PaperlayerView exact revision on arXiv

Source-linked AI summary

The Pandora multi-algorithm approach to automated pattern recognition of cosmic-ray muon and neutrino events in the MicroBooNE detector

MicroBooNE collaboration, R. Acciarri, C. Adams, R. An, J. Anthony, J. Asaadi, M. Auger, L. Bagby, S. Balasubramanian, B. Baller, C. Barnes, G. Barr, M. Bass, F. Bay, M. Bishai, A. Blake, T. Bolton, L. Camilleri, D. Caratelli, B. Carls, R. Castillo Fernandez, F. Cavanna, H. Chen, E. Church, D. Cianci, E. Cohen, G. H. Collin, J. M. Conrad, M. Convery, J. I. Crespo-Anadon, M. Del Tutto, D. Devitt, S. Dytman, B. Eberly, A. Ereditato, L. Escudero Sanchez, J. Esquivel, A. A. Fadeeva, B. T. Fleming, W. Foreman, A. P. Furmanski, D. Garcia-Gomez, G. T. Garvey, V. Genty, D. Goeldi, S. Gollapinni, N. Graf, E. Gramellini, H. Greenlee, R. Grosso, R. Guenette, A. Hackenburg, P. Hamilton, O. Hen, V Hewes, C. Hill, J. Ho, G. Horton-Smith, A. Hourlier, E. -C. Huang, C. James, J. Jan de Vries, C. -M. Jen, L. Jiang, R. A. Johnson, J. Joshi, H. Jostlein, D. Kaleko, G. Karagiorgi, W. Ketchum, B. Kirby, M. Kirby, T. Kobilarcik, I. Kreslo, A. Laube, Y. Li, A. Lister, B. R. Littlejohn, S. Lockwitz, D. Lorca, W. C. Louis, M. Luethi, B. Lundberg, X. Luo, A. Marchionni, C. Mariani, J. Marshall, D. A. Martinez Caicedo, V. Meddage, T. Miceli, G. B. Mills, J. Moon, M. Mooney, C. D. Moore, J. Mousseau, R. Murrells, D. Naples, P. Nienaber, J. Nowak, O. Palamara, V. Paolone, V. Papavassiliou, S. F. Pate, Z. Pavlovic, E. Piasetzky, D. Porzio, G. Pulliam, X. Qian, J. L. Raaf, A. Rafique, L. Rochester, C. Rudolf von Rohr, B. Russell, D. W. Schmitz, A. Schukraft, W. Seligman, M. H. Shaevitz, J. Sinclair, A. Smith, E. L. Snider, M. Soderberg, S. Soldner-Rembold, S. R. Soleti, P. Spentzouris, J. Spitz, J. St. John, T. Strauss, A. M. Szelc, N. Tagg, K. Terao, M. Thomson, M. Toups, Y. -T. Tsai, S. Tufanli, T. Usher, W. Van De Pontseele, R. G. Van de Water, B. Viren, M. Weber, D. A. Wickremasinghe, S. Wolbers, T. Wongjirad, K. Woodruff, T. Yang, L. Yates, G. P. Zeller, J. Zennamo, C. Zhang

arXiv:1708.03135v1hep-exphysics.data-an

TL;DR

LArTPC imaging requires automated pattern recognition capable of separating particle energy deposits, especially amid MicroBooNE’s substantial cosmic-ray background. This paper develops a multi-algorithm Pandora reconstruction for cosmic-ray and neutrino events and evaluates it on simulated final states, finding measurable performance degradation from cosmic-ray remnants while retaining functional reconstruction.

  • Problem

    LArTPC reconstruction requires automated identification of particle energy deposits, while MicroBooNE’s long drift exposure produces substantial cosmic-ray background.

  • Method

    The paper applies over one hundred decoupled Pandora algorithms and tools, each targeting specific event topologies, to reconstruct cosmic-ray muon and neutrino interactions.

  • Results

    86.0% of selected quasi-elastic CC events are reconstructed correctly, while cosmic-ray remnants cause 5.1%–13.7% degradation across specified event topologies.

  • Takeaways & Limitations

    The Pandora approach provides fully automated reconstruction that remains functional for MicroBooNE events despite substantial cosmic-ray backgrounds.

