Comparisons of features used in machine-learning-based calibration and classification of topological cell clusters in data and MC simulations in the ATLAS calorimeters

Clusters of topologically connected cell signals (topo-clusters) provide the basic calorimeter signals in ATLAS. In addition to a basic representation as a massless (pseudo-) particle at the electromagnetic (EM) and the standard hadronic (LCW) energy scale, they have shape, location, and signal characteristics sensitive to the signal origin, the nature of the incoming particle flow, and the general collision environment characterized by significant levels of pile-up. Machine-learning (ML) networks have been developed to calibrate topo-clusters and to identify those arising from hard-scattering signals in jets. The reliability of such ML-based approaches depends critically on the quality of the MC simulation producing the network inputs (features).

We present a first comparison of these features between ATLAS Run-2 data and MC. The selection criteria and reweighting procedures used to ensure comparable phase spaces between data and MC are described, and the level of agreement between data and MC across the feature set is reported.