Speaker
Description
Long lived particles (LLP) are ubiquitous in many Standard Model extensions, and could provide solutions to long-standing problems in modern physics. LLPs would present unique signatures, such as decays in flight far from the interaction point. New trigger and reconstruction techniques are required to search for such events. We propose using the LHCb muon detector as a sampling calorimeter.
In this work, machine learning based techniques are developed and compared, in order to detect the decays of such particles in the muon detector. The models are designed to be implemented in the software triggers, making use of the GPU accelerated hardware used in LHCb. In order to englobe a wide range of models, the techniques are chosen for their generality. Anomaly detection approaches, such as various types of autoencoders and Siamese neural networks, are benchmarked. Such models necessitate only (or mostly) background data to efficiently select signal events, leading to much improved model independence. Ultimately, such techniques can lead fewer trigger lines to develop, train and maintain, while improving the physics performances. The performances of the models are very promising, yielding similar or better signal efficiencies than BDTs and neural network classifiers.
Finally, a clustering method is designed to extract the signal shower from the background hits. Using properties of the cluster, kinematic information are retrieved using neural networks.
