Machine Learning Approaches for ALERT Track Reconstruction at Jefferson Lab*

Understanding how the structure of bound nucleons is modified within the nuclear medium remains a central challenge in hadronic physics. Deeply Virtual Compton Scattering (DVCS) offers a powerful probe of the partonic structure of light nuclei, providing access to their three-dimensional distributions via Generalized Parton Distributions (GPDs) and enabling studies of associated in-medium stimulated effects.

The recently completed CLAS12 experiment at Jefferson Lab employed the newly built A Low Energy Recoil Tracker (ALERT) detector to study tagged DVCS by impinging an 11~GeV polarized electron beam on $^4$He or $^2$H. The ALERT detector is comprised of a hyperbolic drift chamber (AHDC) and a time-of-flight (ATOF) array to provide an effective separation of low-momentum nuclear-recoil fragments, including $^1$H, $^2$H, $^3$H, $^3$He, and $^4$He, down to 70~MeV/c across a broad kinematic range. 

 Recent advances in artificial intelligence (AI), particularly in model architectures, have demonstrated strong performance in high-rate experiments with significant background noise, such as the AHDC case. To this end, AI-based techniques have been developed to enhance AHDC track-finding efficiency, purity, and processing speed relative to conventional algorithms, as well as to identify recoil fragments and optimize matching between AHDC tracks and ATOF hits. In this talk, I will present a highlight of the ALERT physics program and provide an update on the ongoing development and optimization of AI-assisted track reconstruction and particle identification for both simulated and real data.