Charged Particle Tracking Hands-On Dustin Anderson, Steve Farrell, - - PowerPoint PPT Presentation

▶
charged particle tracking hands on
SMART_READER_LITE
LIVE PREVIEW

Charged Particle Tracking Hands-On Dustin Anderson, Steve Farrell, - - PowerPoint PPT Presentation

Nov 22, 2022 45 likes •607 views

Charged Particle Tracking Hands-On Dustin Anderson, Steve Farrell, Dorian Kcira, Jean-Roch Vlimant Content Hands-on Logistics Experimental setup Charged particle trajectories Charged particle tracking Algorithms in use

slide-1
SLIDE 1

Charged Particle Tracking Hands-On

Dustin Anderson, Steve Farrell,

Dorian Kcira, Jean-Roch Vlimant

slide-2
SLIDE 2

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 2

Content

  • Hands-on Logistics
  • Experimental setup
  • Charged particle trajectories
  • Charged particle tracking

➔Algorithms in use

  • Other approaches
slide-3
SLIDE 3

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 3

Hands-on Logistics

slide-4
SLIDE 4

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 4

Schedule

  • Monday 1h30
  • Wednesday 1h20 + 1h30
  • Thursday 1h20 + 1h30
  • Friday

➔Wrap up presentation

slide-5
SLIDE 5

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 5

Register

  • Add your name and github username on this googledoc

➔https://docs.google.com/spreadsheets/d/1s3QIJvgrfyKD9D ➔We will update it with the machine name and the

gpu to be set

  • Navigate to the hub url provided

➔You will have to allow the github app access

  • Start a session

➔This will spawn a jupyter session on the machines at

caltech

  • Register to the slack channel

➔https://join.slack.com/dshep2017/shared_invite/MTgwNzE

slide-6
SLIDE 6

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 6

Setup

  • Open a new terminal in jupyter (new : top right)

➔git clone

https://github.com/HEPTrkX/heptrkx-dshep17.git to get the handson material

  • Set CUDA_VISIBLE_DEVICES to the integer between

0 and 7 in first cell of NB

  • Tracking hands-on are in the hands-on/ directory
  • The data is imported from /inputdata/
  • Basic tutorials are available under tutorial/

➔ Chances are that root_numpy won't work at this time

slide-7
SLIDE 7

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 7

Dataset

  • Kudo to Yetkin Yilmaz, David Rousseau, Balazs Kegl, Isabell

Guyon, Mikhail Hushchyn from the RAMP challenge during https://ctdwit2017.lal.in2p3.fr/

  • Simple 2D geometry

➔ ~4k events (event_id index) ➔ Distribution of tracks with poisson distribution with mean 10

(cluster_id index)

➔ Flat pT distribution between 300(100) MeV and 1GeV ➔ 9 layers (layer index) ➔ Granular in phi (iphi index) ➔ 2D hits (x,y global position)

  • Challenge in preparation, only simplified generator can be

used for now

slide-8
SLIDE 8

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 8

Dataset

slide-9
SLIDE 9

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 9

Hands-on

  • Get people started, for those who need to get started.
  • Brainstorm on way to attack the problem
  • Simple starting kit model
  • Track candidate prediction with convolutional neural

nets

  • Track parameters prediction with CNN and LSTM
  • Hit association prediction with sequence to sequence
  • ...
slide-10
SLIDE 10

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 10

Experimental Setup

slide-11
SLIDE 11

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 11

The Large Hadron Collider LHC

8.5 kilometers

slide-12
SLIDE 12

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 12

Collision at the LHC

Bunch crossing

  • 1011 protons per bunch
  • Bunch crossing every 25 ns (40MHz)
  • Average number of proton-proton interaction per bunch

crossing in ALTAS-CMS : 25-45

slide-13
SLIDE 13

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 13

ATLAS

slide-14
SLIDE 14

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 14

ATLAS Inner Detector

http://atlas.cern/discover/detector/inner-detector 2T solenoid magnetic field along the beam line

slide-15
SLIDE 15

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 15

ATLAS Visualization

Shown trajectories are reconstructed objects

slide-16
SLIDE 16

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 16

CMS

slide-17
SLIDE 17

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 17

CMS Tracker

http://cms.web.cern.ch/news/tracker-detector 3.8T solenoid magnetic field along the beam line

slide-18
SLIDE 18

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 18

CMS Visualization

Shown trajectories are reconstructed objects

slide-19
SLIDE 19

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 19

LHCb

slide-20
SLIDE 20

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 20

LHCb Trackers

Vertex Locator Tracker Trigger Tracker (TT)

