Pandora pattern recognition p for LArTPCs L. Escudero for - - PowerPoint PPT Presentation

▶
pandora pattern recognition
SMART_READER_LITE
LIVE PREVIEW

Pandora pattern recognition p for LArTPCs L. Escudero for - - PowerPoint PPT Presentation

Dec 03, 2022 231 likes •530 views

Pandora pattern recognition p for LArTPCs L. Escudero for the Pandora Team & the MicroBooNE and DUNE collaborations Calibration and Reconstruction for LAr TPC Detectors Thanks to John Marshall, Andy Blake and Steve Green for

slide-1
SLIDE 1

1

Pandora pattern recognition for LArTPCs

γ

p

µ γ

Calibration and Reconstruction for LAr TPC Detectors

Thanks to John Marshall, Andy Blake and Steve Green for contributing (voluntarily or involuntarily) to these slides

  • L. Escudero

for the Pandora Team

& the MicroBooNE and DUNE collaborations

slide-2
SLIDE 2

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

2

Pandora Pattern Recognition

A single clustering approach is unlikely to work for complex topologies (mix of track and shower-like clusters) Instead, the Pandora project is a novel method of pattern recognition, which tackles this project from its beginning in ILC and LHC using an advanced multi- algorithm approach:

  • Build up events gradually
  • Each step is incremental - aim not to

make mistakes (hard to undo)

  • Deploy more sophisticated algorithms

as picture develops

  • Build physics and detector knowledge

into algorithms

  • Possible thanks to the Pandora

Software Development Kit for Pattern Recognition (Eur. Phys. J. C 2015, 75: 439) for all use cases

John Marshall Andy Blake Mark Thomson

slide-3
SLIDE 3

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

3

Pattern Recognition

Low level Reconstruction Noise Filtering Signal Processing Hit Reconstruction Pattern Recognition Clustering 2D -> 3D Particle Hierarchy High level Reconstruction Track Fitting Calorimetry Particle ID Our input: collection of 2D hits Our output: hierarchy of reconstructed particles This is us! Reconstruction path to Physics

slide-4
SLIDE 4

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

4

Pattern Recognition

Pattern Recognition Clustering 2D -> 3D Particle Hierarchy This is us!

Building up events gradually, with chains of small algorithms, harnessing a number of powerful capabilities:

  • 120+ algorithms and

tools

  • Use of multiple parallel

hypothesis

  • Iterative reconstruction

techniques

  • Allowing incorporation
  • f ML/DL methods to

make algorithm decisions

slide-5
SLIDE 5

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

5

Pandora LAr TPC Pattern Recognition

LArTPC reconstruction: high quality, complex images, and lengthy drift times (i.e. long exposures) meaning a significant cosmic-ray background for surface detectors.

  • Our aim is to provide

automated pattern recognition for general usage (any particle, any topology)

  • Pandora algorithms are

reusable for different single- phase LArTPCs

  • Focus in recent years mainly on

MicroBooNE and now protoDUNE, but expanded to also SBND, ICARUS, DUNE FD

slide-6
SLIDE 6

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

6

Two different chains with tens of algorithms each:

  • Cosmic ray reconstruction (Pandora Cosmic)
  • Neutrino interaction/test beam reconstruction (Pandora Neutrino)

Output

Andy Smith

Pandora Consolidated Output

Input

*using SVM/BDT trained models

Harnessing the chains of algorithms in an intelligent manner to provide a consolidated output:

slide-7
SLIDE 7

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

7

Pandora LAr TPC Pattern Recognition

A little bit deeper into some of the steps in the algorithm chains…

2D track reconstruction

a) We start by producing a list of 2D clusters (per plane) that represent continuos, unambiguous lines of hits, starting/stopping at each branch or ambiguity. b) Then series of cluster-merging and cluster-splitting algorithms evaluate the list of 2D clusters and change them based on topological information, carefully aiming at safe merges, improving completeness without compromising purity.

