Project: Find Michels! (INSS 2025)

Welcome to the AI/ML project for INSS 2025. The challenge is simple:

Identify the location (i.e. pixels) of a Michel electron in 1 million images of particles.

Dataset

You will deal with a 3D image of particle trajectories that look like below.

There are two types of data:

1. "xyze" image data (left image above)

   • An array of the shape (N,4) for N 3D voxels. float32 type.
   • Each element contains (x,y,z,energy) information representing 3D location and amount of energy deposition by a particle.
   • The energy deposition is sampled regularly in space (i.e. "voxelized") in the 768x768x768 space. (x,y,z) represent the voxel index along each axis.
   • The deposited energy is in MeV scale.

2. "label" data (right image above)

   • An array of the shape (N) for N 3D voxels (corresponding to "xyze" data). int32 type.
   • Each element contains a semantic label as an integer value from 0 to 4 for the subject voxel.
     • 0 ... "shower"
     • 1 ... "track"
     • 2 ... Michel electron!
     • 3 ... "delta ray"
     • 4 ... scattered energy depositions

The xyze and label data are stored in separate files. The filenames should tell you which information is stored.

Files

There are 4 sets of files.

• A) 1M challenge
  • challenge_xyze_1e6.h5 (link, warning 65 GB)
• B) 10k challenge
  • challenge_xyze_1e4.h5 (link, warning 681 MB)
• C) 10k training
  • train_xyze_1e4.h5 (link, warning 669 MB)
  • train_label_1e4.h5 (link, warning 172 MB)
• D) Example 100 events
  • example_xyze_100.h5 (link, 6.4 MB)
  • example_label_100.h5 (link, 1.7 MB)
  • example_solution_100.h5 (link, 36 KB)

The 100 event files (D) is a subset of 10k training datasets.

To obtain these files:

• You can download from the link above
• You can copy from Kazu's local flash drive while he is at INSS Aug. 12 (Tue) - 15 (Fri).
• If you have S3DF computing account, they can be found under /sdf/data/neutrino/public_html/inss/
  • Kazu can help copying files to other sites (NERSC, Aurora, FNAL, etc.)

Accessing data in files

Take a look at this example notebook.

Challenge Task

• By the deadline, identify the location of Michel electron pixels in all images
• The goal is to analyze 1M images (file A).
  • However, if that not possible, you can also submit your solution based on 10k images (file B).
• Submit your solution in single HDF5 file format
  • The file should contain a variable length array of type int32, the length of the array should be the number of images.
  • The array elements should indicate the 3D voxel index of Michel electron pixels in the original N long array in the xyze data.
    • You can find an example solution format store in D) file, example_solution_100.h5 (link).
  • The solution file should be small and you can upload to this google dirive folder.

Bonus Task

• Write a toy MC simulation for Michel electron energy spectrum.
• Generate a simulated energy spectrum with high statistics (so it's smooth!)
• Analyze the 10k training dataset (file C) and draw the energy spectrum of Michel electrons.
• Compare the shape of two spectrums and comment on the diffrences.
• Propose potential source of discrepancies.

Performance Criteria

We will evaluate your solution based on the F1 score.

The results are separated for 1M images (file A) and 10k images (file B). The former demonstrates scalability over the latter, important for ever-growing neutrino detectors!

When you submit your solution, Kazu will try to get back to you as soon as possible. This means you can submit a solution for 10k sample dataset first to see how you perform (if you want to) before submitting a final solution for 1M dataset.

How you might proceed:

Use 100 example dataset (file D) to get a feeling for the challenge.

Use 10k training dataset (file C) with labels to develop your model and validate its generalizablility.