
network(NN) package available at
Karlsruhe. Concentrate on inclusive search in
low statistic regime as an example. (At higher
statistics the kinematic end points can be used)
3Remind mSUGRA cross sections
Contributions of different production
channels normalized to the total SUSY cross
section
Main production channels
? tot pb
tan?50 A00
Total mSUGRA cross section(LO)
KNLO 1.3-1.7
4mSUGRA topologies
mSUGRA averaged observables in m0-m1/2 plane
tan?50 A00
CMS FAMOS
5SM backgrounds
most important SM backgrounds to the inclusive
mSUGRA search
W,Z leptonic decays in bosons production
6Event selections
CMS experience so far.
- Define robust observables (MET, Heff, Njets,
ETj1,2,3, etc) which are well modelled and the
uncertainties are well controlled. - Use set of selection cuts for each observable
to remove major backgrounds, or use a Likelihood
to combine many observable in one. - Selection cuts can be optimized for a
particular SUSY model using for.ex. genetic
algo, decision tree or ?2 minimization.
Selection cuts in CMS Physics Technical Design
Report (PTDR) analysis
7Selection efficiencies for the CMS PTDR analysis
tan?50 A00
mSUGRA m0-m1/2 plane
JetsMETm
JetsMET
SS 2mJetsMET
OSSFJetsMET
Trileptons
OS 2tJetsMET
8CMS discovery reaches with fixed cuts
Needs bkg rejection 10 6 Typical rejections
for LM1 point with METgt100 GeV 10 3-10
5 (ttbar, wj,qcd) ET jet1gt 100 GeV 10 3
(wj,zj,qcd) Njets gt3 10 2
(DY,Zj,Wj,qcd)
Small efficiency already at trigger level.
Inclusive trigger efficiencies in mSUGRA plane
Main trigger streams for SUSY
L1 in respect to MC
HLT in respect to L1
9Find observables
The selection efficiency depends on the used
observables. How to select them?
Two basic types 1. related to the mass scale
MET, sumET, JetET, PTlepton, Heff ,... (
sensetive to the energy scale errors) 2. related
to topology Njets, Nleptons, Invmass(OSSF,
etc),asymmetry, angules, cos(J1MET),
cos(OSSFMET), Ehad/Etot, MET/ETtot etc
(sensetive to the reconstraction errors and
efficiencies)
Examples LM1 (60,250,10) - ttbar (inclusive)
The observables can correlate, have systematic
uncertainties, have features in distributions,
etc. Needs a detailed study.
Use Neural Net (phi-t) to qualify the
observables
- Individual preprocessing , preselection and
conditioning of each observable - Decorrelation of the whole set of observables to
get rid of linear relations - Train NN for the nonlinear part
See backup slide for details...
10 Preprocessing
Example LM1(60,250,10)-ttbar
Observables I no preprocessing
Observables I (simple set) METx, METy
Njets(gt30GeV), ETj1,ETj2,ETj3,??1,2,3 decorrelate
linear relations
target metx mety Etj1 Etj2 Etj3 ?? ?? ?? Njets
With preprocessing flatenning, splines
target metx mety Etj1 Etj2 Etj3 ?? ?? ??
Preprocessing enhances selective power
significantly
Observables II (large set) MET,Meff,sumET,Njets
Nj,ETj1,ETj2,ETj3,?1,2,3 cos(j1j2),
cos(j1met),cos(j2met) Nlept, Ptl1, Ptl2,
Ptl3,MinvOSSF, Assym, ?1,2,3, more than 30
variables. (only most significant
used) Individual preprocessing
Observable II preprocessed and conditioned
11Observables in different mSUGRA regions
efficiency
NNout for CMS SUSY points - ttbar
LM7
NNout can be treated as a model probability
P(md) NNout(0)50 NNout(1)100 One can set
a cut on Nnout expecting a defined significance
contamination
LM1
LM5
LM9
Worst Focus Point
LM10
HM3
Best High Mass
Can sort observables by significance for each
mSUGRA model
12Observables for different backgrounds
Usually most efficient to have for each bkg
channel(class of events) the own list of
observables (and own NN). They can be combined
according to contributions.
for.ex. LM9- different bkgs
13Uncertainties in the Simulation Model
Have to understand the SM backgrounds and
detector before looking for SUSY
LHPDF uncertainties in the trilepton search
Theory QCD multijets PS versus ME,
nonperturbative effects change the jets ET
distribution can reduce rejection by factors when
using Meff., Cross sections LO, NLO. PDF
uncertainties 5-10 State radiations,
underlying event, beam remnants increases MET.
