Development and Analysis of Machine Learning in Neutron Clustering Reconstruction in the LHC

A new radiation hard reaction plane detector (RPD) will be installed in ATLAS for the upcoming LHC run 3 heavy ion collisions, located in the far forward region of the LHC and inserted between two modules of the Zero Degree Calorimeters (ZDC). The detector measures reaction plane geometry for nuclear collisions with sufficient spatial resolution to perform directed flow analyses. As part of this, the RPD will measure the spatial distribution of spectator neutrons emerging from each collision as they interact with sixteen channels, excited via Cherenkov radiation, although there is no internal structure with which these channels reconstruct neutron clustering. In previous runs, RPD clustering reconstruction has been accomplished through the naive assumption of neutrons being clustered around the center of mass of an incoming shower, which frequently does not match up with truth values. This study provided an alternative method by utilizing gradient-boosted decision trees and neural networks, both fully-connected and convolutional, to find neutron clustering in GEANT4 simulated data using the energy readings of each channel. Analyzing these models provided information about relative channel importance and implicit biases in the construction of the RPD.