Example of cluster-merging algorithm “in action”

Can’t do justice in a few slides, please find more in the Pandora MicroBooNE paper (Eur. Phys. J. C 78, p82 2018)

slide-8
SLIDE 8

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

8

Pandora LAr TPC Pattern Recognition

A little bit deeper into some of the steps in the algorithm chains…

Cluster-matching ambiguities are identified by tools and used to “diagonalise” the tensor Tools modify 2D clusters as appropriate and then run again from the beginning on the updated tensor. Example tool: 2D clusters in the three planes are compared to find those representing same particle, exploiting the common drift-time coordinate and our understanding of wire plane

  • geometry. Results are stored for each 3x2D

combination of clusters in a rank-three tensor.

3D track reconstruction

The Pandora Rotational Coordinate Transformation System is a very powerful feature, allowing 2D->3D and 3D->2D projections

slide-9
SLIDE 9

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

9

Pandora LAr TPC Pattern Recognition

Neutrino Interaction Vertex

Downstream usage:

  • Split 2D clusters at projected vertex position, and use

vertex to protect primary particles This allows us to have good performance in interactions with many final state particles…

Jack Anthony

A little bit deeper into some of the steps in the algorithm chains…

A key algorithm is the one to select the most appropriate 3D vertex position from a list

  • f candidate vertices. Used first a simple score, then a more sophisticated one (with

topological and charge asymmetry information) for each candidate. Now a multivariate approach (SVMs) is used for MicroBooNE. This is an example of how powerful the multi-algorithm approach is, by breaking down pattern recognition into small problems, even allowing to use Machine and Deep Learning methods to solve some of them!

Raquel Castillo NuInt talk

slide-10
SLIDE 10

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

10

Pandora LAr TPC Pattern Recognition

Shower Reconstruction

3D - Reuse ideas from the track tensor, using envelopes and their projection to match 2D shower clusters 2D - A key step in the pattern recognition is to characterise 2D clusters as track/shower like using topological information, to identify shower spines and allow to grow branches (nearby shower-like clusters), whereas prevent doing so around track-like clusters.

A little bit deeper into some of the steps in the algorithm chains…

slide-11
SLIDE 11

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

11

Pandora LAr TPC Pattern Recognition

3D hit/cluster reconstruction

For each 2D hit, sample clusters in

  • ther views at the same X, find uin, vin,

win and use analytic expression to find the most consistent 3D space point by minimising a χ2 function. Iteratively, produce smooth trajectory

Y X Z Y X Z

Finally, walking backwards from interaction vertex, use 3D clusters to

  • rganise

particles into hierarchies (building parent- daughter links)

and 3D particle hierarchy 2D/3D Particle refinement

Several algorithms deal with remnants to improve particle completeness, (esp. sparse showers). Sliding linear fits are used to define 2D envelopes and 3D cones for picking up small clusters/fragments.

slide-12
SLIDE 12

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

12

Pandora Performance Metrics

Efficiency: Fraction of target MCParticles (reconstructable*) with at least

  • ne matched reconstructed

particle (fulfilling conditions based on matched hits)

* see more precise definitions in paper

Correct event: Really strict metrics, events only correct if each MC target (reconstructable*) particle is matched to exactly one reconstructed particle, with correct parent-daughter links.

Event1: correct Event2: incorrect *labels correspond to true MC particle type

slide-13
SLIDE 13

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

13

Pandora Performance in MicroBooNE

The Pandora multi-algorithm approach to automated pattern recognition of cosmic-ray muon and neutrino events in the MicroBooNE detector (Eur. Phys. J. C 78, p82 2018)

Performance on selection of Exclusive Final States

“Correct” here really strict, means ~perfect event!

CC RES: μ+p+π+

slide-14
SLIDE 14

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

14

Pandora in Analysis & Detector Physics

Aleena Rafique

Comparison of νμ −Ar multiplicity distributions

  • bserved by MicroBooNE

to GENIE model predictions (paper)

Multiplicity 2, w/o particle ID

Detector calibration using through going and stopping muons in the MicroBooNE LArTPC (public note)

These are just some examples, see more in MicroBooNE public docs web

Pandora reconstruction is used in multiple MicroBooNE analyses, and also in calibration and detector physics studies (calorimetry, SCE, lifetime, diffusion…)

Colton Hill

Automated Selection

  • f νe from NuMI

using MicroBooE (NuInt talk)

slide-15
SLIDE 15

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

15

Exciting times for protoDUNE!

Pandora reconstruction

  • n protoDUNE

real data!

