Imaging Using Machine Learning for the LDMX Electromagnetic Calorimeter

Abstract: LDMX is a fixed target experiment designed to search for light dark matter. The experiment will search for dark matter signatures using missing momentum and energy in events with electrons scattering in the target. Part of the experimental setup in this search is a high granularity electromagnetic calorimeter to accurately measure the energies of recoil electrons. In this thesis artificial neural networks are trained on simulated events in the electromagnetic calorimeter to classify events based on their particle contents, and to separate two showers from each other in the calorimeter. For the first task two sets of events are generated, one with a single electron that either undergoes bremsstrahlung or not, and one with multiple primary electrons. Two different neural networks models are used to classify these events. A basic convolutional neural network model achieves a classification accuracy of $96\%$ on the first set of data, with an AUC of 0.99, and an accuracy $99.8\%$ on the second set. A graph neural network model achieves a classification accuracy of $83\%$ on the first set, with an AUC of 0.91, and an accuracy $99.3\%$ on the second set. Another graph neural network model is trained on a subset of the first data set to label the individual hits in the calorimeter by whether they come from and electron shower or a photon shower. This network achieves an average accuracy of $86.6 \%$ over all events and correctly labels $90.5 \%$ of the energy in these events. The neural network models are compared to each other and evaluated based on their performance, limitations, and computing requirements.