Speaker
Description
Several nuclear physics studies using the CLAS12 detector rely on the accurate reconstruction of neutrons and photons from its electromagnetic calorimeter system. These studies are often limited by an overabundance of false neutral clusters produced by the existing reconstruction software, COATJAVA, which reduces the reliability of neutral particle measurements and restricts the physics reach of multiple analyses.
We present a new AI-based clustering method for CLAS12 calorimeter hits built on the object condensation framework. The model uses GravNet layers to learn a latent representation of the full detector topology through local message-passing, which then serves as positional encoding for a Transformer encoder that captures long-range contextual information across the calorimeter. This hybrid architecture allows the network to contextualize both local and global features, improving its ability to handle the hodoscopic geometry of the CLAS12 calorimeter. Our study also marks the first application of AI clustering techniques to hodoscopic detectors, showing potential for use in many other experiments.
Evaluated on one million simulated $e^{-}+p$ collision events, the method significantly improves cluster trustworthiness: the fraction of reliable neutron clusters increases from 6.99% to 17.48%, and photon clusters from 51.76% to 56.93%, compared to the COATJAVA baseline. Ongoing work focuses on validating the model against discrepancies between Monte Carlo simulation and experimental data, with the goal of deploying AI-assisted clustering within the CLAS12 reconstruction pipeline.