Choose timezone
Your profile timezone:
EP-nu Group Meeting
→
Europe/Zurich
304/1-001 (CERN)
EP-nu Group Meeting
EP-NU group meeting, 3 November 2026 - Minutes
SUMMARY
Two talks and one discussion item. Maksym Ovchynnikov showed that raw Pythia gets the hadronic final-state composition of GeV-mass FIP decays wrong by up to three orders of magnitude, and presented a data-constrained correction (exHad) that reaches a factor of 2 on e+e- data. The group saw a direct link to neutrino generators, which use the same string model at W of a few GeV, and agreed to follow up on testing it against GENIE's KNO-based hadronization for DUNE.
Emanuele Villa presented first feasibility studies for a muon DIS cross-section measurement at TeV energies in FASERnu. At truth level the analysis looks possible, with about 3,600 selected events already in the 9.5 fb⁻¹ of processed Run 3 data. Main feedback: rethink the signal definition (report Q² smearing rather than subtract a hard Q² cut as background).
The EHN1 compute pilot (repurposed CMS HLT nodes) got positive feedback from its users, but cooling problems during LS3 keep the machines offline. Implementation options for a longer-term group compute model will be brought to a future group meeting.
1. HADRONIC DECAYS OF GeV-MASS FIPs: THE PYTHIA ANGLE (M. Ovchynnikov, with V. Kryshtal)
Based on arXiv:2609.16104. Above ~2 GeV, FIP searches hand the decay quarks/gluons to Pythia's Lund string model, tuned to e+e- data at 91 GeV. At 2-5 GeV this extrapolation fails badly.
Results:
- Against BaBar/BESIII e+e- -> hadrons data at 2-3 GeV, raw Pythia overpredicts the simplest channels (π+π-, K+K-, π+π-π⁰) by 10-30x and underpredicts 3(π+π-) by 1-3 orders of magnitude. Data are dominated by 4 and 6 pions; Pythia favours 2-3 mesons.
- For a 2 GeV ALP, raw Pythia puts most decays into parity-forbidden or isospin-suppressed states and gives η(')ππ only 1.4%, vs ~80% from exclusive calculations.
- Correction uses three ingredients: reject final states violating isospin, C or G parity; fix channel-group rates at m* = 2.3 GeV from data/exclusive calculations; channel probabilities fall as s^-2 (seen in e+e- data). Only the composition changes; total hadronic width and lifetime stay fixed.
- 389/397 e+e- measurements below charm threshold lie within a factor of 2. A factor of 2 is the target: appropriate for a non-first-principles model and sufficient for BSM searches.
- Covers dark photon, B-L vector, ALP, Higgs-like scalar, HNL. Public as exHad (github.com/maksymovchynnikov/exHad; Python/C++, Pythia decay handler, JSON/HepMC output), integrated in EventCalc-SHiP.
- At SHiP (2 GeV, cτ = 10 m) the accepted 2-track fraction drops from 11% to 1.8% (dark photon) and 8.3% to 0.6% (ALP). Tracks+photons becomes the largest accepted category (14% and 9%). Sensitivity studies based on raw Pythia (SHiP, LHCb downstream triggers, dark showers, CalRatio) should be revisited; LHCb triggers without π⁰ reconstruction are directly affected.
- Neutrino generators (GENIE, NEUT) use the same string model above W ~2-3 GeV. Open question whether Pythia gets the composition right there; H/D2 bubble-chamber data could test it.
- Next step: a local fragmentation model with these constraints built in (quantum numbers conserved at each string break). Preference is to get this into baseline Pythia (in contact with a Pythia developer); if that is slow, will discuss integration with the authors.
Discussion:
- Model variation band: several parameter sets describing exclusive data equally well, plus the s^-2 exponent varied 1.75-2.25. Affects the overall rate more than the composition, which is what matters for experiments.
- Laura raised the DUNE use case: low-W, low-Q² hadronization on argon has never been measured, is one of the largest uncertainties, and DUNE currently relies on KNO scaling. The correction reweights/rejects Pythia final states after generation, so a generator calling Pythia (e.g. GENIE) could pick it up directly. Below the Pythia region resonances dominate, so any treatment there must be a hybrid matching low-W and high-W descriptions at both ends.
- Proposed comparison: KNO scaling vs constrained Pythia at W ~1.5-2 GeV, varying the tune to estimate uncertainties (factor 2 is too coarse for oscillation analyses).
- Factorization: changing hadronization should not change the inclusive cross section, but GENIE's low-W multiplicities are tuned to neutrino data and also split the non-resonant rate. New GENIE dials varying the hadronization model were implemented last week. Factorization issues deferred to the afternoon discussion.
- Pythia ignores parity because it was built for high-W events with 20-60 hadrons, where per-event conservation averages out.
- Reweighting: Pythia channel rates can be wrong enough for weights to blow up (e.g. ηππ at 10^-3); importance sampling enhances those channels by up to 100x. For neutrinos, a similar check would show whether a dedicated rejection scheme in Pythia is needed.
- Non-minimal BSM (e.g. Z' produced in neutrino scattering, 0.6-2.5 GeV): no fundamental obstacle; Pythia is just not maintained for that mass range.
2. FIRST STUDIES ON MUON DIS IN FASERnu (E. Villa, ETH Zurich)
Preliminary. Goal: first measurement of muon DIS at TeV energies (100 GeV to ~5 TeV), following arXiv:2506.13889, with charm/intrinsic-charm studies later.
Results:
- Run 4 plans: on-axis emulsions, on-axis electronic detector (FORTVNE), off-axis FASERcal (3D segmented scintillator + calorimeters + spectrometer).
