Performance of mass-decorrelated jet substructure observables for hadronic two-body decay tagging in ATLAS

Mass-decorrelated jet substructure observables have the potential to increase the sensitivity of searches for new physics in final states with high-$p_{\mathrm{T}}$ hadronically decaying resonances, both by minimising the sculpting of background jet mass distributions and enabling more robust background estimation. A broad study of different analytical and multivariate mass-decorrelation techniques is performed in Monte Carlo simulation: designed decorrelated taggers (DDT), fixed-efficiency $k$-NN regression, convolved substructure (CSS), adversarially trained neural networks (ANN), and adaptive boosting for uniform efficiency (uBoost) are studied and compared. Performance is evaluated using metrics for background rejection and mass-decorrelation.