Identification of tau leptons using a convolutional neural network with domain adaptation

Aram Hayrapetyan, Vladimir Makarenko, A. Tumasyan, W. Adam, Janik Walter Andrejkovic, L. Benato, Thomas Bergauer, M. Dragicevic, Cristina Giordano, Priya Sajid Hussain, M. Jeitler, Natascha Krammer, Ang Li, D. Liko, Mark Matthewman, Ivan Mikulec, J. Schieck, Robert Schöfbeck, D. Schwarz, Maryam Shooshtari, M. Sonawane, W. Waltenberger, Claudia-Elisabeth Wulz, X. Janssen, Hyejin Kwon, Daniel Ocampo Henao, T. Van Laer, Pierre van Mechelen, Jas Bierkens, Nordin Breugelmans, J. D’Hondt, Soumya Dansana, A. De Moor, M. Delcourt, Felix Heyen, Y. Hong, Pavlo Kashko, Steven Lowette, I. Makarenko, D. Müller, Juhee Song, Stefaan Tavernier, M. Tytgat, G. P. Van Onsem, S. Van Putte, D. Vannerom, Bugra Bilin, B. Clerbaux, Aloke Kumar Das, Isabelle de Bruyn, G. De Lentdecker, Hugues Evard, Laurent Favart, P. Gianneios, Mohammad A. Khalilzadeh, Fakhri Alam Khan, A. Malara, Muhammad Aamir Shahzad, Laurent Thomas, M. Vanden Bemden, Catherine Vander Velde, P. Vanlaer, Fengwangdong Zhang, M. De Coen, Didar Dobur, Gul Gokbulut, J. Knolle, Luka Lambrecht, David Marckx, K. Skovpen, N. Van Den Bossche, Jan van der Linden, Jules Vandenbroeck, Liam Wezenbeek, Samuel Bein, Anna Benecke, Agni Bethani, G. Bruno, A. Cappati, J. De Favereau De Jeneret, C. Delaere, A. Giammanco, Ahmet Oguz Guzel, V. Lemaitre, J. Lidrych, Paul Malek, Paola Mastrapasqua, S. Turkcapar, G.a. Alves, M. Barroso Ferreira Filho, E. Coelho, Carsten Hensel, T. Menezes De Oliveira, C. Mora Herrera, P. Rebello Teles, Mariana Soeiro, E. J. Tonelli Manganote, A. Vilela Pereira, Walter Luiz Aldá Júnior, Helena Brandao MalbouissonJournal of Instrumentation 2025 · DOI

TL;DR

The CMS collaboration has developed DEEPTAU v2.5, an improved deep learning algorithm for identifying tau leptons in LHC data. By incorporating domain adaptation, this new version significantly reduces misidentification of jets as tau leptons and improves agreement between simulation and real-world data, crucial for precision physics analyses.

What Problem Does This Paper Address?

Identifying tau leptons (τ) is essential for many physics analyses at the Large Hadron Collider (LHC), particularly those involving Higgs boson decays and searches for new physics. However, tau leptons often decay into hadrons and neutrinos, making their signatures complex and easily mimicked by other particles like quark or gluon jets, as well as misreconstructed electrons and muons. Distinguishing genuine tau leptons from these backgrounds is a significant challenge. Previous methods relied on machine learning, but discrepancies between simulated data (used for training) and real collision data led to performance degradation, especially in regions with high purity of genuine tau leptons.

Methods

The paper introduces DEEPTAU v2.5, an upgraded version of their deep convolutional neural network for tau lepton identification. The core improvements are:

  • Domain Adaptation by Backpropagation: This is the key novelty. It's a technique that trains an "adversarial subnetwork" alongside the main tau identification network. This subnetwork tries to distinguish between real collision data and simulated data. By forcing the main network to perform well while fooling the adversarial network, it learns features that are robust across both data types, reducing discrepancies.
  • Refined Training Workflow: This includes an updated and more balanced training dataset, optimized hyperparameters, and a revised selection of input variables.
  • Convolutional Neural Network Architecture: The underlying CNN architecture remains largely the same as the previous version (v2.1), leveraging both high-level input variables and particle information from "grids" in the detector's pseudorapidity-azimuth (η-φ) space.

Datasets

  • Training: Simulated proton-proton collision data from 2018, and real collision data from 2018 containing Z → ττ decays (used for domain adaptation).
  • Validation & Calibration: Simulated and real collision data from 2018 (at 13 TeV) and 2022 (at 13.6 TeV).

Key Insights and Ideas

  • Domain Adaptation is Crucial for Real-World Performance: The paper strongly emphasizes that even sophisticated ML models trained on simulation can falter when applied to real data due to subtle differences in how particles and detectors behave. Domain adaptation directly tackles this by making the model less sensitive to these simulation inaccuracies.
  • Adversarial Training for Data-Simulation Alignment: The use of an adversarial subnetwork to actively discriminate between data and simulation is a powerful way to enforce consistency. The goal is to make the tau identification performance the same whether the input is real data or simulation.
  • Improved Jet Rejection is a Major Win: The significant reduction in jet misidentification probability (30-50%) is a direct benefit of the improved domain adaptation and training, making it easier to isolate tau lepton signals in complex events.
  • Calibration is Key for Physics Analyses: The paper details extensive efforts to calibrate the algorithm's performance in simulation against real data using scale factors. This is essential for translating simulation-based predictions into precise measurements.

