The ATLAS experiment at the Large Hadron Collider (LHC) features a large calorimeter system for the reconstruction of the energy and direction of particles emerging from the proton–proton (and heavy ion) collisions. Its principal signals are clusters of topologically connected cell signals (topo-cluster) that are intended to represent the energy deposits of interest. Their signals are subject to significant fluctuations from a background generated by pile-up in a collision environment characterized by the high proton bunch intensities and the high bunch crossing frequency in the LHC. Calibrating the topo-cluster potentially reduces the effects from pile-up, in addition to correcting for the non-compensating signal character of the various calorimeters in the system. Such a calibration can employ various observables associated with each topo-cluster as inputs. The ones that are sensitive to the origin of the measured signal are of particular interest. They include the cluster location, its spatial extension and shapes, and the relevance, density, compactness and timing of its signal.
In this talk we present a recent approach where a Bayesian neural network is not only trained to predict the calibration for each individual topo-cluster from a selected set of inputs but also to learn the uncertainty of this prediction. After a brief introduction to the ATLAS calorimeter system, the motivation behind the selection of the inputs is discussed, together with the network setup and training procedures. This is followed by summarizing the precision and accuracy achieved in testing the model. The talk concludes with an interpretation of the predicted uncertainties in the context of the particularities of the signal formation in certain regions of the calorimeter.
The presented results are published in SciPost Phys. 19 (2025) 6, 155 (arXiv: 2412.04370 [hep-ex]).