Longitudinal Vector Boson Scattering with Deep Machine Learning - - PowerPoint PPT Presentation

Longitudinal Vector Boson Scattering with Deep Machine Learning

Jake Searcy, Lillian Huang, Marc-Andre Pleier, Junjie Zhu

1

VBS and the Higgs

Without a Higgs the matrix element for Longitudinal VBS (VLVL) grows with energy until it becomes strongly coupled

A Higgs fixes this

2

The Questions: Does our Higgs fix this? How can we find out?

Longitudinal Scattering

Longitudinal bosons grow with M(WW) if there is no Higgs

Higgs-less Higgs-126 GeV

3

Extra useful info http://www.sciencedirect. com/science/article/pii/0550321385 900380

No Higgs With Higgs

Fraction of Longitudinal W Bosons

• Look in Same

Sign W∓W∓

• Two same sign

leptons

• Two jets

○ high M(j,j) ○ high Δn(j,j)

4

Event Signature

• Phys. Rev. Lett. 113, 141803

Same Sign W∓W∓

Di-leptonic

• Aim: Longitudinal scattering
• Pros

○ Easy to find ○ Signal is clean

• Cons

○ Two neutrinos

• Other options

○ OS WWjj-All top ○ Semileptonic WWjj-not yet sensitive to SM ■ ZZjj,WZjj-Very few events

• What can we do with 2

neutrinos

• Study parton level truth for

W∓W∓

5

• Phys. Rev. Lett. 113, 141803
• Phys. Rev. Lett. 114, 051801

How do we do it? Some Proposals

• K. Doroba, J. Kalinowski, J. Kuczmarski, S. Pokorski, J. Rosiek, M. Szleper, S.

Tkaczyk

Use specially built RpT variable

http://xxx.lanl.gov/pdf/1201.2768v2

• A. Freitas, J. S. Gainer

Use Matrix Element Analysis to differentiate different Higgs models

http://xxx.lanl.gov/pdf/1212.3598v2.pdf

We want to try and measure the longitudinal fraction directly

RpT = pT(Lep1)pT(Lep2) / pT(Jet1)pT(Jet2)

6

Measuring VLVL

• We’ve seen the first signs of VBS in W+W+

○ Next step is to see VLVL ■ Then can we measure VLVL at high M(W,W)?

• Effect of polarization is on the θ* distribution

Boost to W rest frame

θ*

7

e+ νe

W+ Direction

Cos(θ*) distributions - 1D

• Fits give polarization

fractions

• Of course can’t do this

in real events because

• f the two missing

neutrinos

• Do we have any

sensitivity with measurable quantities?

8

http://arxiv.org/pdf/1203.2165v2.pdf

... . . . Machine Learning Neural Networks

Really common in HEP to use multivariate techniques classification (discret estimation)

Signalness

. . . . . .

Event Inputs Outputs Hidden Layers

Just a simple f(xi)➝Output Train weights so this mapping gives you the best discriminate between signal and background Squash output between (0,1)

9

... . . . Machine Learning Neural Networks

You can also train NN to approximate continuous functions (Regression)

Signalness

. . . . . .

Event Inputs Outputs Hidden Layers

Squash outputbetween (0,1) Don’t squash output

10

... . . . Machine Learning Neural Networks

You can also train NN to approximate continuous functions (Regression)

My favorite truth value

. . . . . .

Event Inputs Outputs Hidden Layers

Just a simple f(xi)➝Output Train weights so this mapping gives you the best approximation of the function you want (minimum error2) My other favorite truth value

11

Goal

• 1. Take measurable quantities in Same Sign

W∓W∓, (pt,eta,phi, leptons and jets + met)

• 2. Train Neural Network to output the two

true values of Cos(θ*) (one of each W)

• 3. Fit Neural Networks Cos(θ*)

approximation to measure Longitudinal Fractions

Select Signal Events Subtract Background Fit Data Run Neural Network

12

Training the Neural Network: Deep Learning

• Deep learning is simply extending a simple

neural network with many layers

○ Conceptually simple, computationally a little difficult

• Has had a lot of success in recent days

○ In HEP and elsewhere ○ P.Baldi, P. Sadowski, D. Whiteson http://arxiv.org/pdf/1410.3469.pdf

“The deep networks trained on the low-level variables performed better than shallow networks trained on the high-level variables engineered by physicists, and almost as well as the deep network trained high-level variables,”

• Since we have only one “High-level variable” this is

exactly what we want.

13

Train Neural Network

• Use Deep Learning

○ Network with 20 Layers 200 Nodes ○ Instead of shallow one-layer ■ Though Results look pretty good with 1 layer ■ Gain ~20% with Deep learning

• Validate on independent data

○ Far from perfect, but certainly usable

14

Fit Neural Network

• 6 templates

○ ++,--,+-.LL,+L,-L

• Combine into 3

○ Transverse-Transverse ○ Transverse-Longitudinal ○ Longitudinal-Longitudinal

1 ab-1 Fit

15

Other Fit option

Also possible to do all six at once

16

Real Life

• Have to separate

Signal and Background

• Get a little help from

the NN

• Still need to make

event level cuts

• Finite detector

resolution

• ATLAS CUTS

○ M(j,j) > 500 GeV ○ dY(j,j) > 2.4 ○ MET > 30 GeV ○ Lepton pT > 25 GeV ○ Jet pT > 30 GeV

17

Sensitivity

Parton Level Delphes Simulation ATLAS Cuts ATLAS Cuts + Delphes

18

Some Comparisons

Compare against variable from Doroba et al. RpT = pT(Lep1)pT(Lep2)/pT(jet1)pT(jet2)

Log Scale Linear Scale

19

Fit Comparison

Precision at 3ab-1 68% Parton with Cuts NN: 7.5+2.6

• 2.4%

RpT: 7.5+11.3

• 7.5%

NN has > 4x the sensitivity to the LL

• fraction. (Equivalent

to 16x the data or hundreds of years of running the LHC!)

20

Future Studies

21

• A few nice properties of this regression that could

be explored more

• What differential

measurements possible?

• Train on Delphes

samples

• Can be combined

with cuts to enhance LL fraction (i.e. Doroba et. al.)

Conclusions

• Regression is a great tool for pulling out “hidden” information

○ Many applications beyond this one

• Measuring the properties of VBS possible in the same sign

WW state ○ Neural Networks can more than double the sensitivity of the current state of the art methods.

• Limits can be made as early as the first measurements.
• Strong bounds will require the High Lumi LHC
• Differential distributions could be possible
• Better understanding (generation) of the polarization

distributions will help

• Extra Thanks to

○ Olivier Mattelaer ○ Sally Dawson

22