The Pandora multi-algorithm approach to pattern recognition 16th September 2019 Reconstruction & Machine Learning in Neutrino Experiments workshop - DESY Andrew Smith - For the MicroBooNE & DUNE collaborations
Overview Hope to answer the following questions: ● Why is pattern-recognition important for neutrino physics? ○ Where does pattern-recognition fit into the field of reconstruction? ○ What is Pandora’s approach to pattern-recognition for Liquid Argon Time ○ Projection Chamber experiments? - MicroBooNE, ProtoDUNE & DUNE How can I get involved / learn more? ○ 2
LArTPC detectors LArTPC = Liquid Argon Time Projection Chamber Neutrino interacts with Ar atom ● Resultant particles ionize Ar along their trajectory ● and also produce scintillation light Drift time Ionization charge drifts towards wires and ● produces a signal on all three planes Charge ⊙ Wire number E-field Single-phase 3 ⨯ images LArTPC . . . . . . A neutrino interaction image from one wire plane in MicroBooNE s s e e n n Very high resolution calorimeter - millimeter-scale ● neutrino → Ar a a l l p p e e r r Can resolve individual particles down to low energies ● i i w w ⨯ ⨯ 3 3 3 ⨯ 2D views ⇒ 3D imaging ● Filled with MicroBooNE detector liquid argon JINST 12 (2017) no.02, 3 P02017
The MicroBooNE experiment Booster LINAC High rise protons → ← ~1 GeV neutrinos ICARUS MicroBooNE SBND Target hall 470m Experiment Collaboration of 179 scientists from 31 worldwide institutions ● Taking data since October 2015 ● ~10 5 neutrino interactions in this time ● Additional off-axis neutrinos from second beam: NuMI ● Physics ● Study previous anomalous results: excess of ν e at low energies ● Measure suite of precision ν on Ar cross sections ● Hardware and software R&D for future experiments such as DUNE 4
DUNE and ProtoDUNE DUNE = Deep Underground Neutrino Experiment Collaboration of over 1000 scientists from more than ● ● 30 countries Groundbreaking for DUNE long baseline ○ neutrino facility in 2017 Physics Neutrino oscillation - Do neutrinos violate CP? ● ProtoDUNE = prototype for DUNE ● Astroparticle physics, supernovae ● Based at CERN, first data 2018 ○ Proton-decay? ● Important test-beam data for DUNE ○ 5
From images to Physics The reconstruction pipeline Low-level image Pattern Particle fits: Calorimetric Particle Hit finding processing recognition Tracks, Showers reconstruction Identification 6
From images to Physics The reconstruction pipeline Low-level image Pattern Particle fits: Calorimetric Particle Hit finding processing recognition Tracks, Showers reconstruction Identification • Hits • • • Collection • Current out of wire Wire position Drift time • Induction • Drift time Induction The raw waveforms are processed to deconvolve detector effects and remove noise A hit is produced for peaks in the Cartoon wire responses to ionization processed waveforms. These form charge drifting past a wire on the input to the patrec stage induction and collection planes 7
From images to Physics The reconstruction pipeline Low-level image Pattern Particle fits: Calorimetric Particle Hit finding processing recognition Tracks, Showers reconstruction Identification • Hits • Secondary • vertex • Collection Clusters of hits • represent individual Current out of wire Wire position Drift time particles • Induction • Pattern Drift time Neutrino interaction recognition Wire position vertex Induction The main job of the patrec is to: Cluster the hits together to Drift time represent individual particles Cartoon wire responses to ionization Hits from the collection plane for a simulated neutrino Identify the hierarchical relationship charge drifting past a wire on interaction in MicroBooNE, before and after patrec between particles induction and collection planes 8 Eur.Phys.J. C78 (2018) no.1, 82
The Pandora project SDK Application Algorithms Pandora General purpose open-source framework for ● pattern recognition Initially used for future linear collider ● experiments, but now well established on many LArTPC experiments too! GitHub Software μBooNE Repository development kit Algorithms github.com/PandoraPFA Eur.Phys.J. C75 Eur.Phys.J. C78 (2015) no.9, 439 (2018) no.1, 82 9
