Machine Learning for Anomalous Particle Detection in Ultraperipheral Relativistic Collisions of Heavy Ions

We present possible strategies for the detection of rare particle decays in diffractive and ultraperipheral collisions by means of anomaly detection. Ultraperipheral collisions are collisions of relativistic nuclei where the impact parameters are greater than the sum of the two radii of the nuclei and are suited to study exclusive processes. Standard searches for new or rare particle decays in ultraperipheral collisions rely on predefined decay topologies or available Monte Carlo simulations. By implementing anomaly detection through the usage of autoencoders, it is possible to flag anomalous events without having to define specific selection criteria. Our autoencoder designs are trained with toy samples of processes observed in ultraperipheral collisions in the ALICE detector at the Large Hadron Collider. Realistic experimental particle identification capabilities have been included. The designs are tested using an independent sample of typical events that has been injected with rare events. The test sample combines the various processes included in approximately the same ratios as they are observed in ALICE data. The autoencoder is able to flag those injected rare events as anomalous. It is also possible to use a search based on anomaly detection to establish exclusion limits on the production of new resonances, or exotica, in UPCs. This approach demonstrates the applicability of a new technique for rare particle searches in the current and future data sets at collider experiments.