Energy reconstruction in a liquid argon calorimeter cell using convolutional neural networks

TL;DR

This paper demonstrates that convolutional neural networks, trained with a novel loss function, can outperform traditional optimal filter techniques in reconstructing particle energy in a liquid argon calorimeter, especially under high pileup conditions.

Contribution

The study introduces a novel loss function for CNN training and shows its effectiveness in improving energy reconstruction accuracy over existing methods.

Findings

CNNs outperform optimal filter in energy accuracy

CNNs maintain performance under high pileup conditions

Novel loss function enhances CNN training effectiveness

Abstract

The liquid argon ionization current in a sampling calorimeter cell can be analyzed to determine the energy of detected particles. In practice, experimental artifacts such as pileup and electronic noise make the inference of energy from current a difficult process. The beam intensity of the Large Hadron Collider will be significantly increased during the Phase-II long shut down of 2025-2027. Signal processing techniques that are used to extract the energy of detected particles in the ATLAS detector will suffer a significant loss in performance under these conditions. This paper compares the presently used optimal filter technique to convolutional neural networks for energy reconstruction in the ATLAS liquid argon hadronic end cap calorimeter. In particular, it is shown that convolutional neural networks trained with an appropriately tuned and novel loss function are able to outperform the optimal filter technique.