pith. machine review for the scientific record. sign in

arxiv: 2605.12407 · v1 · submitted 2026-05-12 · ✦ hep-ex

Recognition: 2 theorem links

· Lean Theorem

Novel Machine Learning Methods to Improve Z Pole Integrated Luminosity at Future Colliders

Brendon Madison

Pith reviewed 2026-05-13 03:09 UTC · model grok-4.3

classification ✦ hep-ex
keywords machine learningluminosity measurementZ poleBhabha scatteringdiphotonbeam deflectionparticle identificationfuture colliders
0
0 comments X

The pith

Machine learning techniques enable the required 10^{-4} precision in luminosity measurement for future Z-pole electron-positron colliders.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper explores using machine learning to improve the accuracy of integrated luminosity measurements at future electron-positron colliders operating at the Z pole. These colliders require a precision better than one part in ten thousand to meet their physics goals. The work focuses on two issues: rejecting background events in the diphoton channel using boosted decision trees, and correcting for beam deflection biases using both standard and new regression methods. If successful, these methods would allow reliable luminosity monitoring using both small angle Bhabha scattering and diphoton events.

Core claim

The paper demonstrates that gradient boosted decision trees classify events to reject neutral hadron backgrounds in the diphoton luminosity channel using existing and upgraded detectors, but only the upgraded luminosity calorimeter rejects small angle Bhabha scattering at the required δL/L < 10^{-4}. A newly developed Adaptive Symbolic Memetic Regression (ASMR) algorithm outperforms boosted decision trees for event-by-event beam deflection correction, reducing the uncertainty to 5×10^{-6}.

What carries the argument

Gradient boosted decision tree classification for particle ID in forward trackers and LumiCal, combined with Adaptive Symbolic Memetic Regression for beam deflection correction.

Load-bearing premise

Machine learning models trained on simulated background and beam deflection data will generalize to real future collider conditions with the same performance, and that all relevant backgrounds have been correctly modeled in simulation.

What would settle it

A comparison of the ML-corrected luminosity value against an independent calibration such as Z boson production rates on actual future collider data, checking whether the total uncertainty stays below 10^{-4}.

Figures

Figures reproduced from arXiv: 2605.12407 by Brendon Madison.

Figure 1
Figure 1. Figure 1: Forward region layout used in this study, based on the ILD forward region used at ILC [4]. The existing ILD LumiCal is shown on the bottom while the GLIP LumiCal, which is further document in other studies, is shown on the top. [1, 6] and beam electromagnetic deflection. To optimize these, this paper deploys two machine￾learning algorithms: ROOT TMVA’s Gradient Boosted Decision Tree (BDTG), and Adaptive Sy… view at source ↗
Figure 2
Figure 2. Figure 2: MAE scaling of BDTG and ASMR with parameter count for the benchmark test that used the W-M function, with arbitrary units and 2 × 10−4 intrinsic error added using Gaussian smearing. 1 2 3 4 5 ) par (N 10 Number of Parameters, log −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 /mrad) θ( 10 Mean-Absolute Error, log Beam Deflection - MAE Scaling of ML Reconstruction of e =-0.03 ± 0.02 1 =-0.48 ± 0.07, p 0 x, MAE 1 + p 0 BD… view at source ↗
Figure 3
Figure 3. Figure 3: Regression performance for the event-by-event beam deflection reconstruction of the outgo￾ing electron using the BDTG and ASMR algorithms. 5 Updated uncertainty picture and outlook Recent updates on beam polarization error propagation for diphotons, including higher order corrections, have reduced the effect of the beam polarization uncertainty [18]. This reduction, combined with using both beam polarimete… view at source ↗
read the original abstract

Future $e^+e^-$ colliders at the Z pole place strong demands of $\frac{\delta L}{L}<10^{-4}$ on the integrated luminosity measurement. Small angle Bhabha scattering (SABS) remains the standard channel, while diphoton ($\gamma\gamma$) events provide a complementary measurement. This contribution summarizes recent work on two dominant uncertainties. First, we investigate backgrounds to the diphoton channel and find that SABS and low-invariant-mass neutral hadrons are the most significant backgrounds. A gradient boosted decision tree (BDTG) is used to classify events by particle ID. The classification results show the existing and upgraded forward tracker and luminosity calorimeter (LumiCal) designs reject neutral hadrons but only the LumiCal upgrade can reject SABS at $\frac{\delta L}{L}<10^{-4}$. Second, we solve the beam deflection bias problem on an event-by-event basis using two machine learning algorithms. A BDTG and the newly written Adaptive Symbolic Memetic Regression (ASMR) are trained on beam deflection data. ASMR outperforms BDTG and provides a reduced uncertainty of $5\times10^{-6}$ for beam deflection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes machine learning methods to reduce uncertainties in Z-pole integrated luminosity measurements at future e+e- colliders. It applies a gradient boosted decision tree (BDTG) to classify and reject backgrounds (SABS and low-mass neutral hadrons) to the diphoton channel using existing and upgraded forward tracker/LumiCal designs, claiming that only the LumiCal upgrade achieves the required δL/L < 10^{-4}. It further introduces Adaptive Symbolic Memetic Regression (ASMR) alongside BDTG to correct beam deflection bias event-by-event on simulated data, reporting that ASMR outperforms BDTG and yields a beam-deflection uncertainty of 5×10^{-6}.

