Speaker
Description
Employing machine-learned local calibrations for the basic calorimeter signals in the ATLAS experiment at the Large Hadron Collider (LHC), which are formed by clustering topologically connected cell signals (topo-clusters), shows indications of significant performance improvements in terms of accuracy and precision. The most successfully trained model so far is a dense neural network (DNN) employing a heteroscedastic loss function and implementing a Gaussian mixture in a probabilistic approach. This model is found to provide the best performing calibration in the highly stochastic signal environment dominated by the pile-up that is characteristic for the proton–proton collisions at the LHC, indicating a significant mitigation of these pile-up contributions and the associated fluctuations. In addition, independent classification networks have been studied with the goal to reduce the effect of pile-up on reconstructed calorimeter jets by reweighting the topo-cluster signal contribution to the jet kinematics. The corresponding weighting functions are based on the amount of signal arising from pile-up in a given topo-cluster. This talk will present the status of the latest developments for both calibration and classification, as well as a brief outlook on future applications in ATLAS.