Go to the Indico Home Page Oldest event Older event Up to category Newer event Newest event Printable version
 
Focus on:  Hide Contributions
  • Compact style
  • Indico style
  • Indico style - inline minutes
  • Indico style - numbered
  • Indico style - numbered + minutes
  • Indico Weeks View
    • Switch to the management area of this event
    Login

    Tau Reconstruction Meeting

    → Europe/Zurich
    Laurits Tani (National Institute of Chemical Physics and Biophysics (EE)), Michel Hernandez Villanueva (Brookhaven National Laboratory (US))
    Description

    Zoom: https://cern.zoom.us/j/65046467552?pwd=AopXEaX4fYHPFTxRmIf7iQx9q8dXr8.1

    Live notes
    Hide

    FCC Tau Reconstruction - October 9, 2026

    Present: Laurits Tani, Norman Seeba, Sofia Giappichini, Karl Ehataht, Michel Hernandez Villanueva, Mohadese, Federico Melis, Attilio Andreazza,

    Apologies: Maria Cepeda

    Introduction

    • Two contributions today
    • We were asked if there could be a dedicated tau reconstruction session during the next reconstruction workshop at CERN
      • https://indico.cern.ch/event/1719015/ 
      • Wed morning, 90 min = intro + 4 speakers
      • We need to define speakers

    ParticleTransformer is all you need (Norman)

    • Focused on hadronic tau decays
    • The main motivation is hadronic taus are difficult to separate from jets
    • Using a Transformer to reconstruct tau from jets, 4 tasks
      • Tau ID
      • Decay mode identification
      • Electric charge reconstruction
      • Kinematic reconstruction (visible 4-momentum)
    • Two approaches
      • SingleParTau - Each task is solved independently through per task training
      • MultiParTau - A single model is trained to solved all four tasks
    • Dataset: Fu𝜏ure v3
      • e+e- samples with Z->ττ, and Z->qq events, simulated with Pythia 8 at sqrt(s) = 91 GeV
      • CLD, full detector simulation by Geant4
    • Each tau jet represents a set of reconstructed constituents, feature vector
      • Relative kinematics (Δη, Δφ, log(pT), ΔR)
      • ParticleID
      • Impact parameters
    • Evaluating SingleParTau vs MultiParTau, with the later outperforming slightly
    • Decay mode classification: Different decay modes are not equally populated in the dataset
      • The “Overall” category represents the average performance across all channels weighted by their branching fractions
      • Good performance for well populated decay modes, poorly for rare decays
    • Results: Kinematic reconstruction - The visible transverse momentum is well reconstructed with the jet density
      • At higher values the spread grows due to the limited number of high-pT jets in the dataset

    Questions, comments

    • Why log scale in pT and E?
      • It helps the ML algorithm, slightly better performance
    • You could try to request a dataset with normalized BRs, to see if the performance in rare decays is recovered
      • Nice idea, but not high priority, since most people will care about primary decay modes and we don’t want to lose performance on them
    • Impact parameter is respect to the primary vertex?
      • Yes
    • Let’s request a dataset with alternative models in the tau decay (KKMC+Tauola) to asset the impact of the tau decay dynamics in the performance

    Heuristic tau reconstruction (Sofia)

    • Focused as well in hadronic tau decays
    • A C++ function that loops through the jets and their constituents
      • Marks jets with electrons or muons as non-taus
      • Find the leading constituent
      • Add other constituents with pt > 1 GeV or Δ𝜃<0.2
      • Define tau decay mode based on constituents - 1 or 3 prong + photons
    • Tested the 𝜋0 reconstruction inside the function
      • Not successful due to how Delphes reconstruct photons, not investigated in depth
      • In the end (for the specific use case), specific number of photons in the tau jets was not very informative
    • Integrated in FCCAnalyses in the pre-edm4hep1 branch
      • PR merged and the findTauInJet can be found under the ReconstructedParticle analyzer
      • Not tested if working with newer edm4hep samples

    Questions

    • Are there any plans to implement as Gaudi algorithm?
      • Not from Sofia side, anyone interested is welcome to work on it
    • About documentation, well documented in the code, can this be propagated to a web interface?
      • To be followed with Juraj, no plans from Sofia
    • Let’s consider what benchmarks can be defined from this algorithm

    AOB

    • Who would be the 4th speaker?
      • Let’s talk with Aurelien and see if Miguel can talk about the mass reconstruction and meaningful information for their study
    • Laurits talked with Uli and particleID could be added soon to PF
    • CLD asked if we could perform a sensitivity study of |Vus| with only neutral kaons
      • Michel will talk with people about it, but probably no, or heavily systematic limited (Emilie Passemar is an excellent option to ask)
    • Testing the Thrust implementation in FCCAnalyses; I think there is a memory leak in the current implementation (cgroup out of memory: Killed process).
      • Michel will follow with Juraj and open a PR.
    There are minutes attached to this event. Show them.
      • 12:00 → 12:10
        Introduction 10m
        Speakers: Laurits Tani (National Institute of Chemical Physics and Biophysics (EE)), Michel Hernandez Villanueva (Brookhaven National Laboratory (US))
        Introduction.pdf
      • 12:10 → 12:30
        ParticleTransformer is all you need for reconstructing hadronic tau leptons 20m
        Speaker: Norman Seeba (National Institute of Chemical Physics and Biophysics (EE))
        Taumeeting261009.pdf
      • 12:30 → 12:50
        Heuristic tau reconstruction 20m
        Speaker: Sofia Giappichini (KIT - Karlsruhe Institute of Technology (DE))
        FCCTau_methods_261009.pdf
      • 12:50 → 13:20
        Round table 30m