1T Magnetic Field

Outer Tracker

https://lhcb-public.web.cern.ch/lhcb-public/en/Detector/VELO-en.html https://lhcb-public.web.cern.ch/lhcb-public/en/Detector/Trackers-en.html

slide-21
SLIDE 21

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 21

LHCb Visualization

Add a couple of event displays Shown trajectories are reconstructed objects

slide-22
SLIDE 22

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 22

A Charged Particle Journey

slide-23
SLIDE 23

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 23

First order effect : electromagnetic elastic interaction of the charge particle with nuclei (heavy and multiply charged) and electrons (light and single charged) Second order effect : inelastic interaction with nuclei.

slide-24
SLIDE 24

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 24

Magnetic Field

  • Magnetic fieldB acts on charged

particles in motion : Lorentz Force

  • The solution in uniform magnetic field is

an helix along the field : 5 parameters

  • Helix radius proportional to the

component of momentum perpendicular to B

  • Separate particles in dense

environment

➔ Bending induces radiation :

bremsstrahlung

➔ The magnetic field has to be known to

a good precision for accurate tracking

  • f particle
slide-25
SLIDE 25

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 25

Multiple Scattering

  • Deflection on nuclei (effect from

electron are negligible)

  • Addition of scattering processes
  • Gaussian approximation valid for

substantial material traversed

Gaussian Approximation

slide-26
SLIDE 26

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 26

Bremsstrahlung

  • Electromagnetic radiation of

charged particles under acceleration due to nuclei charge

  • Significant at low mass or high

energy

  • Discontinuity in energy loss

spectrum due to photon emission and track curvature

➔ Can be observed as kink in the

trajectory or presence of collinear energetic photons

slide-27
SLIDE 27

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 27

Energy Loss

  • Momentum transfer to electrons when

traversing material (effect of nuclei is negligible

  • Energy loss at low momentum

depends on mass : can be used as mass spectrometer

ALICE Experiment

slide-28
SLIDE 28

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 28

Summary on Material Effects

  • Collective effects can be estimated

statistically and taken into account in how they modify the trajectory

  • Bremstrahlung and nuclear interactions

significantly distort trajectories

slide-29
SLIDE 29

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 29

High Luminosity LHC The Challenge

slide-30
SLIDE 30

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 30

Cost of Tracking

  • Charged particle track reconstruction is one of the most CPU consuming

task in event reconstruction

  • Optimizations (to fit in computational budgets) mostly saturated
  • Large fraction of CPU required in the HLT. Cannot perform tracking

inclusively at CMS and ATLAS. Online tracking strategy for LHCb.

slide-31
SLIDE 31

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 31

HL-LHC Challenge

<PU>=7 <PU>=21

<PU>=140-200 Circa 2025

  • CPU time extrapolation into HL-LHC era far surpasses growth in

computing budget

  • Need for faster algorithms
  • Approximation allowed in the trigger
slide-32
SLIDE 32

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 32

Algorithms In Use

slide-33
SLIDE 33

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 33

Event Reconstruction

From individual measurements in sub-detectors to kinematics and properties of particles created in collisions

slide-34
SLIDE 34

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 34

In a Nutshell

Seeding Kalman Filter

  • Particle trajectory bended in a

solenoid magnetic field

  • Curvature is a proxy to

momentum

  • Particle ionize silicon pixel

and strip throughout several concentric layers

  • Thousands of sparse hits
  • Lots of hit pollution from low

momentum, secondary particles

  • Explosion in hit combinatorics in both seeding and stepping pattern recognition
  • Highly time consuming task in extracting physics content from LHC data
slide-35
SLIDE 35

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 35

In a Nutshell

  • Hits preparation
  • Seeding
  • Pattern recognition
  • Track fitting
  • Track cleaning