Key Results

  • Reduced Jet Misidentification: DEEPTAU v2.5 reduces the probability of quark and gluon jets being misidentified as tau leptons by 30-50% compared to the previous version, for given reconstruction and identification efficiencies.
  • Improved Data-Simulation Agreement: Domain adaptation significantly reduces discrepancies between data and simulation, particularly in the highest discriminator score bins, leading to a reduction in residual corrections by 5%.
  • Enhanced Performance: The refined training workflow and domain adaptation lead to improved classification performance overall.
  • Calibrated Performance: Scale factors for tau lepton identification efficiency and energy scale have been measured for both 2018 and 2022 data, showing good agreement with unity and improved consistency compared to the previous version.

Limitations and Future Work

  • 2022 Data Challenges: The 2022 datasets are expected to show weaker data-to-simulation agreement due to different detector conditions and collision energies not fully captured by the 2018-trained domain adaptation. Retraining the algorithm with 2022 data would likely improve this.
  • High-pT Tau Leptons: While the algorithm is generally robust, the domain adaptation was not specifically targeted at high-pT tau leptons. The paper notes that consistency of scale factors with unity is observed for both v2.1 and v2.5 in this regime, suggesting the domain adaptation's impact might be less pronounced there.
  • Missing Charged Hadrons Decay Modes: The paper mentions that efficiency measurements have not been performed for decay modes with missing charged hadrons, and future analyses will need to derive appropriate corrections for these.
  • Future Analyses: The algorithm has been introduced for use in CMS physics analyses starting from 2022 data, and its improved performance is expected to benefit a wide range of physics studies.

Summary updated

Abstract

Abstract A tau lepton identification algorithm, DeepTau , based on convolutional neural network techniques, has been developed in the CMS experiment to discriminate reconstructed hadronic decays of tau leptons (τ h ) from quark or gluon jets and electrons and muons that are misreconstructed as τ h candidates. The latest version of this algorithm, v2.5, includes domain adaptation by backpropagation, a technique that reduces discrepancies between collision data and simulation in the region with the highest purity of genuine τ h candidates. Additionally, a refined training workflow improves classification performance with respect to the previous version of the algorithm, with a reduction of 30–50% in the probability for quark and gluon jets to be misidentified as τ h candidates for given reconstruction and identification efficiencies. This paper presents the novel improvements introduced in the DeepTau algorithm and evaluates its performance in LHC proton-proton collision data at √( s ) = 13 and 13.6 TeV collected in 2018 and 2022 with integrated luminosities of 60 and 35 fb -1 , respectively. Techniques to calibrate the performance of the τ h identification algorithm in simulation with respect to its measured performance in real data are presented, together with a subset of results among those measured for use in CMS physics analyses.
BibTeX
@article{hayrapetyan2025identification,
  author = {Aram Hayrapetyan and Vladimir Makarenko and A. Tumasyan and W. Adam and Janik Walter Andrejkovic and L. Benato and Thomas Bergauer and M. Dragicevic and Cristina Giordano and Priya Sajid Hussain and M. Jeitler and Natascha Krammer and Ang Li and D. Liko and Mark Matthewman and Ivan Mikulec and J. Schieck and Robert Schöfbeck and D. Schwarz and Maryam Shooshtari and M. Sonawane and W. Waltenberger and Claudia-Elisabeth Wulz and X. Janssen and Hyejin Kwon and Daniel Ocampo Henao and T. Van Laer and Pierre van Mechelen and Jas Bierkens and Nordin Breugelmans and J. D’Hondt and Soumya Dansana and A. De Moor and M. Delcourt and Felix Heyen and Y. Hong and Pavlo Kashko and Steven Lowette and I. Makarenko and D. Müller and Juhee Song and Stefaan Tavernier and M. Tytgat and G. P. Van Onsem and S. Van Putte and D. Vannerom and Bugra Bilin and B. Clerbaux and Aloke Kumar Das and Isabelle de Bruyn and G. De Lentdecker and Hugues Evard and Laurent Favart and P. Gianneios and Mohammad A. Khalilzadeh and Fakhri Alam Khan and A. Malara and Muhammad Aamir Shahzad and Laurent Thomas and M. Vanden Bemden and Catherine Vander Velde and P. Vanlaer and Fengwangdong Zhang and M. De Coen and Didar Dobur and Gul Gokbulut and J. Knolle and Luka Lambrecht and David Marckx and K. Skovpen and N. Van Den Bossche and Jan van der Linden and Jules Vandenbroeck and Liam Wezenbeek and Samuel Bein and Anna Benecke and Agni Bethani and G. Bruno and A. Cappati and J. De Favereau De Jeneret and C. Delaere and A. Giammanco and Ahmet Oguz Guzel and V. Lemaitre and J. Lidrych and Paul Malek and Paola Mastrapasqua and S. Turkcapar and G.A. Alves and M. Barroso Ferreira Filho and E. Coelho and Carsten Hensel and T. Menezes De Oliveira and C. Mora Herrera and P. Rebello Teles and Mariana Soeiro and E. J. Tonelli Manganote and A. Vilela Pereira and Walter Luiz Aldá Júnior and Helena Brandao Malbouisson},
  title = {{Identification of tau leptons using a convolutional neural network with domain adaptation}},
  journal = {Journal of Instrumentation},
  year = {2025},
  doi = {10.1088/1748-0221/20/12/p12032},
}
Figure