The three readout planes The event data model give three “views” of the same interaction Cluster of hits We encapsulate our current understanding of the ● Neutrino interaction event via the event data model U plane vertex After the patrec is finished, these are the objects ● which are available in LArSoft for downstream Parent-Daughter link defines hierarchy analysis Wire position Reconstructed Hits particles are the principal output V plane 2D Clusters EM shower Interaction 3D Each colour Particles Vertices Trajectories represents a different particle Hierarchies W plane Drift time (common to all planes) 10 Event data model Understanding of the event A cartoon LArTPC pattern recognition problem
Pandora’s multi-algorithm approach Break the problem up into smaller well defined tasks and Iteration is used to allow 2-way ● ● develop targeted algorithms for each task information flow between algorithms E.g. Cluster together two hits if … ○ Algorithm A Algorithms update our current understanding of the Algorithm Information event by modifying the event flow data Information flow Algorithm B Algorithm complexity varies from simple cuts up to more ● advanced machine learning techniques Iteration provides powerful feedback ● loops - a technique that Pandora The application runs ~100 algorithms to gradually build our ● frequently utilizes understanding until a complete picture of the event develops 11
Pandora’s algorithms for neutrino interactions 2D clustering Vertex finding Matching clusters between views using SVMs U plane Vertex Candidate cluster split candidates V plane W plane Final result Electromagnetic showers Candidate shower Protected track Wire position branches clusters Candidate shower spines Drift time Neutrino interaction vertex Simulated unresponsive channels MicroBooNE simulation 12 Neutrino interaction vertex
Performance - case study For performance in other channels, see: Eur.Phys.J. C78 (2018) no.1, 82 Reconstruction efficiency Matched True MC Shared Reco reconstructed particle hits hits hits particle Purity = # Shared hits / # Reco hits Low energy particles produce few Completeness = # Shared hits / # MC hits hits - causing efficiency drop True momentum [GeV] Fraction of events Fraction of events 5cm Wire position Neutrino interaction vertex ↑ logarithmic-scale ↑ logarithmic-scale Drift time 13 Purity Completeness
Pandora’s algorithm chains Neutrino / test-beam chain Cosmic-ray chain As described on previous slides, algorithms are Optimised to reconstruct cosmic-ray muons ● ● designed for neutrino or test-beam interactions More strongly track-oriented than the neutrino / ● Identify the primary interaction vertex early in the test-beam chains ● patrec to inform later algorithms Includes algorithms to identify and reconstruct ● Includes special chains of algorithms for delta-rays of energetic cosmic-rays ● electromagnetic showers Each algorithm chain works well on the types of interactions it’s designed for For surface detectors, we need a way of dealing with events that contain both neutrino/test-beam interactions and cosmic-rays A Pandora-reconstructed cosmic-only data event in Solution: “Consolidated approach” MicroBooNE 14
Handling cosmic-rays Use two different algorithm chains to handle ● A Pandora-reconstructed neutrinos / test-beam interactions and cosmic-rays test-beam ProtoDUNE data event Cosmic chain Daughter particles: 1 Track, 4 Showers Tag clear CRs 3D slicing Neutrino / Cosmic chain test-beam chain Parent reconstructed particle Slice ID “Slice” is reconstructed using 15 the test-beam algorithm chain
Further information and tutorials 16
Papers and documentation The Pandora SDK paper ● Details the design of the software development kit and how algorithms interface with the ○ application that is running Pandora (e.g. larpandora) The Pandora MicroBooNE paper ● Gives details of Pandora’s algorithms in MicroBooNE at the time of publication, but generally ○ applicable to other LArTPC experiments too All Pandora code is self-documented using doxygen and is available on github ● https://github.com/PandoraPFA ○ 17
Recent workshops & hands-on exercises Multi-day Pandora workshop in Cambridge, UK - 2016 ● Talks about how the algorithms work and step-by-step exercises about how you might develop ○ a new algorithm using Pandora! LArSoft workshop in Fermilab - 2019 ● LArSoft workshop in Manchester, UK - 2018 ● Workshop on advanced computing & machine learning, Paraguay - 2018 ● Talks and exercises about running and using Pandora within LArSoft, including tutorials on using ○ Pandora’s custom event display Experiment specific resources: ● ProtoDUNE analysis workshop, CERN - 2019 ○ MicroBooNE Pandora workshop, Fermilab - 2018 ○ 18
Recommend
More recommend