  • Takeaways & Limitations

    The pattern recognition does not yet exploit dE/dx information to resolve collinear muons and protons, and cosmic-ray backgrounds degrade performance.

Abstract

from arXiv · show

The development and operation of Liquid-Argon Time-Projection Chambers for neutrino physics has created a need for new approaches to pattern recognition in order to fully exploit the imaging capabilities offered by this technology. Whereas the human brain can excel at identifying features in the recorded events, it is a significant challenge to develop an automated, algorithmic solution. The Pandora Software Development Kit provides functionality to aid the design and implementation of pattern-recognition algorithms. It promotes the use of a multi-algorithm approach to pattern recognition, in which individual algorithms each address a specific task in a particular topology. Many tens of algorithms then carefully build up a picture of the event and, together, provide a robust automated pattern-recognition solution. This paper describes details of the chain of over one hundred Pandora algorithms and tools used to reconstruct cosmic-ray muon and neutrino events in the MicroBooNE detector. Metrics that assess the current pattern-recognition performance are presented for simulated MicroBooNE events, using a selection of final-state event topologies.

1 Introduction

MicroBooNE uses Pandora to develop and evaluate automated pattern recognition for LArTPC events, addressing the challenge of identifying particle energy deposits. The paper focuses on a multi-algorithm reconstruction strategy applied to cosmic-ray and neutrino interactions.

  • MicroBooNE provides a testbed for automated LArTPC reconstruction using real data and supports neutrino cross-section and low-energy-event studies.
  • Pandora addresses particle energy-deposit identification with many decoupled algorithms, each targeting a specific event topology.
  • The paper applies generic Pandora algorithms specifically to MicroBooNE and evaluates selected final-state topologies for neutrino reconstruction.

2 The MicroBooNE detector

MicroBooNE is a single-phase LArTPC whose imaging depends on ionisation signals recorded across three wire planes. Reconstruction is challenged by cosmic-ray backgrounds, long drift exposure, and detector imperfections.

  • The detector’s surface location and drift times of up to a few milliseconds create substantial cosmic-ray muon contamination in neutrino events.
  • Partially correlated noise, unresponsive channels, and imperfect input hits can affect fine-detail pattern recognition.

3 Inputs and outputs

LArPandora translates MicroBooNE event data into Pandora objects, runs configured algorithms, and translates reconstructed particles back into the LArSoft event record. The output organizes tracks and showers into particle-flow hierarchies with associated clusters, space points, and vertices.

  • LArPandora bridges the LArSoft and Pandora event-data models while initiating algorithms and returning reconstructed pattern-recognition results.
  • Initialization registers geometry, algorithm, tool, and plugin components and uses PandoraSettings to define each event’s algorithm sequence and configuration.
  • Per event, the module translates hits, processes them with Pandora, extracts reconstructed particles, writes them to the event record, and resets the instance.
  • PFParticles represent distinct tracks or showers and link 2D clusters, 3D SpacePoints, vertices, and parent-daughter particle-flow relationships.

4 Algorithm overview

MicroBooNE uses separate PandoraCosmic and PandoraNu reconstruction paths, with shared algorithms but different topology priorities. Cosmic candidates are identified first, then their associated hits are removed before neutrino reconstruction.

  • PandoraCosmic is track-oriented for cosmic-ray muons and delta rays, whereas PandoraNu identifies a neutrino vertex and reconstructs emerging tracks and showers.
  • PandoraCosmic and PandoraNu process MicroBooNE data in two passes, with cosmic-ray tagging between them and a cosmic-removed hit collection feeding PandoraNu.

4.1 Cosmic-ray muon reconstruction

PandoraCosmic reconstructs cosmic-ray muons through staged 2D clustering, topology-based refinement, and 3D matching across the three readout views. Its tools resolve ambiguities by modifying cluster groupings, splitting or merging clusters, and recovering missing segments.

