Machine learning-based optimisation of Higgs coupling measurements in the H → 4l decay channel with ATLAS Run 3 data

Cross-section measurements for the different Higgs boson production and decay processes constitute a key area in the exploration of Higgs properties, with a high sensitivity to potential physics beyond the Standard Model. Due to its exceptionally clear signal, the decay of a Higgs boson into a ZZ$^*$ pair with a subsequent decay of each Z boson into two signature light leptons, H → ZZ$^∗$ → 4 l , is one of the most important channels for the Higgs property measurements. To increase the sensitivity of the Higgs cross-section measurements, the Simplied Template Cross Section (STXS) approach has been developed, where the final state is divided into exclusive phase space regions. Optimized classification of events according to kinematic production regions is vital to improve the signal sensitivity. The previous round of STXS measurements in the H → 4 l channel with the Run 2 ATLAS dataset employed a event classification using a Recurrent Neural Network (RNN) based approach. With the new Run 3 dataset at a centre-of-mass energy of 13.6 TeV, an alternative Neural Network approach based on permutation invariant Deep Sets is explored for this classification, which is motivated by permutation invariant symmetries between the H → ZZ$^∗$ → 4 l final state particles. A direct comparison between the predictive behaviour and the signal-/background separation is made between the Deep Set and RNN models in all STXS bins. Similar comparisons are made between Deep Sets trained on Run 2 and Run 3 data. The STXS classification schemes used in the RNN training during Run 2 and Run 3 are both considered for the Deep Set Neural Networks. Improved or comparable signal-to-background separations were observed for the Deep Set models compared to the RNN models in all kinematical STXS classification regions. Training on Run 3 data lead to improvements in the Deep Set performance compared to Run 2 data. The Run 2 iteration of the STXS binning scheme is in general observed to be more favourable for the Deep Set performance. The Deep Set approach can be considered as a viable alternative to the previous RNN approach.