Energy reconstruction with artificial neural networks on LDMX simulations

Abstract: It is clear from evidence such as rotational curves and cosmic microwave background measurements that dark matter exists. The light dark matter experiment (LDMX) will search for dark matter in the sub-GeV range. It will do this using missing-momentum measurements of electrons interacting with a Tungsten target. The electron will recoil and be measured in the electromagnetic calorimeter (ECal) of the experiment. The accuracy of this measurement is vital for the result of the experiment. Therefore, the ECal design will draw from the Phase-II high granularity upgrade of the Compact muon solenoid (CMS) forward ECal. This thesis have investigated the possibility of using artificial neural networks (ANNs) to improve the energy resolution of the ECal. This was performed on simulation data based on the LDMX framework. Both convolutional neural networks (CNNs) and dense neural networks (DNNs) were trained on the data and compared with a linear fit between ECal readout energy and the original electron energy. The analysis have shown that CNNs can improve the energy resolution of the ECal compared to both the DNN and linear fit who perform similarly. Some inconsistencies in how the models performed on different energies was discovered. Finally, solutions to this and suggestions for future work is discussed.