Deep Learning from MAGIC Raw Data

Jarred Green

MAGIC SOBO Meeting — 30 June 2025

Max Planck Institute for Physics

jgreen@mpp.mpg.de

Outline

Current Reconstruction Method
Graphs
Current Methods
ICRC and Next Steps

Jarred Green - jgreen@mpp.mpg.de

A way to include as much information as possible?

1. Current Reconstruction

Particle Type
Energy
Arrival Direction

Jarred Green - jgreen@mpp.mpg.de

Proton

Gamma

Muon

Find a new ways to represent raw data

1. Current Reconstruction

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

Lets go from images

to graphs

Credit: J. Leskovec (Stanford CS224)

for MAGIC?

GNNs can effectively deal with

irregular camera geometry
sparse data
non gridlike pixels
arrays with different cameras (ie CTA)
information on each node (pixel) and edge (connection)

Jarred Green - jgreen@mpp.mpg.de

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

time

Same event with a simple threshold in p.e.

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

time

Real data event

Can we 'steal' this timing information from OFF data?
YES!

"time calibration pasting"

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

Can we 'steal' this timing information from OFF data?
YES!

"time calibration pasting"

✅ Contains all signal and background information
❌ Huge graphs with 100K+ nodes
❌ Training ~days

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

time

Can we make these graphs smaller somehow?

cleaning!

✅ Small graphs with 200-2K nodes
✅ Training ~hours
❌ Maybe loosing some signal information

2. Graphs

an IceCube case study

Created their own GNN framework for Neutrino detection
15-20% improvement in reconstruction of energy, zenith, direction at low energies
Can do realtime reconstruction onsite with 1 GPU

Credit: R. Abbasi et al. [arXiv:2209.03042]

Jarred Green - jgreen@mpp.mpg.de

2. Graphs

Jarred Green - jgreen@mpp.mpg.de

Goal: get calibrated data with timing information

MARS - sorcerer

Python package magic-gnn

.parquet or LMDB files exported, about ~1GB / min of data

stereo event matching

Transformers with
cleaned graphs

GNNs on cleaned and uncleaned graphs

Calibration

MExportCSV

MExportParquet

data standardization

time calibration pasting

Python package magic-gnn

✅

✅

2️⃣

1️⃣

3. Current Methods

Jarred Green - jgreen@mpp.mpg.de

Training Data: 150K MC events, za05to35, ST0307

GNNs with cleaned graphs

Real Data: ST0307 Triangulum II (OFF)

Training with GraphNet (IceCube)
Dynedge architecture
Telescopes talk with cross-attention
Outputs 'gammaness'

3. Classification

1️⃣

Jarred Green - jgreen@mpp.mpg.de

Training Data: 2M gamma events, za05to35, ST0307

Transformers on
MAGIC graphs

Real Data: ST0307 Triangulum II (OFF)

Similar idea to ChatGPT
Pixels in time are treated like 'sentences'
Keeping a smaller GNN built into the model

3. Direction

Loss: Angular error plus FoV error
No other feature extraction or physics information

2️⃣

Training now

- Running since ~12hrs at MPP

- Current 68% error at ~1º

4. What to show at ICRC

Focus on methods only
No need for performance plots
- Results will continue to improve over the next month(s)
Save results for publication and TeVPA
Preliminary Theta2 of Crab Nebula?

Jarred Green - jgreen@mpp.mpg.de

Nudge direction reconstruction model towards DISP methods
- Provide with global features or tuned physics-informed models/losses
Test on Crab Nebula sample
- Need drive reports?
Try both architectures with Energy reconstruction
Foundation model (all tasks together)
Collaborate with IceCube and ML experts at TUM