Speaker
Description
High-energy particle collisions at the LHC produce prompt photons, which carries clean information about the partons, quarks and gluons, inside the colliding protons or nuclei. The dominant contamination in any such measurement comes from neutral pions ($\pi^0$), which decay into two photons which are boosted so close together that they strike the detector at a very mimimum distance
and appear as a single photon-like cluster. Distinguishing these prompt photons from $\pi^0$ decay pairs is therefore essential before a direct-photon signal can be extracted from the $\pi^0$ background.
The ALICE Forward Electromagnetic Calorimeter (FoCal-E) is a highly granular, 20-layer silicon-tungsten sampling calorimeter covering $3.2 < \eta < 5.8$, designed for exactly this task: where tungsten's short radiation length keeps showers compact and the finely segmented silicon layers resolve their lateral structure at the pixel level. This granularity, makes it possible to extract shower-shape information fine enough to separate merged $\pi^0$ showers from genuine single-photon showers on an event-by-event basis.
Using a Geant4 simulation of FoCal-E, single-photon and $\pi^0 \rightarrow \gamma\gamma$ events at 100~GeV were generated, and seven transverse shower-shape observables, including the Transverse Shower Profile (TSP) at four radial powers ($1,2,3,4$), lateral RMS widths ($\sigma_x, \sigma_y$), and the energy-weighted spread $E/\Delta$, were computed at pixel \
layers 5 and 10. These 14 features were used to train a Boosted Decision Tree (BDT) classifier, which achieves an AUC of 0.9595 and a $\gamma$-rich-region purity of 0.9194 on an independently generated 80{,}000-event test sample,which the model has never seen before, demonstrating strong discriminating power from shower-shape variables alone. Work is now extending to pixel-level and graph-based representations (CNN/GNN) to exploit the full 2D and longitudinal energy profile without relying on handcrafted variables.