Trigger-Level Multiple Electron Event Classification with LDMX using Artificial Neural Networks

Abstract: Artificial neural networks is a powerful tool for classifying and identifying patterns in large amounts of data. One of the possible tasks of these networks is classification of data into categories. LDMX is a fixed target experiment that is designed to search for light dark matter particles using missing energy and momentum of electrons passing through and scattering of a target. In the current design the number of electrons passing through is counted using the trigger scintillator system. In this thesis, four different types of artificial neural networks were trained to count the number of electrons present in events using simulated data from both the electromagnetic calorimeter and trigger scinitllator to explore the viability of using neural networks at LDMX. These were convolutional neural networks, recurrent neural networks, graph neural networks, and a combined convolutional and recurrent neural network. Three datasets were generated: one using only data from the electromagnetic calorimeter, one that combined the electromagnetic calorimeter and the trigger scinitllator, and one that combined the electromagnetic calorimeter and a modified design for the trigger scinitllator data. To be viable the accuracy of the models needed to be equal to or exceed 0.99999. None of the models were able to exceed or match this accuracy with the highest accuracy reached by a convolutional and recurrent model training with the electromagnetic calorimeter and a modified design for the trigger scinitllator data of 0.962 and 0.9575. The graph neural network was not able to be trained with data in time and neither was all of the combined convolutional and recurrent neural networks. The different benefits and drawbacks of the models were then evaluated and compared to each other.