Speaker
Description
Accurate reconstruction of neutral particles (neutrons and photons) is essential for a broad range of nuclear physics measurements. However, the current COATJAVA reconstruction software at CLAS12 produces an overabundance of false neutral clusters, necessitating conservative selection cuts that reduce reconstruction efficiency. To address this limitation, we present an AI-based clustering model designed specifically for the CLAS12 electromagnetic calorimeter (ECal).
The model combines GravNet layers for local detector topology encoding with a Transformer encoder to capture long-range hit relationships, utilizing the Object Condensation framework for end-to-end cluster prediction. Evaluated on 1 million simulated e⁻+p collision events, our approach demonstrates significant improvements in particle reconstruction trustworthiness: neutron trustworthiness increases from 8.88% to 30.73%, and photon trustworthiness improves from 51.07% to 64.73%. This represents the first application of AI-based hit clustering to hodoscopic detectors at CLAS12. The model's ability to effectively suppress false clusters while maintaining detection efficiency provides a powerful tool for future physics analyses requiring precise neutral particle identification.