Research
Graph neural network methods for spatial reconstruction in ground-based gamma-ray astronomy.
§1 Current Research
MSc Thesis · CIC-IPN · 2024–present
The thesis investigates the reconstruction of extensive air shower (EAS) core positions from sparse hit patterns on dual-layer water Cherenkov detector (WCD) arrays, as proposed for the Southern Wide-field Gamma-ray Observatory (SWGO). The central contribution is BiLayerEdgeConv, a graph neural network architecture that treats the electromagnetic and muonic detector layers as separate graph structures and fuses their learned representations through a bidirectional cross-attention mechanism. The work is supervised by Dr. J.A. Martínez Castro (CIC-IPN) and Dr. I.D. Torres Aguilar (INAOE).
Each detector layer is modelled as a k-nearest-neighbor graph over station coordinates. The upper branch applies EdgeConv operators to electromagnetic signals — integrated charge, arrival timing, and PMT amplitude — while the lower branch processes muonic responses from the shielded detector volume. After independent feature extraction, a bidirectional multi-head cross-attention module allows each layer's node embeddings to attend to the other layer's spatial context. Learned gating coefficients control the information flow, enabling the network to weight inter-layer correlations adaptively. The fused representation is passed to a regression head that predicts the shower core position (x, y).
Training and evaluation use CORSIKA-generated showers propagated through a Geant4 detector simulation. The architecture achieves a mean absolute error of 45 m on the core position task — within one station spacing of the 161-tank reference array — and a median error of 29.5 m, representing sub-station-pitch resolution. The full model comprises 119,845 trainable parameters, making it compact enough for deployment in online reconstruction pipelines.
Table 1 — BiLayerEdgeConv Architecture
| Component | Description |
|---|---|
| Upper branch | EdgeConv on electromagnetic layer (charge, timing, PMT amplitude) |
| Lower branch | EdgeConv on muonic layer (shielded detector response) |
| Fusion | Bidirectional multi-head cross-attention with learned gating |
| Output | Core position (x, y) regression |
Table 2 — Preliminary Results
| Metric | Value | Context |
|---|---|---|
| MAE | 45 m | Within one station spacing |
| Improvement | −20% | vs Centre-of-Gravity baseline |
| Median error | 29.5 m | Sub-station-pitch resolution |
| Parameters | 119,845 | Compact architecture |
§2 Publications
- [1] Martínez Castro J.A., Tat'y Mwata-Velu, Mpangi Musungu E., Misonia H., et al. “Suppressor Of Cytokine Signaling Members In Lung Adenocarcinoma: Unveiling Expression Patterns, Posttranslational Modifications, And Clinical Significance.”Journal of Population Therapeutics & Clinical Pharmacology, Vol. 30, No. 18, 2023.
doi:10.53555/jptcp.v30i18.3309 - [2] Misonia H., Martínez Castro J.A., Torres Aguilar I.D. “Reconstruction of EAS Core Position for SWGO via Deep Learning with Spatio-Temporal Bilayer Fusion.”MSc thesis, CIC-IPN, 2025 (in preparation).
§3 Collaborations
SWGO— The Southern Wide-field Gamma-ray Observatory is a next-generation ground-based detector under development in South America, targeting the 100 GeV–PeV energy range with a double-layer water Cherenkov detector array. The collaboration spans 80+ institutions across 14 countries. Misonia contributes to the core position reconstruction pipeline as a student member.
swgo.org
HAWC— The High-Altitude Water Cherenkov Observatory operates at 4,100 m elevation on Sierra Negra, Puebla, Mexico, observing TeV gamma-rays and cosmic rays. Simulation frameworks developed for HAWC (HAWCSim, CORSIKA) inform the architectural decisions in the SWGO thesis work. Misonia participates through the CIC-IPN laboratory.
hawc-observatory.org
§4 Tools & Methods
| Category | Tools |
|---|---|
| Deep learning | PyTorch, PyTorch Geometric, EdgeConv, GATConv |
| Attention | Multi-head cross-attention, learned gating, multi-task learning |
| Simulation | CORSIKA, Geant4, HAWCSim |
| Languages | Python, C++, Bash |
| Infrastructure | SLURM, Git, Weights & Biases, LaTeX |