Abstract
We present an application of unsupervised learning for zero-bias detection of
rare particle decays and exotic hadrons in low-background environments such as
those characteristic of diffractive events and ultraperipheral pp, p--A, or
A--A collisions at the CERN Large Hadron Collider (LHC), or in e--A collisions
at the ePIC experiment at the future Electron-Ion Collider (EIC). Using a toy
dataset simulating the decays of known resonances, including
$\ensuremath{{\mathrm J}/\psi}\xspace$ and {\ensuremath{\psi'}\xspace}, as well
as more exotic candidates, we implement an autoencoder neural network to
identify anomalies in the decay kinematics. The autoencoder, trained solely on
typical events, is designed to reconstruct normal decays with low error while
flagging anomalous decays based on the reconstruction error. We demonstrate
that the autoencoder successfully separates typical decays from rare exotic
events, with peaks in the invariant mass distribution corresponding to the
injected rare signals. Our method shows promise in detecting rare, unpredicted
processes in large-scale collider data, offering an effective approach for
discovering new physics beyond the Standard Model.