A Machine Learning Approach to Local Energy Calibration in the ATLAS Calorimeters

The intention of an energy calibration scheme is to provide a calorimeter signal for physics object reconstruction in ATLAS which is calibrated outside an assumption about the type of object. This is of particular importance for final-state objects with a significant hadronic signal content, such as jets. The energy calibration method used in run 1 of The ATLAS detector was called Local Hadronic Cell Weighting (LCW) Calibration [1]. The LCW calibration aims at the cluster-by-cluster reconstruction of the calorimeter signal on the appropriate (electromagnetic or hadronic) energy scale. This method used bins and a lookup table to calibrate the energy. In this, the cluster energy resolution is expected to improve by using other information in addition to the cluster signal in the calibration. The calorimeter signal inefficiencies that calibration includes non-compensating calorimeter response, signal losses due to clustering, and signal losses due to energy lost in inactive material. The LCW method worked well for high energy and is the current standard. However, LCW calibration had poor response for energies less than 100 GeV. A Machine Learning (ML) approach to energy calibration could help to alleviate the issues that arise with LCW calibration. The hope of this approach is that it will be better at calibrating the energy for all energies. Also, due to the amount of data that ATLAS has, LCW calibration is computationally more expensive than a ML based model. In tandem with such a scheme, this paper outlines a project where we calibrate energy based on Geant4 simulation data. The data from the full simulations are stored in ROOT tuples, in a tree format. Each row vector in the dataset contains the signal, moments and information related to the composition for one topo-cluster. Since this data is produced through simulation, we have access to information, such as the topo-cluster true energy, which is not available in an experimental situation. This makes such a simulated format valuable for developing robust ML enabled energy calibration schemes. The remaining part of the paper is organized as follows. We first introduce how we devise ML frameworks that are capable of multi-dimensional regression. Then we describe the various cuts of data and learning experiments we conducted along with results and discussion. We summarize the findings and remarks on future works in the last section.