Speaker
Description
The principal ATLAS calorimeter signals are clusters of topologically connected cell signals. These clusters do not only provide an energy and direction measurement, but also have sufficient shape and other structural information allowing to calibrate to individually calibrate them. The standard approach is to use multi-dimensional binned look-up tables to retrieve scale factors for this calibration. Using a machine-learned calibration overcomes some of the limitations of the standard approach, which are mostly introduced by the loss of correlations between the observables used as input, and bin transition effects. A Bayesian neural network (BNN) has been designed to learn a calibration based on the response of the topo-clusters in fully simulated multi-jet final states in the proton-proton collisions at the Large Hadron Collider (LHC), including the effects of pile-up on the calorimeter signal in ATLAS under operational conditions observed in LHC Run 2 (2015-2018). This network also predicts the uncertainty on this learned calibration. The performance of this calibration is compared to previously used and explored topo-cluster calibrations. The learned uncertainties are validated by a comparison to independently derived uncertainty predictions and their interpretation in the context of detector signal features is discussed.