Unbinned energy correlator unfolding in the context of top quark mass measurements

Abstract

This work explores the novel approach of unfolding the threefold energy correlator with generative machine learning methods. In recent investigations energy correlators (EEEC) were found to have a top quark mass sensitive peak which is also accessible analytically. To extract the top quark mass by comparing data with analytic calculations, the measured correlators must be unfolded to correct for detector effects. Normalizing flows with conditional invertible neural networks (CiNN) and conditional flow matching (CFM) are two generative machine learning methods, able to perform unbinned unfolding. In this work, simulated data from boosted tt_bar events was used to construct the correlators with emphasis on feasibility with machine learning methods, while allowing the comparison to theory. In contrast to latest studies, the jet p_T dependency of the mass peak, was solved via the construction of a new observable without this dependency. Configurations of the two models were found enabling the unfolding of these correlators. The concept of unfolding EEEC could be demonstrated and showed promising results. Supplementary investigations regarding the generalization capabilities of the models showed a mass bias towards distributions originating from samples with m_t = 172.5 GeV for both models.