  • Two-dimensional reconstruction: PandoraCosmic separates hits by readout plane and creates initial 2D clusters representing continuous, unambiguous hit lines.The EventPreparation and TrackClusterCreation algorithms provide the inputs for subsequent topological processing.
  • Two-dimensional reconstruction: Cluster-merging algorithms improve completeness, while cluster-splitting algorithms improve purity when topology indicates multiple particles or discontinuities.Merging considers proximity and pointing relationships in the full event context; splitting responds to direction changes, intersections, or converging clusters.
  • Three-dimensional track reconstruction: 3D track reconstruction stores compatibility information for every u-v-w cluster combination in a rank-three tensor and uses it to identify matching ambiguities.The tensor also records connections and cluster reuse across multiple combinations, enabling event-level ambiguity resolution.
  • Three-dimensional track reconstruction: ClearTracks first forms particles from unambiguous three-view groupings, requiring common x-overlap above 90% of each cluster’s x-extent.Subsequent tools address more complex configurations, including delta-ray ambiguity and 1:2:2 cluster matches.
  • Three-dimensional track reconstruction: Additional tools split overshot clusters, merge undershot fragments, recover missing track segments, and add missing hits when view matching is otherwise unambiguous.After 3D reconstruction, unused clusters are reclustered and topologically refined as likely delta-ray fragments.

4.2 Neutrino reconstruction

PandoraNu isolates neutrino interactions from residual cosmic-ray activity, selects a three-dimensional interaction vertex, and reconstructs track-like and electromagnetic shower-like particles. Vertex scoring and iterative cross-view reconstruction organize hits into neutrino-centered particle hierarchies.

  • Neutrino reconstruction: PandoraNu divides reconstructed 3D hits into slices using proximity and direction metrics to isolate neutrino interactions from residual cosmic-ray remnants.Each slice ultimately produces one reconstructed neutrino particle with a hierarchy of daughter particles.
  • Three-dimensional vertex reconstruction: The neutrino path identifies a vertex from candidate positions generated by comparing overlapping 2D clusters across different readout planes.Candidate positions are produced by comparing cluster endpoints and evaluating cluster positions at shared x coordinates.
  • Three-dimensional vertex reconstruction: Candidates must lie on or near hits or registered detector gaps in all three views, after which the highest-scoring candidate is selected.The EnergyKickVertexSelection score combines energy-kick, asymmetry, and beam-deweighting components.
  • Track and shower reconstruction: Track-like and shower-like clusters are distinguished using geometric and vertex-related features, with long shower-like clusters serving as spines for recursive branch addition.ShowerGrowing records branch-to-spine association strengths and makes additions in the context of the overall event topology.
  • Track and shower reconstruction: ThreeDShowers matches shower-like clusters across readout planes using overlap and relationship information, iteratively altering 2D reconstruction to remove ambiguities.Fits characterize cluster extents, while a rank-three tensor supports cross-view matching and formation of 3D shower particles.

5 Performance metrics

The paper defines pattern-recognition performance by matching reconstructed particles to target MCParticles through shared hits, while excluding non-reconstructable topologies. Efficiency, completeness, purity, and event-level correctness provide complementary interpretations of reconstruction quality.

  • Performance metrics compare reconstructed particles with target MCParticles using shared hits assigned through the particle hierarchy.Visible particles are selected as targets, and downstream hits are folded into the relevant target MCParticle.
  • Hits associated with downstream particles of far-travelling neutrons or isolated, diffuse topologies are excluded when they are not reconstructable targets.The evaluation uses the MCParticle hierarchy to identify hits that should not characterize target particles.
  • Matches require at least five shared hits, at least 50% purity, and at least 10% completeness.These thresholds suppress weak associations, ambiguous assignments, and low-quality matches involving small fragment particles.
  • The matching procedure first assigns strongest one-to-one matches, then assigns remaining reconstructed particles to their highest-sharing target even if that target already has matches.Each target and reconstructed particle is initially matched at most once; later assignments can create multiple reconstructed particles for one target.
  • An event is correct only when exactly one reconstructed particle corresponds to each target MCParticle, making the correct-event fraction a sensitive performance measure.A target with no match after the one-to-one stage is considered lost and cannot gain a match during the final assignment stage.