Steven Green

Very efficient reconstruction in terms of memory and CPU time (tipically <1 minute for protoDUNE data events)

slide-16
SLIDE 16

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

16

Exciting times for protoDUNE!

  • Using beam trigger information (trigger recob::Track providing position and

direction) performance on real data can be explored

  • Efficiency: Fraction of target events where the trigger is active and we

reconstruct at least one test beam particle

  • This folds in: cosmic vs test beam separation, reconstruction, tagging
  • Main reason for inefficiency is cosmic contamination, resulting in test beam

hierarchy reconstructed as cosmic ray protoDUNE SP DATA protoDUNE SP simulation

Note: Total number of hits (U+V+W)

slide-17
SLIDE 17

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

17

And getting ready for DUNE FD

So stay tuned!

Lorena Escudero

Preliminary performance on DUNE FD (1x2x6) simulation

NC DIS CC DIS ALL BUT DIS ALL BUT DIS

Note: Total number of hits (U+V+W)

x [drift] w [wire]

Pandora LArTPC algorithms designed to be reusable. Good performance already achieved using MicroBooNE algorithms in DUNE FD - specific tuning will follow

slide-18
SLIDE 18

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

18

Summary & Plans

  • The Pandora multi-algorithm approach enables us to build

up reconstructed images of very complex interactions

  • Pandora uses both pure 2D and pure 3D approaches, in

addition to algorithms where multiple 2D points are used to find test 3D positions, or 3D positions are projected into 2D

  • Machine Learning methods can be incorporated in this

approach to drive algorithm decisions.

  • Full particle hierarchies are delivered, with tagging of

cosmic-ray muons and neutrinos/test beam particles providing a consolidated output

  • Pandora delivers necessary output for physics analysis that

really exploit imaging capabilities of LArTPCs! We are actively working on the SBN program (mainly MicroBooNE) and protoDUNE and DUNE, and expecting to grow the team in 2019!

slide-19
SLIDE 19

Thanks!

19

slide-20
SLIDE 20

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

20

Pandora Slices

Pandora uses the concept of 3D slice internally since its (LAr) beginning: ○ They represent topologically distinct collection of hits (grouped by proximity and pointing info) ○ They become a candidate neutrino or beam-particle interaction in the pattern recognition ○ They are produced after the unambiguous cosmic-rays have already been identified

slide-21
SLIDE 21

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

21

Pandora Performance Metrics

Example DIS event

slide-22
SLIDE 22

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

22

ProtoDUNE Input Information

slide-23
SLIDE 23

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

23

ProtoDUNE Trigger Information

slide-24
SLIDE 24

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

24

ProtoDUNE data performance

slide-25
SLIDE 25

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

25

Pandora LAr TPC Pattern Recognition

Stitching

A little bit deeper into some of the steps in the algorithm chains…

In detectors with multiple drift volumes like protoDUNE, Pandora can determine the true particle time if it crosses an enclosed cathode (or anode) plane. By shifting pairs

  • f reconstructed particles in different drift volumes an equal amount in drift time,

cosmic rays (with a different T0 to the target 𝜉/TB) can be identified.

slide-26
SLIDE 26

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

26

Exciting times for protoDUNE!

  • Pandora shows a good performance

in test beam reconstruction in the assessment studies using simulation

  • Main causes of inefficiency are

contamination from cosmic rays and halo

  • There is also a small inefficiency from

Pandora slicing and test beam ID

Pandora test beam reconstruction performance assessment

slide-27
SLIDE 27

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

27

BDT for protoDUNE test beam ID

slide-28
SLIDE 28

Lorena Escudero, Calibration and Reconstruction for LArTPCs Detectors

28

On the one hand…

Deal with low energy, sparse showers, small particles. Also missing or poor information in one or more views

CCRES w/ π0 γ2

x [drift] w [wire]

MicroBooNE simulation CCRES w/ π0

LAr TPC Challenges (IMHO)

Pandora tuning for MicroBooNE proves to be efficient for reconstructing low energy showers. In addition, gaps treatment added to handle unresponsive detector regions in the pattern recognition (effective

  • verlap)

On the other hand…

Overlap of multiple particles commonly occurs in high energetic DIS events like the one in previous page, affecting not only pattern recognition but also physics measurements such as dE/dx