Speaker
Description
Particle identification (PID) of cosmic‐ray nuclei using silicon strip detectors is employed in many space-born experiments, such as AMS-02, DAMPE, and HERD. However, the detector response exhibits strong dependence on the relative hit position between strips, and the nonlinear effects of front-end electronics under large charges make it difficult to establish explicit analytic expressions for the signal–charge relation. As a result, conventional heavy nuclei PID approaches often rely on external detector information to correct the cluster signals.
We present a novel unsupervised learning algorithm for heavy nuclei PID based on an AutoEncoder architecture. By introducing a newly designed histogram loss, our method enables direct learning of the high-dimensional cluster features obtained from real data without the need for labels. This approach allows simultaneous extraction of particle charge and hit position, demonstrating the potential of unsupervised learning in addressing long-standing challenges in learning high-dimensional physical information, such as Jet tag.
