Photon Reconstruction of Axion-Like Particles with Graph Neural Networks at Beamdump Experiments

Axion-Like Particles (ALPs) are hypothetical weakly interacting light particles predicted
by theories Beyond the Standard Model which could be mediators between a dark sector
and the Standard Model. Some of these theories predict light ALPs which decay into two
photons and could be detected at future beamdump experiments such as LUXE - New
Physics search at Optical Dump.

To investigate the properties of such ALPs, an accurate reconstruction of the common
photon vertex from the hits measured in the detector can aid the search significantly. For
this purpose, the photon shower direction needs to be reconstructed precisely, combining
techniques from cluster and track reconstruction. This task is a prime candidate for modern
methods of reconstruction based on Machine Learning such as Graph Neural Networks.

This thesis presents a Graph Neural Network composed of GravNet and GarNet, which
is able to reconstruct the decay vertex of ALPs from the sparse detector hits of the two
photon showers. The performance of the network is assessed on a data-set simulated with a
high granularity calorimeter.