Significance. If the reported background rejection and uncertainty reductions hold under realistic conditions, the work could supply concrete tools for meeting the stringent luminosity precision needed for Z-pole programs at future colliders. The introduction of ASMR as a novel regression technique is a methodological contribution that may offer advantages over standard BDTG in this domain.

major comments (1)
  1. [Abstract and results] Abstract and results sections: The quantitative claims (neutral-hadron rejection, SABS rejection at δL/L < 10^{-4}, and beam-deflection uncertainty of 5×10^{-6}) rest entirely on Monte Carlo samples whose generation, size, training/validation splits, and particle-ID response modeling are not described. No cross-checks against LEP or other existing e+e- data are reported, and no systematic variations (generator tunes, misalignment, beam-parameter uncertainties) are propagated into the final δL/L figures. This is load-bearing because the central claim is that these ML methods achieve the required precision for a future collider.
minor comments (2)
  1. [Abstract] The acronym ASMR is introduced without an explicit definition or reference to its algorithmic details at first use.
  2. [Abstract] Notation for the luminosity uncertainty δL/L is used inconsistently in the abstract without a clear definition of how it is computed from the classification or regression outputs.

Simulated Author's Rebuttal

1 responses · 1 unresolved

We thank the referee for the careful and constructive review. We address the major comment on the description and validation of our Monte Carlo results below.

read point-by-point responses

  1. Referee: [Abstract and results] Abstract and results sections: The quantitative claims (neutral-hadron rejection, SABS rejection at δL/L < 10^{-4}, and beam-deflection uncertainty of 5×10^{-6}) rest entirely on Monte Carlo samples whose generation, size, training/validation splits, and particle-ID response modeling are not described. No cross-checks against LEP or other existing e+e- data are reported, and no systematic variations (generator tunes, misalignment, beam-parameter uncertainties) are propagated into the final δL/L figures. This is load-bearing because the central claim is that these ML methods achieve the required precision for a future collider.

    Authors: We agree that the manuscript requires expanded documentation of the Monte Carlo setup to substantiate the reported performance. In the revised version we will add a dedicated methods subsection specifying the event generators and tunes employed, the total number of simulated events, the training/validation/test splits for both the BDTG and ASMR algorithms, and the detailed modeling of particle-ID response in the forward tracker and LumiCal. We will also perform and report additional systematic studies by varying generator tunes, introducing realistic misalignment scenarios, and scanning beam-parameter uncertainties, then propagating these variations into the final δL/L values. Cross-checks with LEP data are feasible only for the existing detector geometries and standard SABS/diphoton selections; we will include such comparisons where they exist. For the proposed LumiCal upgrade, however, no equivalent real data are available, so validation necessarily remains simulation-based. revision: partial

standing simulated objections not resolved
  • Direct experimental cross-checks of the upgraded LumiCal design against LEP or other existing data, since no such upgraded detector has been operated.

Circularity Check

0 steps flagged

No circularity; ML results are empirical outputs from simulated data

full rationale

The paper applies standard supervised ML (BDTG classification for particle ID and BDTG/ASMR regression for beam deflection) to Monte Carlo samples of SABS, neutral-hadron, and beam-deflection events. Reported figures (neutral-hadron rejection, SABS rejection at δL/L < 10^{-4}, and 5×10^{-6} deflection uncertainty) are direct evaluation metrics on held-out test events, not quantities defined in terms of themselves or obtained by renaming fitted parameters. No equations, self-citations, or ansatzes appear that would make any central claim tautological with its inputs. The derivation chain is therefore self-contained as conventional ML benchmarking.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities can be identified from the abstract alone; full text would be needed to audit simulation assumptions or ML hyperparameters.

pith-pipeline@v0.9.0 · 5502 in / 1169 out tokens · 125156 ms · 2026-05-13T03:09:31.042709+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    B. Madison,Integrated luminosity with 100 ppm precision, methods for √s precision of 1 ppm, and beyond standard model sensitivity using photonic events, at e +e− higgs factories,2510.18666

    work page arXiv

  2. [2]

