Abstract
High energy particle physics is the study of our universe's most fundamental particles and forces. One area of research in the field of particle physics is ultraperipheral collisions. Ultraperipheral collisions are collisions of relativistic nuclei in which the impact parameter is greater than the sum of the two radii; this leads to electromagnetic interactions rather than strong interactions. Standard searches for new or rare particles in ultraperipheral collisions rely on predefined decay topologies or available Monte Carlo simulations. A separate analysis must be conducted for each particle and/or decay topology. This thesis presents possible strategies for the detection of rare particles by means of anomaly detection through the usage of autoencoders, a type of unsupervised machine learning. This method makes it possible to flag exotic events without having to define specific selection criteria and allows for multiple simultaneous searches in a single analysis.
Two autoencoder implementations are investigated. They are trained with simulated samples of processes observed in ultraperipheral collisions by the ALICE detector at the Large Hadron Collider, and realistic experimental particle identification capabilities have been included. These autoencoders are then tested using an independent sample of the same processes that has been injected with rare events. The test sample combines the various processes in approximately the same ratios as they are observed in ALICE data. The autoencoders show success in flagging exotic decays with efficiency and purity values comparable to, or better than, traditional searches.