Abstract
"Title: Development of MIDDLE, a novel method for the identification of the decays of heavy flavour hadrons to muons, to measure the fragmentation properties of the b-quark and the mass of the top quark. Description: The MIDDLE project aims to develop a novel Deep Learning–based method for identifying muons originating from the semileptonic decays of heavy-flavour hadrons (b- and c-hadrons) within hadronic jets. By exploiting detailed detector information—such as impact parameters, track quality, calorimetric deposits, and isolation profiles—MIDDLE will achieve superior background rejection compared to existing Boosted Decision-Tree approaches, while maintaining minimal bias from jet kinematics. The tool will be implemented within the ATLAS software framework and optimized for both Run-2 (√s=13 TeV) and Run-3 (√s=13.6 TeV) datasets, leveraging approximately 250 fb⁻¹ of LHC data for final measurements" The MIDDLE project is designed to revolutionize heavy-flavour muon identification by developing a state-of-the-art deep-learning classifier that dramatically improves signal purity and background rejection in hadronic jets using detailed inner-detector and calorimeter information . Building on this foundation, the MIDDLETOP extension will precisely reconstruct top-quark decay chains—distinguishing muons from b→μX, b→c→μX, and c→μX processes—and exploit charge and angular correlations to further enhance classification performance . Taking advantade from these advanced tagging tools, the project will deliver the first measurement of b-quark fragmentation observables in tt¯ events at the LHC, providing critical benchmarks for theoretical models and Monte Carlo generators. Finally, by combining superior muon-tag purity with robust top-reconstruction techniques, MIDDLE/MIDDLETOP aims to achieve a top-quark mass determination with an uncertainty below 500 MeV—setting a new standard for precision in direct mₜ measurements. MIDDLE and MIDDLETOP will be delivered as fully calibrated, publicly documented tools within ATLAS, each culminating in a high-impact publication. WP1/2 papers will describe the algorithms and their performance, while WP3/4 will present world-first measurements of b-quark fragmentation in tt¯ events and the most precise single direct top-mass determination. Beyond these flagship analyses, the tools are poised to enhance W/Z+HF cross-section measurements, improve boosted top/Higgs tagging by incorporating leptonic inputs, and support searches for new physics—thereby offering a versatile resource for the particle-physics community well beyond the project’s 24-month lifespan.
Results achieved
: The PRIN 2022 20223N7F8K project, titled “Development of MIDDLE, a novel method for the identification of the decays of heavy flavour hadrons to muons, to measure the fragmentation properties of the b-quark and the mass of the top quark” – CUP I53D23000820006, was carried out as part of Mission 4, Component 2, Investment 1.1 of the National Recovery and Resilience Plan, with financial support from the European Union – NextGenerationEU. The project, running from September 28, 2023, to February 28, 2026, involved INFN Roma Tor Vergata, the University of Rome Tor Vergata, and the University of Bologna, and was developed in the context of the ATLAS experiment at CERN's Large Hadron Collider. The overall goal was to develop new artificial intelligence techniques to identify muons produced in the decays of hadrons containing heavy quarks within hadronic jets and use that information to improve precision measurements of the fragmentation properties of the b quark and the top quark mass. The first main result was the development of TightNN, a new algorithm for identifying so-called soft muons, based on deep neural networks. The algorithm uses various reconstructed muon properties to distinguish muons produced in the decays of hadrons containing heavy quarks from background contributions, particularly those arising from the decays of light hadrons. At the same signal efficiency, TightNN improves background rejection by up to a factor of two compared to techniques previously used in ATLAS. The method has been integrated into the experiment's official software through ONNX interfaces and is therefore available to the entire ATLAS Collaboration and for future physics analyses. Calibration of its performance on collision data is also underway. The project also produced a new model dedicated to identifying low transverse momentum muons, in the 3–18 GeV range. The model, developed using machine learning techniques based on Gradient Boosted Decision Trees, achieved about a 30% improvement in rejecting backgrounds from pions and kaons wrongly identified, while keeping the efficiency for real muons unchanged. This tool is also being integrated into ATLAS's official muon reconstruction and identification software. Another finding involves extending the MIDDLE approach to events containing top–antitop quark pairs. A multi-class classification method, called MIDDLETOP, was developed to match the observed muon to the correct decay chain and identify whether it comes from a top or antitop quark. The method distinguishes, among other things, direct decays of hadrons containing b quarks, decays of hadrons containing c quarks produced from b hadrons, decays of c hadrons without a b ancestor, and other production mechanisms. This classification helps reduce ambiguities and more precisely control uncertainties in measurements based on soft muons. Using the developed tools, a new ATLAS analysis was carried out on the fragmentation properties of the b quark in top–antitop events through observables built from soft muons. The work included defining and optimizing the event selection, studying observables sensitive to different fragmentation models, comparing Monte Carlo generators, producing dedicated simulated samples, developing unfolding procedures to correct for detector effects, and evaluating the main sources of systematic uncertainty. The analysis has been integrated into the official ATLAS Top Physics group program and forms the basis of a scientific paper currently in preparation. Studies have also been carried out for a new measurement of the top quark mass using purely leptonic observables, which use muons produced in b-hadron decays. The sensitivity of the measurement has been improved by taking advantage of the data collected during Run 2 and Run 3, the performance of the new tagger, the topological reconstruction techniques developed in the project, and the analysis infrastructure set up for the fragmentation measurement. The project activities have also contributed to an ATLAS publication dedicated to measuring the top quark mass through decays containing a J/ψ meson. The overall results include validated scientific software integrated into the official ATLAS frameworks, analysis methodologies, simulated samples, technical documentation, contributions to scientific publications, and training activities. The funding supported the recruitment of three postdoctoral researchers and helped train young researchers, PhD students, and undergraduates through research activities, scientific qualification, and thesis work. The activities and results were presented within the ATLAS Collaboration, at national and international workshops, and at the National Congress of the Italian Physical Society. Initiatives were also carried out for schools and the general public, including ATLAS masterclasses, events for Researchers' Night, outreach meetings, and informal presentations focused on particle physics and artificial intelligence applications. A dedicated website ensured continuous visibility of the project's goals, activities, and results. All the planned scientific goals have been achieved. The completion of publications beyond the administrative end of the project is ensured by integrating the activities into the ATLAS Collaboration's scientific program, by the availability of the acquired computing and storage infrastructures, and by the ongoing involvement of the participating institutions. Therefore, the results will remain usable for future precision measurements and for new applications of machine learning in high-energy physics. The activities were carried out in compliance with the Do No Significant Harm principle, with equal opportunity principles, and with open access policies. The publications resulting from the project are, or will be, made available in Open Access, according to the CERN and ATLAS Collaboration policies.Project details
Unibo Team Leader: Matteo Franchini
Unibo involved Department/s:
Dipartimento di Fisica e Astronomia "Augusto Righi"
Coordinator:
INFN-Istituto Nazionale di Fisica Nucleare(Italy)
Total Unibo Contribution: Euro (EUR) 57.728,00
Project Duration in months: 24
Start Date:
28/09/2023
End Date:
28/02/2026