Several Times

slide-36
SLIDE 36

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 36

Hit Preparation

  • Calculate the hit position from barycenter of charge

deposits

  • Use of neural net classifier to split cluster in ATLAS
  • Access to trajectory local parameter from cluster

shape

  • Remove hits from previous tracking iterations
  • HL-LHC design include double layers giving more

constraints on the local trajectory parameters

slide-37
SLIDE 37

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 37

Seeding

  • Combinatorics of 2 or 3 hits

with tight/loose constraints to the beam spot or vertex

  • Seed cleaning/purity plays

in an important in reducing the CPU requirements of sub-sequent steps

➔ Consider pixel cluster

shape and charge to remove incompatible seeds

  • Initial track parameters from

helix fit

slide-38
SLIDE 38

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 38

Hough Space Binning

  • Project hits within a search window onto a reference plane
  • Find clusters of hits
  • Done in innner and outer tracker of LHCb thanks to low

density of hits

slide-39
SLIDE 39

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 39

Pattern Recognition

  • Use of the Kalman filter

formalism with weight matrix

  • Identify possible next layers

from geometrical considerations

  • Combinatorics with compatibles

hits, retain N best candidates

  • No smoothing procedure
  • Resilient to missing modules
  • Hits are mostly belonging to one

track and one track only

  • Hit sharing can happen in dense

events, in the innermost part

  • Lots of hits from low momentum

particles

slide-40
SLIDE 40

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 40

Kalman Filter

  • Trajectory state propagation

done either

✔ Analytical (helix, fastest) ✔ Stepping helix (fast) ✔ Runge-Kutta (slow)

  • Material effect added to

trajectory state covariance

  • Projection matrix of local helix

parameters onto module surface

➔ Trivial expression due to local

helix parametrisation

  • Hits covariance matrix for pixel

and stereo hits properly formed

✗ Issue with strip hits and

longitudinal error being non gaussian (square)

slide-41
SLIDE 41

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 41

Track Fitting

  • Use of the Kalman filter

formalism with weight matrix

  • Use of smoothing

procedure to identify

  • utliers
  • Field non uniformity are

taken into account

  • Detector alignment

taken into account

slide-42
SLIDE 42

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 42

Cleaning, Selection

  • Track quality

estimated using ranking or classification method

➔Use of MVA

  • Hits from high quality

tracks are remove for the next iterations where applicable

  • Efficiency shoul
slide-43
SLIDE 43

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 43

Performance

  • It is not mandatory to reconstruction 100% of all tracks

➔ Tracks within jets are crucial ➔ Tracks in very dense jet helps with resolution and

identification

  • Tracking in the high level trigger could suffer

performance degradation in favor of speed-up

  • Predominance of low momentum tracks from
  • verlapping collisions

➔ No need to reconstruct them all ➔ Would be good enough to reject the hits

slide-44
SLIDE 44

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 44

Other Approaches

Material from Connecting The Dots https://indico.hephy.oeaw.ac.at/event/86/ https://ctdwit2017.lal.in2p3.fr/

slide-45
SLIDE 45

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 45

CA at CBM

slide-46
SLIDE 46

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 46

Associative Memory

Giacomo Fedi, CTD 2016

slide-47
SLIDE 47

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 47

Parallelism

slide-48
SLIDE 48

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 48

Hough Transform

slide-49
SLIDE 49

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 49

Hough Transform

slide-50
SLIDE 50

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 50

Hough Transform

slide-51
SLIDE 51

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 51

RANSAC

slide-52
SLIDE 52

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 52

Broken Lines

A.Schoning, Heidelberg University

slide-53
SLIDE 53

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 53

Broken Lines

A.Schoning, Heidelberg University

slide-54
SLIDE 54

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 54

Tracklets

Louise Skinnari (Cornell University)

slide-55
SLIDE 55

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 55

Combine hough transform, image templates and graph representation to build relations between track and hits. Use of decision tree to remove ambiguities. Ellipse are hits Diamonds are tracks

“Disconnecting the Graphs”

slide-56
SLIDE 56

05/08/17 DS@HEP2017, Fermilab, vlimant@cern.ch 56

New computing paradigm Kalman filter for a simplified tracking Progress on understanding precision

Neuromorphic Hardware