Detector simulation Detector responce often
requires MC corrections (Ej5.6/ET ) and
calibrations (JES10). Long term unstabilities
can affect the results, etc.
Fake jets rate vs ET in ttbar events
- Reconstruction
- Jets JES errors can be reduced the gamma/Z-jets
balancing technique, but the stoch. term
remains sET 1.25/?ET . - Fake jets due to noise and underlying events
- MET sMET SET , have contribution from
unclustered jets tails, beam remnants, etc. The
minimum MET 30 GeV. - Leptons Electron identification is complicated,
depends on Calo. - Muons and Electrons are affected by fakes Fe10-4
, Fm10-5 per W/Z jet significantly
increase bkg for SUSY
CMS MET resolution versus ET
14Validate Simulation Model with NN
We can compare distribution of all observables
one by one, but which part is important for the
SUSY selection?
Validate the observables relevant for the SUSY
search only.
Data -driven calibration 1. preselect a data
sample for a particular channel(class of
events) for .ex. Zjets( by invariant mass),
ttbat (top reconstruction), etc. 2. setup a NN
with observables used for SUSY discovery. Use
Sim data for the tested channel as a signal and
Real Data as bkg.
Nnout between truly ttbar(red) and smeared
ttbar (black)
Example ttbar and ttbar smeared with the
uncertainties typical (x3 higher) for CMS
These events have to be studied in details, they
can fake the signal!
15Background suppression with NN
Preselection efficiency in m0-m1/2 plane
m1/2
Cut most of trivial bkgs by simple cuts.
Split the rest of events on different classes
(can join different bkg. channels) and train
separate NN. The linear correlation between
variables can be removed and observables
preselected.
m0
TTbar rejection efficiency for SUSY (fixed cut on
NNout)
Train NNs for each point SUSY(m0 ,m1/2) -ttbar.
The non uniformity will contribute to the
uncertainties and all NN benefits will dissapear.
10 nonuniformity, But the NN cuts can be
optimized to equalize responses But even more
model dependency
16Compare NN selection and fixed cuts
Train 400 NN(preselect) for mSUGRA(mo,m1/2)-
(ttbar,qcd,wj) compare with JetMET selection
in CMS PTDR.
PTDR METJet selections METgt200 Njgt2,
ET180,110,30 ??lt1.7 ??(metj2)gt20
NN with observables (12 observables) MET, Heff,
ETj1, ETj2, ETj3, Ptl Nl, Nj,Njb cos?(j1j2),
cos?(j1met), cos?(j2met)
SNsig/sqrt(Nbkg)
Significance plot (no syst.) in m0- m1/2 mSUGRA
plane , tanb10
Lint10fb-1
Lint10fb-1
Red area 5s reach
Visually, little of improvement, but remember
? m1/2 4
17SUSY region separation with NN
S2-(S1S3)
When backgrounds are gone
NN can be used to identify the mSUGRA
model already at low statistics by looking on
event topology.
NN out
For.ex 3 mSUGRA regions train NNs for each
combination
Selection efficiencies of 3 NNs in mSUGRA
plane(fixed NNout cut)
S1/S2S3
S2/S1S3
S3/S1S2
tb50
18More model identification
Can identify SUSY production channels by event
topology
Train NNs for 3 individual channels against
others for one test point (m0500 m1/2200,
tb50)
Selection efficiencies in mSUGRA plane of NN
trained for different production channels at (
m0500 m1/2200 tb50)
MC and Rec fractions
The selection efficiency E is a convolution of
NN and preselection efficiencies Eie (presl)ki
e (NN)k . Find the fractions F i from the
observed Di DiEi X Fi and compare with the
expectation from MC(normalized to some chann.)
Better 20 accuracy!
Check another point(m02500 m1/2600, tb50)
with the NN trained for ( m0500 m1/2200, tb50)
Larger errors, but still can constrain the ratios
of different production channels
19Summary
The search for SUSY at the beginning of LHC at
Lintlt10fb-1 would require a qualification off all
observables and validation of the simulation
model. This can be done with Neural Network
tools. The selection cuts can be optimized
iteratively with the NN.
The topology of the selected events , and
therefore the SUSY model, can be constrained
already at small statistics using a NN topology
selector. More detailed reconstruction of SUSY
parameters can be done with the kinematic end
points analysis
We will find it...
20BackUp Qualify the observables
Processing MET in LM1-ttbar (inclusive)
Same for the cos(j1j2) not as good
Observables sorted by significance (contribution
to the NNout in )
Enough to keep only first 10 in NN...