- Generation: POWHEG NLO + NNPDF4.0 (MSHT20 comparable) for Q > 1.65 GeV; GENIE for 1 < Q < 1.65 GeV (resonance/QE contamination); Pythia 8 hadronizes both. GENIE scaled to POWHEG at the boundary. Muons flat in log E from 10 GeV to 4.2 TeV, FLUKA spectrum applied in Calypso/Geant4. ~800k muons, DIS cross section x1000 and reweighted -> 60k interactions.
- Selection: Q² > 2.72 GeV², E'_rec > 100 GeV, >= 3 visible tracks above 1 GeV (implies W >= 2 GeV, so no explicit W cut given 20-30% momentum resolution), vertex >= 100 plates from both ends.
- Momentum smearing (arXiv:2602.17575) validated only for 100-plate tracks. Projecting to 700 plates (unvalidated): contamination 20% -> 6%, efficiency 82% -> 91%. Angular smearing negligible.
- DIS secondaries cannot pass the 100 GeV cut (W < 100 GeV), so no feedback background.
- Reco cannot yet go down to 3 visible tracks; efficiency ~29% at 5 tracks, ~10% at 8. Reaching 3 tracks is a priority.
- Yield: ~3,600 selected events in the 9.5 fb⁻¹ already processed (100-plate resolution, 50% reco efficiency), out of ~250 fb⁻¹ in Run 3.
- Backgrounds: neutrinos easy unless vertex overlaps a muon track; charged hadrons reduced by 2 orders of magnitude (to check with full reco); neutral-hadron vertices overlapping muon tracks possibly the only significant background; cosmics negligible. Track inclination vs line of sight suggested as an extra cut.
- Vertex-finding efficiency ~50%; testing a Kalman-filter tool from an ETH student.
- Next: full MC production request after validation, first σ(E) measurement, then differential and charm.
Discussion:
- Emulsion displays show reconstructed tracks merged after scanning; gaps are neutrals or reco failures. DORAEMON interested in FASERnu data for ML plate alignment.
- GENIE is the sample scaled at the POWHEG/GENIE boundary; they differ most near 1 GeV, which barely enters the signal region.
- Smearing applied to incoming and outgoing muon; vertex identified by hadronic activity.
- Signal definition (main feedback): a hard true-Q² cut is an arbitrary definition of DIS. Report in reconstructed Q² (and x) with true-Q² contributions per bin from MC instead of background subtraction, as neutrino experiments do.
- Maksym: hadronization drives multiplicity, energies and angular distributions, so it matters for the track cut and differential results. exHad does not yet cover scattering.
- Possible overlap with highest-energy existing muon DIS data (likely Tevatron, 1990s) to be checked. EIC community also interested.
3. EHN1 COMPUTE MACHINES: PILOT FEEDBACK
Eight former CMS HLT nodes (48 cores, a few TB each) in the EHN1 extension mezzanine, set up by Alessandro. Four in active use (Laura, Dario, Pierre, Gajendra); two unused, two not installed. Ran for 1-2 months, currently offline.
User feedback:
- Analysis (DUNE stack): parallel running over many files is much faster than Fermilab GPVMs. Main issue: downtime with results stuck on the machine; keep important outputs on EOS/AFS.
- MC production (neutrino generator stack, up to tens of millions of events): large productivity gain over Condor, no priority loss. AFS home makes setup identical to lxplus; outputs to EOS. Pain points: no internet without proxy, login via lxplus tunnel, tmux sessions dying daily on AFS token expiry. Suggested fix: local home sourcing settings from AFS.
- Gajendra plans to read out NP02/NP04 test-stand data and run light analysis.
Infrastructure: machines are off due to cooling. The rack cooling circuit has a long-standing leak, and LS3 refurbishment has EHN1 on a temporary chiller due to be removed in winter. Room temperature rose noticeably with all eight on. Nearly all interventions so far were infrastructure-related.
Discussion:
- Options: full cluster with mini Condor pool, smaller easier-to-maintain set, or drop it. Any option needs a named maintainer; a cost-benefit analysis is needed to decide on funding that effort.
- ProtoDUNE machines are not a model (unmaintained, winding down).
- The incoming applied fellow in computing could help once a concrete plan exists.
- Joining HTCondor would need an IT agreement as for LHC experiments; out of scope. Grid covers DUNE-scale production, Condor medium-large jobs; these machines fill the single-user 48-core-on-demand niche, which covers most current use.
- Hyper-K: fast CPU access would help start interaction-model work. Setting up a Hyper-K (possibly container-based) stack on one machine would test a different workflow.
- No GPUs, so ML workflows cannot be tested.
ACTION ITEMS
- Follow up with Maksym on applying exHad-style constraints to neutrino hadronization: KNO vs constrained Pythia at W ~1.5-2 GeV, check reweighting weights (group + Maksym)
- Rework muon DIS signal definition: reconstructed Q² (and x) with true-Q² contributions per bin (Emanuele)
- Check overlap with existing high-energy muon DIS data, likely Tevatron (Emanuele)
- Validate 700-plate resolution projection; push reco down to 3 visible tracks (Emanuele + FASER reco)
- Submit full MC production request after validation (Emanuele)
- Prepare 2-3 implementation schemes for group compute on the EHN1 machines, present at the next group meeting or the one after (Alessandro + 2)
- Discuss compute needs and funding with Francesco
- Document the neutrino-generator software stack used on the machines (Laura)
- Try a Hyper-K software stack on one machine
- Test local home with settings sourced from AFS to fix tmux sessions (Laura + Alessandro)
- Continue the pilot as cooling allows; decide whether to give the unused machines to other users
There are minutes attached to this event.
Show them.