How Neutrino Astronomy May Inform the Next-Generation of IACT Event Reconstruction Techniques

Jarred Green

IFAE, 28 October 2025

Max Planck Institute for Physics

jgreen@mpp.mpg.de

The Data
The Challenge
Some New Ideas
Neutrinos?
ML Playground

Outline

Jarred Green - jgreen@mpp.mpg.de

γ ray

Particle Type
Energy
Arrival Direction

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

}

Dimensionality Reduction

Jarred Green - jgreen@mpp.mpg.de

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

One Pixel, ~30ns

Time [ns]

Counts

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

{
	'size': 546.88637,
	'xc': -3.78983,
	'yc': 7.29427,
	'length': 7.76769,
	'width': 1.76480,
	'delta': -0.56875,
	 ...
}

"Hillas Parameters" (1977!)

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

{
	'size': 546.88637,
	'xc': -3.78983,
	'yc': 7.29427,
	'length': 7.76769,
	'width': 1.76480,
	'delta': -0.56875,
	 ...
}

Random Forests

Jarred Green - jgreen@mpp.mpg.de

1. The Data

0. Raw data

1. Calibration

2. Cleaning

3. Parameterization

4. ML

5. Science

Particle Type
Energy
Arrival Direction

Detections

Skymaps

Spectra

Jarred Green - jgreen@mpp.mpg.de

1. The Data

1039 \ \text{pixels} \\ \times 50 \ \text{time slices} \\ \times 2 \ \text{telescopes} \\ = 103900 \ \text{params}

{
	'size': 546.88637,
	'xc': -3.78983,
	'yc': 7.29427,
	'length': 7.76769,
	'width': 1.76480,
	'delta': -0.56875,
	 ...
}

Particle Type
Energy
Arrival Direction

Dimensionality

reduction

tl;dr

Jarred Green - jgreen@mpp.mpg.de

1. The Data

Proton

Gamma

Muon

103900 \ \text{params per event (raw)}

\sim 300 \frac{\text{triggers}}{\text{s}}

Jarred Green - jgreen@mpp.mpg.de

\text{S/N ratio} \sim \frac{1}{10000}

2. The Challenge

How do we go from shower images back to the original properties of the particle?

2. The Challenge

How do we go from shower images back to the original properties of the particle?

The current methods work extremely well!

3. Some New Ideas

{
	'size': 546.88637,
	'xc': -3.78983,
	'yc': 7.29427,
	'length': 7.76769,
	'width': 1.76480,
	'delta': -0.56875,
	 ...
}

Can we work directly with the images?

Convolutional Neural Networks

We want to extract new features from the images in a way that encodes spatial information as well

Source: Ujjwal Karn

A type of NN which specializes in computer vision

Convolutional Neural Networks

A type of NN which specializes in computer vision

We want to extract new features from the images in a way that encodes spatial information as well

Source: Ujjwal Karn

✨ Dimensionality Reduction

Example image

A cute 3x3 matrix

the convolution of the two

Convolutions

Act like filters on an image!

CNNs in gamma-ray astronomy

Case Study:

CNNs with MAGIC

Hexagonal pixels

to square

Source: Miener et al 2022, Jacquemont et al 2019

CNNs in gamma-ray astronomy

Case Study:

CNNs with the
Large-Sized Telescope

Source: Vuillaume et al 2021

Now we can work directly with hexagonal pixels

4. Neutrinos?

4. Neutrinos

Particle Type
Energy
Arrival Direction

👀

Some things to notice:

3D in space
Time resolution
Not all 'nodes' are always active (sparse)
Nodes are irregularly spaced

4. Neutrinos

Some things to notice:

3D in space
Time resolution
Not all 'nodes' are always active (sparse)
Nodes are irregularly spaced

Graph representation!

4. Neutrinos

Graphs

Images can be recontextualized as graphs

nodes and edges in a regular grid structure

Graphs

But graphs don't have to conform to a particular shape

Naturally handle irregular geometrics
Can do convolution-like operations on graphs too!

A GNN framework for Neutrino observatories
Easy to adapt to other detectors

With Graph Neural Networks, IceCube can improve their reconstruction metrics by 13-20%

Source: IceCube 2022

4. Neutrinos

A GNN framework for Neutrino observatories
Easy to adapt to other detectors

With Graph Neural Networks, IceCube can process events in realtime at nearly 300 Hz on a single GPU

Source: IceCube 2022

4. Neutrinos

5. ML Playground

Jarred Green - jgreen@mpp.mpg.de

irregular camera geometry
sparse data
non gridlike pixels
arrays with different camera types

5. ML Playground

Jarred Green - jgreen@mpp.mpg.de

Both telescopes can be embedded in the same single graph

5. ML Playground

Jarred Green - jgreen@mpp.mpg.de

Particle Classification with

Model:

Binary classification
DynEdge with 12 NN
Separate each telescope
Cross-attention
5-10% of training data

DynEdge

Confidence of gamma classification

Gammas

Protons

PRELIMINARY

5. ML Playground

Jarred Green - jgreen@mpp.mpg.de

Direction Reconstruction with

Model:

2nd place winner of IceCube Kaggle Competition
Transformer + DynEdge
Very robust model for neutrino direction reconstruction

DeepIce

Radial Error [deg]

Cumulative Fraction

MC 68% angular offset

<0.08 deg

PRELIMINARY

5. ML Playground

Jarred Green - jgreen@mpp.mpg.de

Coming soon

Energy reconstruction results
Full comparison of MC performance with standard pipeline
Detection with real data 🤠

MAGIC Raw Data

Contains full time evolution of events
Saves on data processing
Are naturally represented as graphs

Graph Neural Networks

Apply domain knowledge from IceCube
Basic neutrino models work on gamma-ray data
Robust performance with minimal modifications

Evidence for the benefits
of cross-domain collaboration

Jarred Green - jgreen@mpp.mpg.de

Future Motivation

70+ Telescopes!!

Cherenkov Telescope Array

Jarred Green - jgreen@mpp.mpg.de

Thank you!

Jarred Green - jgreen@mpp.mpg.de

Credit: G. Ceribella