    Abramowicz et al.,A linear collider vision for the future of particle physics, 2503.05437

    H. Abramowicz et al.,A linear collider vision for the future of particle physics, 2503.05437

    work page arXiv

  3. [3]

    Abramowicz et al.,The ILD Detector: A V ersatile Detector for an Electron-Positron Collider at Energies up to 1 TeV,2506.06030

    ILD collaboration, H. Abramowicz et al.,The ILD Detector: A V ersatile Detector for an Electron-Positron Collider at Energies up to 1 TeV,2506.06030

    work page arXiv

  4. [4]

    Abramowicz et al.,International Large Detector: Interim Design Report,2003.01116

    ILD ConceptGroupcollaboration, H. Abramowicz et al.,International Large Detector: Interim Design Report,2003.01116

    work page arXiv 2003

  5. [5]

    Behnke, J

    T. Behnke, J. E. Brau, B. Foster, J. Fuster, M. Harrison, J. M. Paterson et al.,The International Linear Collider Technical Design Report - V olume 1: Executive Summary,1306.6327

    work page arXiv

  6. [6]

    Madison and G

    B. Madison and G. Wilson,Novel position reconstruction methods for highly granular electromagnetic calorimeters,EPJ Web Conf.315(2024) 03007

    work page 2024

  7. [7]

    Jadach, W

    S. Jadach, W. Płaczek and B. F. L. Ward,BHWIDE 1.00: O(α) yfs exponentiated monte carlo for bhabha scattering at wide angles for lep1/slc and lep2,Phys. Lett. B390 (1997) 298–308

    work page 1997

  8. [8]

    Jadach, B

    S. Jadach, B. F. L. Ward, Z. Wa¸ s, S. A. Yost and A. Siodmok,Multi-photon monte carlo event generator KKMCee for lepton and quark pair production in lepton colliders,Comput. Phys. Commun.283(2023) 108556

    work page 2023

  9. [9]

    Wilson and B

    G. Wilson and B. Madison,Reimagining e +e− collider precision luminosity measurements,EPJ Web Conf.315(2024) 01024

    work page 2024

  10. [10]

    Hocker, J

    A. Hocker, J. Stelzer, H. V oss and K. V oss,TMVA - Toolkit for Multivariate Data Analysis,PoSACA T(3, 2007) 040, [physics/0703039]

    work page arXiv 2007

  11. [11]

    V outsinas, E

    G. V outsinas, E. Perez, M. Dam and P. Janot,Beam-beam effects on the luminosity measurement at LEP and the number of light neutrino species,Phys. Lett. B800 (2020) 135068, [1908.01704]

    work page arXiv 2020

  12. [12]

    WHIZARD: Simulating Multi-Particle Processes at LHC and ILC

    W. Kilian, T. Ohl and J. Reuter,WHIZARD: Simulating Multi-Particle Processes at LHC and ILC,Eur . Phys. J. C71(2011) 1742, [0708.4233]

    work page Pith review arXiv 2011

  13. [13]

    A Brief Introduction to PYTHIA 8.1

    T. Sjostrand, S. Mrenna and P. Z. Skands,A Brief Introduction to PYTHIA 8.1, Comput. Phys. Commun.178(2008) 852–867, [0710.3820]

    work page internal anchor Pith review arXiv 2008

  14. [14]

    Agostinelli et al.,GEANT4—a simulation toolkit,Nucl

    S. Agostinelli et al.,GEANT4—a simulation toolkit,Nucl. Instrum. Meth. A506(2003) 250–303

    work page 2003

  15. [15]

    Schulte,Beam-beam simulations with GUINEA-PIG,

    D. Schulte,Beam-beam simulations with GUINEA-PIG,

    work page

  16. [16]

    James and M

    F. James and M. Roos,Minuit: A System for Function Minimization and Analysis of the Parameter Errors and Correlations,Comput. Phys. Commun.10(1975) 343–367

    work page 1975

  17. [17]

    M. V . Berry and Z. V . Lewis,On the weierstrass–mandelbrot fractal function, Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences370(1980) 459–484

    work page 1980

  18. [18]

    Madison,Beam polarization precision requirements for future e +e− Higgs factories, 2511.09851

    B. Madison,Beam polarization precision requirements for future e +e− Higgs factories, 2511.09851

    work page arXiv

  19. [19]

    I. B. Jelisav ˇci´c, S. Luki´c, G. M. Dumbelovi´c, M. Pandurovi´c and I. Smiljani´c, Luminosity measurement at ilc,Journal of Instrumentation8(Aug., 2013) P08012–P08012

    work page 2013