Waveform MAGIC Telescope Data With Graph Neural Networks
Jarred Green
Geneva, ICRC 2025
20 August 2025
Max Planck Institute for Physics
jgreen@mpp.mpg.de
Outline
- Current Reconstruction Method
- Graphs
- The Data
- ML Playground
- Conclusions
Jarred Green - jgreen@mpp.mpg.de
1. Current Reconstruction
Jarred Green - jgreen@mpp.mpg.de
{
'size': 546.8863769783503,
'xc': -3.7898364035228065,
'yc': 7.294274369102854,
'length': 7.767692666363586,
'width': 1.7648053936769377,
'delta': -0.5687518727314761,
...
}tl;dr
- Particle Type
- Energy
- Arrival Direction
ML
1. Current Reconstruction
1039 \ \text{pixels} \\ \times 50 \ \text{time slices} \\ \times 2 \ \text{telescopes} \\ = 103900 \ \text{data points}
tl;dr
1. Current Reconstruction
1039 \ \text{pixels} \\ \times 50 \ \text{time slices} \\ \times 2 \ \text{telescopes} \\ = 103900 \ \text{data points}
Why Deep Learning?
- Fast on-site inference
- Handles multiple analysis configurations with ease
Jarred Green - jgreen@mpp.mpg.de
A way to include as much information as possible?
1. Current Reconstruction
Jarred Green - jgreen@mpp.mpg.de
Proton
Gamma
Muon
Find a new ways to represent 'raw' data
1. Current Reconstruction
Why waveform data?
- Contains full time evolution of event
- Traditionally too data-heavy to consider
2. Graphs
Jarred Green - jgreen@mpp.mpg.de
Jarred Green - jgreen@mpp.mpg.de
Lets go from images
to graphs
Credit: J. Leskovec (Stanford CS224)
Graphs 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
Why Graphs?
Jarred Green - jgreen@mpp.mpg.de
time
- Sample MC gama
- ⏳ 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
2. Graphs
Jarred Green - jgreen@mpp.mpg.de
Goal: get calibrated data with timing information
Python package magic-gnn
.parquet or LMDB files exported, about ~1GB / min of data
Python package magic-gnn
3. Data
GNN with
calibrated graphs
1️⃣
Transformers
+ GNNs
2️⃣
Jarred Green - jgreen@mpp.mpg.de
3. Data
Training Data:
- 2.5M MC events
- 50/50 gamma-ray / proton split
- GNN framework for Neutrino reconstruction
- 15-20% improvement in reconstruction of energy, zenith, direction at low energies
Jarred Green - jgreen@mpp.mpg.de
4. Playground
✅ Particle ID classification
Model:
- Binary classification task
- DynEdge with 12 NN
- Separate in each telescope
- Cross-attention
- Only need 5-10% of training data
GNN with
calibrated graphs
1️⃣
Jarred Green - jgreen@mpp.mpg.de
4. Playground
✅ Direction reconstruction
In progress:
- Direction classification
- Testing a combination of GNNs with DynEdge and transformers
- Adapted from IceCube's most robust direction reconstruction models
chatbots 🤖
MC 68% containment < 0.08deg
Transformers
+ GNNs
2️⃣
Jarred Green - jgreen@mpp.mpg.de
5. Conclusions
Deep Learning
-
Fast analysis with GPUs
-
Deep models can handle multiple analysis settings
Graph Networks
- Handle irregular camera geometries
- Manages multiple different telescopes
Waveform Data
- Contains full time evolution
- Can save on data processing
Thank you!
Jarred Green - jgreen@mpp.mpg.de
Credit: G. Ceribella
ICRC 2025 - MAGIC Waveform DL
By astrojarred
