Speaker
Description
We present the progress of detector-focused AI studies at CLAS12 and at the future EIC 2nd detector. At CLAS12, we developed an attention-based model that clusters hits in the experiment’s hodoscopic forward calorimeter. In this model, each hit is treated as a token, with positional encodings learned via a dedicated module incorporating several GravNet layers. Utilizing the object condensation framework, this approach has been shown to outperform the current clustering algorithm, significantly enhancing event reconstruction efficacy. In addition, detector optimization studies have begun exploring the design space of a KLM detector at the EIC using the AID(2)E framework. To enable this otherwise computationally expensive project, we developed fast simulations of the detector response using a normalizing flow model. Graph neural networks trained on simulated scintillator responses from the KLM have demonstrated their ability to accurately reconstruct particle energy, providing the critical feedback needed to optimize the overall design.
