Novel Machine Learning Methods to Improve Z Pole Integrated Luminosity at Future Colliders
\firstname Brendon \lastname Madison \fnsep · 2026
Read Novel Machine Learning Methods to Improve Z Pole Integrated Luminosity at Future Colliders by \firstname Brendon \lastname Madison \fnsep (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-12407.txt.
Introduction
Madison reviews two machine-learning attacks on the toughest errors in future Z-pole luminosity measurement. Small-angle Bhabha scattering stays the standard channel, while photon-pair events offer a second route, and both suffer from background and beam effects. A boosted decision tree sorts events by particle type, and only an upgraded luminosity calorimeter rejects the scattering background at the required precision. For beam deflection bias, a new symbolic regression method beats the boosted tree and cuts that uncertainty to a few parts per million.
Catalog record
| CREATOR | \firstname Brendon \lastname Madison \fnsep |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 17749 |
| LICENSE | http://creativecommons.org/publicdomain/zero/1.0/ |
| RIGHTS | CC0 |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.12407
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Novel Machine Learning Methods to Improve Z Pole Integrated Luminosity at Future Colliders \firstname Brendon \lastname Madison \fnsep Abstract Future e + e − e^{+}e^{-} colliders at the Z pole place strong demands of δ ℒ / ℒ < 10 − 4 \delta\mathcal{L}/\mathcal{L}<10^{-4} on the integrated luminosity measurement. Small angle Bhabha scattering (SABS) remains the standard channel, while diphoton ( γ γ \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 δ ℒ / ℒ < 10 − 4 \delta\mathcal{L}/\mathcal{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 × 10 − 6 5\times 10^{-6} for beam deflection. 1 Introduction Precision measurements at the Z pole remain a clear physics case for future e^{+}e^{-} colliders. A Z pole run can deliver improvements in electroweak observables only if the absolute normalization of the data sample, the integrated luminosity, is controlled below the 10^{-4} level 15 ; 1 . We restrict this study in two ways: we consider the case of a linear collider with (|P_{e-}|,|P_{e+}|)=(80\%,30\%) beam polarization, such as the International Linear Collider (ILC) or the linear collider at CERN, and we use the International Large Detector (ILD) design 2 ; 3 ; 5 ; 1 . We also consider an upgrade of the ILD forward region to use a highly granular calorimeter, referred to as the Granular Long Instrument for Precision (GLIP), as done in previous studies 13 ; 15 . Small-angle Bhabha scattering (SABS) remains the standard luminosity channel due to its large cross-section and well established precision theory calculation 8 ; 9 . Diphoton events, e^{+}e^{-}\to\gamma\gamma , are proposed to provide a complementary measurement, but with different detector and beam uncertainties and a comparably simple theory calculation 15 ; 19 . Both are restricted to the forward region, particularly the Luminosity Calorimeter (LumiCal) and forward tracker (FTD), as this leads to better uncertainties in terms of detector energy calibration and metrology 15 . This paper explores two dominant uncertainties: event identification and beam electromagnetic deflection. To optimize these, this paper deploys two machine-learning algorithms: ROOT TMVA’s Gradient Boosted Decision Tree (BDTG), and Adaptive Symbolic Memetic Regression (ASMR), which has been developed for this study 7 . Figure 1: Forward region layout used in this study, based on the ILD forward region used at ILC 3 . 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. 13 ; 15 For a more in-depth introduction to integrated luminosity and its measurement at future e^{+}e^{-} colliders, especially at the Z pole, we recommend previous studies 15 ; 18 . A diagram of the forward instrumentation can be seen in Figure 1 . We have extended the existing ILD FTD to the inner acceptance of the LumiCal with what we refer to as the inner FTD (iFTD) so that we can asses the performance of forward tracking over the entire acceptance of the LumiCal, instead of the existing design where the FTD acceptance would only include wider angles. The ILD LumiCal, with approximately 1X_{0} per layer, and the GLIP LumiCal, with \frac{1}{6}X_{0} per layer, are investigated as LumiCal candidates. The upgrade to a more granular design is motivated by previous studies that show improved energy calibration, position resolution, and particle identification 15 ; 13 . 2 Backgrounds to SABS and diphotons For SABS, four-fermion production is the main contamination but, as reported in previous studies, this can be treated as a correction to Bhabha scattering 15 . For diphotons the main contamination is from Bhabha scattering. The event identification for diphotons and Bhabha scattering was previously done, using a BDTG with both particle identification, detector and reconstruction information. This previous study found that the residual bias of Bhabha scattering contamination on the diphoton integrated luminosity precision is 35\times 10^{-4} for the ILD LumiCal with five layer forward tracker and 0.01\times 10^{-4} for the GLIP LumiCal with five layer forward tracker 15 . We explored numerous other possible background processes, specifically for the Z pole and diphotons. A new background process, radiative hadrons ( e^{+}e^{-}\to q\bar{q}\gamma ), where the photon is an initial state radiation (ISR) photon near beam energy and in the forward region. The recoiling hadronic system is then boosted into the opposing forward region, with an invariant mass \lessapprox 2~\mathrm{GeV} , placing the hadronization in the realm of low energy e^{+}e^{-} colliders. In this regime, radiative return to \omega(\mathrm{782}) or the production of \pi^{0}\eta is common, decaying to narrow and photonic final states. These can then be confused with a single photon in the LumiCal. Using WHIZARD and the cuts from a previous integrated luminosity study, we found a cross-section of \approx 3 pb, compared to the \approx 100 pb cross-section of diphotons 12 . No standard simulation chain is reliable in this regime, as neither WHIZARD nor Pythia are designed to handle low invariant mass quark production and hadronization 12 ; 17 . To alleviate this, we used WHIZARD for event generation, followed by Pythia to handle parton kinematics 12 ; 17 . Existing models and data from low energy e^{+}e^{-} colliders were used to hadronize partons in custom code that was added to GEANT4 to allow for the production of final states from hadronic decays 4 . We find that \approx 20% of the q\bar{q}\gamma events in the forward region can be neutral hadrons faking photons. These rates should be interpreted as order-of-magnitude estimates, as this hadronization simulation is not precise. Table 1: Particle ID confusion matrix, in percentages, for ILD LumiCal and 5 layer forward tracker. True \rightarrow \gamma hadrons e^{-} e^{+} Output \downarrow \gamma 96.88 0.48 1.38 1.26 hadrons 0.06 99.78 0.1 0.06 e^{-} 0.1 0.02 98.62 1.26 e^{+} 0.14 0.02 1.16 98.68 Table 2: Particle ID confusion matrix, in percentages, for GLIP LumiCal and 5 layer forward tracker. True \rightarrow \gamma hadrons e^{-} e^{+} Output \downarrow \gamma 98.22 0.14 0.72 0.92 hadrons 0.02 99.80 0.12 0.06 e^{-} 0.16 0.02 99.24 0.58 e^{+} 0.02 0.02 0.66 99.30 A BDTG was trained on kinematic, spatial, and cluster observables from the forward calorimeter and tracker to classify photons, electrons, positrons, and neutral hadrons 7 . It does not perform event identification as done in a previous study 15 . The results, seen in Tables 2 and 2 , indicate similar neutral hadron and photon separation for both LumiCal designs. The neutral hadron contamination affects integrated luminosity precision by 2\times 10^{-5} for the ILD LumiCal and 6\times 10^{-6} for the GLIP LumiCal. For the ILD LumiCal, the leading effect is still from misidentifying Bhabha