Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
  • Open Access

Deep neural network for pixel-level electromagnetic particle identification in the MicroBooNE liquid argon time projection chamber

C. Adams et al. (MicroBooNE Collaboration)
Phys. Rev. D 99, 092001 – Published 7 May 2019
PDFHTMLExport Citation
Abstract
Authors
Article Text
  • INTRODUCTION
  • MICROBOONE DETECTOR AND PARTICLE IMAGES
  • U-RESNET: TRACK/SHOWER PIXEL-LEVEL…
  • TRAINING U-RESNET
  • DATA SAMPLES FOR TRAINING AND VALIDATION
  • NETWORK PERFORMANCE ON SIMULATION…
  • NETWORK PERFORMANCE WITH DETECTOR DATA…
  • CONCLUSION
  • ACKNOWLEDGMENTS
  • APPENDICES
    • References

    Abstract

    We have developed a convolutional neural network that can make a pixel-level prediction of objects in image data recorded by a liquid argon time projection chamber (LArTPC) for the first time. We describe the network design, training techniques, and software tools developed to train this network. The goal of this work is to develop a complete deep neural network based data reconstruction chain for the MicroBooNE detector. We show the first demonstration of a network’s validity on real LArTPC data using MicroBooNE collection plane images. The demonstration is performed for stopping muon and a νμ charged-current neutral pion data samples.

    • Figure
    • Figure
    • Figure
    • Figure
    • Figure
    • Figure
    • Figure
    14 More
    • Received 23 August 2018

    DOI:https://doi.org/10.1103/PhysRevD.99.092001

    Published by the American Physical Society under the terms of the Creative Commons Attribution 4.0 International license. Further distribution of this work must maintain attribution to the author(s) and the published article’s title, journal citation, and DOI. Funded by SCOAP3.

    Published by the American Physical Society

    Physics Subject Headings (PhySH)

    1. Research Areas
    Particle interactions
    1. Techniques
    Artificial neural networksMachine learningNeutrino detection
    Particles & Fields

    Authors & Affiliations

    Click to Expand

    Article Text

    Click to Expand

    References

    Click to Expand
    Issue

    Vol. 99, Iss. 9 — 1 May 2019

    Reuse & Permissions
    Author publication services for translation and copyediting assistance advertisement

    Authorization Required

    Log In
    Other Options
    ×

    Images

    • Figure 1
      Figure 1

      Two examples of νe (a) and νμ (b) events simulated in the MicroBooNE detector. The pixel color represents the amount of energy deposited per pixel in the 2D projection image. The pattern of ionization differs significantly between νe and νμ.

      Reuse & Permissions
    • Figure 2
      Figure 2

      The MicroBooNE detector schematics showing three wire planes and example 2D projections of V and Y plane waveforms.

      Reuse & Permissions
    • Figure 3
      Figure 3

      U-ResNet architecture diagram. Black arrows describe the direction of tensor data flow. Red arrows indicate concatenation operations to combine the output of convolution layers from the encoding path to the decoding path. The final output has the same spatial dimension as the input with a depth of three, representing the background, track and shower probability of each pixel.

      Reuse & Permissions
    • Figure 4
      Figure 4

      Top: an example image from the training set in which two protons, one electron, and one muon are produced. The gaps along the trajectory of an electron and proton on the left are due to unresponsive wires [6] in the detector. Bottom: the event from the top image that shows PL weighting categories indicated in different colors.

      Reuse & Permissions
    • Figure 5
      Figure 5

      (a) The training loss value as a function of training time using the validation sample. The red line shows the average at a given Epoch computed using 200 the neighboring Epoch points. (b) ICPF for the same sample. The sudden drop in both figures at Epoch 14 is due to lowering of the learning rate by a factor of 10.

      Reuse & Permissions
    • Figure 6
      Figure 6

      The binned ICPF distribution over all images in the test set.

      Reuse & Permissions
    • Figure 7
      Figure 7

      The ICPF error rates for U-ResNet labeling track (proton) and shower (electron) pixels for the benchmark test set are plotted against initial kinematic variables from simulation information. (a) The opening angle between two particles from the 1e1p sample. (b) The electron kinetic energy from the 1e1p-LE sample. (c) The proton kinetic energy from the 1e1p-LE sample.

      Reuse & Permissions
    • Figure 8
      Figure 8

      (a) Binned distribution of ICPF for track and shower pixels where the pixel-level labels are produced by a physicist. The data and simulation distributions are area normalized and represent 100 Michel electron events. There is no event outside the shown range on the horizontal axis. (b) The normalized, binned softmax probability distributions for shower pixels on data and simulation. (c) The same as (b) for track pixels.