6 Performance

Pandora reconstruction performance is evaluated across simulated MicroBooNE neutrino topologies using particle-level matching, efficiency, completeness, purity, and vertex-displacement metrics. Performance is generally strong for track-like particles but degrades for closely spaced or complex shower topologies, with identifiable merging, splitting, and vertex-placement failure modes.

  • CC quasi-elastic interactions: 95.8% of target muons and 87.3% of target protons are matched to exactly one reconstructed particle in the quasi-elastic topology, with 86.0% of events reconstructed correctly.Muon–proton merging is the predominant cause of missing matches, while some particles are split into multiple reconstructed particles.
  • CC quasi-elastic interactions: The proton reconstruction efficiency is lower than the muon efficiency across momenta, especially when the tracks are collinear and merged.dE/dx information might resolve collinear particles, but it is not yet exploited by the pattern recognition.
  • CC quasi-elastic interactions: 68% of quasi-elastic events have vertex displacement below 0.74 cm, while 10.4% exceed 5 cm, mainly because the vertex is placed at the wrong track end.Long tracks pointing back toward the beam source and decay-electron topologies can create competing vertex candidates.
  • CC resonance interactions with charged pion: 95.1% of target muons, 86.8% of target protons, and 80.9% of target pions yield a single reconstructed particle, while 70.5% of resonant-pion events are correct.Muon–pion merging and parent–daughter pion hierarchies account for important matching failures.
  • CC resonance interactions with charged pion: 68% of charged-pion resonance events have vertex displacement below 0.48 cm, while 7.3% exceed 5 cm.The pion track adds pointing information despite increasing event complexity.

7 Impact of cosmic-ray muon background

Cosmic-ray backgrounds substantially challenge MicroBooNE pattern recognition, while the slicing algorithm has topology-dependent effects. Performance remains functional but is lower for events containing sparse showers than for track-only final states.

  • Background conditions: A typical 20.6 cosmic-ray muons overlay each 3.2 ms simulated readout window, modeling the surface-based detector environment.The sample includes muons with at least 30 true hits.
  • Background conditions: The cosmic-ray removal procedure can degrade events when at least 10% of neutrino-induced hits are removed or final-state particle classification changes.Such events are identified as degraded before performance metrics are assessed.
  • Background conditions: Only 76% of cosmic-ray muons are tagged, allowing many remnants to enter PandoraNu and challenge neutrino pattern recognition.The tagging is described as conservative.
  • Slicing effects: For two-shower final states, slicing lowers the correct-event fraction by 5.8% on average across (µ +π0 +Np) final states.Sparse shower elements can be assigned to separate slices and reconstructed as separate neutrino candidates.
  • Slicing effects: Slicing does not degrade events with only track-like final-state particles, while cosmic-ray remnants cause additional degradation across all investigated event types.The slicing algorithm has little impact for track-only topologies, as reflected by overlapping configurations in three plots.
  • Overall performance: With cosmic-ray backgrounds, total degradation reaches 5.1% for CC quasi-elastic (µ +Np), 7.2% for CC resonance (µ + Np), 5.5% for (µ + π+ + Np), and 13.7% for (µ + π0 + Np) final states.Pattern recognition is typically deemed correct for 70% of track-only events and 35% of interactions with two sparse showers.

8 Concluding comments

The paper presents Pandora’s multi-algorithm approach for automated LArTPC reconstruction and evaluates it in simulated MicroBooNE events using strict particle-matching metrics. Cosmic-ray backgrounds pose substantial challenges, while future algorithms and refinements are expected to improve performance.

  • Approach: Pandora uses many decoupled algorithms to gradually reconstruct cosmic-ray muon and neutrino interactions in LArTPC detectors.The approach provides fully automated reconstruction.
  • Evaluation: The study evaluates simulated MicroBooNE events with strict metrics matching reconstructed particles to every true visible final-state particle.The evaluation provides a snapshot of current pattern-recognition performance.
  • Findings: Cosmic-ray backgrounds pose substantial challenges to pattern recognition in the surface-based MicroBooNE experiment.The conclusion identifies this background as a central performance challenge.
  • Outlook: The paper expects improvements from adding new algorithms and refining cosmic-ray muon removal.These improvements are presented as future development directions.
Loading 1708.03135v1…