      Reuse & Permissions
    • Figure 9
      Figure 9

      Peak pixel value distribution for Michel electron images for data and simulation using the 3-pixel differentiation algorithm described in the text. The vertical axis shows the pixel counts while the horizontal axis shows the peak pixel values.

      Reuse & Permissions
    • Figure 10
      Figure 10

      The average ICPF in percent for the Michel electron data versus pixel scaling factor is shown in blue. The pixel-fraction disagreement between physicist and network categorizations is shown in percent in red for track and shower pixels separately.

      Reuse & Permissions
    • Figure 11
      Figure 11

      Three regions for an analysis of the interpixel correlation for a Michel in data. Region 1 contains the minimum ionizing muon trajectory. Region 2 focuses on the end of the muon trajectory where dE/dx increases. Region 3 contains a decay Michel electron.

      Reuse & Permissions
    • Figure 12
      Figure 12

      Minimum ionizing muon track in region 1 of Fig. 11 where all pixels in the image are masked except for a small portion of a muon track shown in the images above. The lower row shows normalized track vs shower score distributions for all non-zero pixels in the image.

      Reuse & Permissions
    • Figure 13
      Figure 13

      Stopping muon track in region 2 of Fig. 11 where all pixels in the image are masked except for the small portion of the muon track next to its stopping point shown in the images above. The lower row shows normalized track and shower score distributions for all nonzero pixels in the image.

      Reuse & Permissions
    • Figure 14
      Figure 14

      Michel electron in region 3 of Fig. 11 where all pixels in the image are masked except for a portion of an electron trajectory shown in the images above. The lower row shows normalized track and shower score distributions for all nonzero pixels in the image. (a) The entire Michel electron image is unmasked. (b) The initial portion of the Michel electron trajectory is masked. (c) Everything except the initial portion of the Michel electron trajectory is masked

      Reuse & Permissions
    • Figure 15
      Figure 15

      The intersection of regions 2 and 3 of Fig. 11 where all pixels in the image are masked except for a small portion of a Michel electron and the Bragg peak from a stopping muon. The middle histogram shows a normalized track and shower score distribution for all nonzero pixels in the image. The right figure shows the normalized score value of the classified pixel category.

      Reuse & Permissions
    • Figure 16
      Figure 16

      (a) The binned distribution of ICPF where the pixel-level labels are produced by a physicist. The data (black) and simulation (red) distributions are area normalized, produced from 100 CCπ0 events. There is no event outside the shown range on the horizontal axis. (b) The normalized, binned softmax probability distributions for shower pixels by the network on data and simulation. (c) The same as (b) for track pixels.

      Reuse & Permissions
    • Figure 17
      Figure 17

      The ICPF mean in percent for CCπ0 data (blue) for varying pixel scaling factor shown on the horizontal axis. A category-wise physicist-network disagreement pixel fraction in percent is shown in red for track and shower pixels separately.

      Reuse & Permissions
    • Figure 18
      Figure 18

      Four example CCπ0 events with highest ICPF values using physicist generated pixel labels. Left: input images to the network. Middle: track (yellow) and shower (cyan) using physicist generated labels. Right: track (yellow) and shower (cyan) labels predicted by the network.

      Reuse & Permissions
    • Figure 19
      Figure 19

      Event displays of νe (upper two rows) and νμ interactions (lower two rows). The left column images are inputs to the network. The middle column shows labeled images based on simulation information. Track pixels are masked in yellow and shower pixels are in cyan. The right column shows the output of the network.

      Reuse & Permissions
    • Figure 20
      Figure 20

      Michel electron event displays from real detector data. Left: input images to the U-ResNet. Middle: track (yellow) and shower (cyan) physicist labels. Right: track (yellow) and shower (cyan) labels predicted by the network.

      Reuse & Permissions
    • Figure 21
      Figure 21

      Neutrino event displays from CCπ0 candidate detector data selected based on activity around the interaction vertex. Left: input images to the network. Middle: track (yellow) and shower (cyan) physicist labels. Right: track (yellow) and shower (cyan) labels predicted by the network.

      Reuse & Permissions
    ×

    Sign up to receive regular email alerts from Physical Review D

    Reuse & Permissions

    It is not necessary to obtain permission to reuse this article or its components as it is available under the terms of the Creative Commons Attribution 4.0 International license. This license permits unrestricted use, distribution, and reproduction in any medium, provided attribution to the author(s) and the published article's title, journal citation, and DOI are maintained. Please note that some figures may have been included with permission from other third parties. It is your responsibility to obtain the proper permission from the rights holder directly for these figures.

    ×

    Log In

    Cancel
    ×

    Search

    Article Lookup

    Paste a citation or DOI
    Enter a citation
    ×