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

This paper presents novel machine learning methods, specifically a Gradient Boosted Decision Tree and a new Adaptive Symbolic Memetic Regression algorithm, to mitigate dominant backgrounds and beam deflection biases in Small Angle Bhabha scattering and diphoton channels, thereby enabling future $e^+e^-$ colliders to achieve the stringent integrated luminosity precision of $\delta L/L < 10^{-4}$ required at the Z pole.

The Gist

Imagine a future particle accelerator as a massive, ultra-precise factory. Its job is to smash electrons and positrons together to study the "Z boson," a fundamental particle that acts like a ruler for the universe's laws. To get a perfect reading from this ruler, the factory needs to count exactly how many collisions happen. This count is called integrated luminosity.

The paper argues that to get a truly perfect measurement, the factory needs to be accurate to within one part in ten thousand. Currently, the tools used to count these collisions have a few "bugs" that make the count slightly fuzzy. The author, Brendon Madison, uses two new types of "smart software" (Machine Learning) to fix these bugs.

Here is a breakdown of the two main problems and the solutions, explained with everyday analogies:

1. The "Fake Photon" Problem (Identifying the Right Particles)

The Problem:
To count the collisions, the detectors look for specific events. One method looks for "Small-Angle Bhabha Scattering" (SABS), which is like spotting two billiard balls bouncing off each other at a very shallow angle. Another method looks for "Diphoton" events, which are like spotting two flashes of light.

However, the detectors sometimes get confused.

• The Mix-up: Sometimes, a neutral hadron (a type of heavy, invisible particle) sneaks in and looks exactly like a flash of light (a photon). It's like a person wearing a perfect disguise walking into a room of photographers; the cameras can't tell they aren't a real celebrity.
• The Old Solution: The current detector design (called ILD) is like a standard security camera. It's good, but it still lets some of these "imposters" through, messing up the count.
• The New Solution: The author tested an upgraded detector (called GLIP) which is like a high-definition, 3D scanner. They used a smart algorithm called a BDTG (a type of decision tree that asks a series of "yes/no" questions) to sort the particles.
  • The Result: The old camera (ILD) still struggles to tell the difference between the real light and the imposters. But the new 3D scanner (GLIP) is so sharp it can spot the imposters and kick them out. This reduces the error significantly, but only if the detector is upgraded first.

2. The "Magnetic Wind" Problem (Beam Deflection)

The Problem:
When the electron and positron beams crash, they don't just bounce off; they create a tiny, invisible "wind" of electromagnetic force. This wind pushes the particles slightly off their intended path, like a strong gust of wind blowing a kite sideways.

• The Old Way: Previously, scientists tried to fix this by calculating the average wind speed for the whole factory and applying one big correction. It's like trying to fix a wobbly table by guessing the average height of the floor and shimming all legs equally. It helps, but it's not perfect because every single "kite" (collision) gets pushed differently.
• The New Way: The author used two new AI tools to fix this on a per-event basis.
  1. BDTG: A standard smart algorithm.
  2. ASMR: A brand-new, custom-built algorithm that acts like a detective trying to find a mathematical formula (a "symbolic" solution) rather than just guessing. It's like a detective who doesn't just say "the wind was strong," but figures out the exact physics equation describing the wind for that specific moment.

The Result:
The new "detective" (ASMR) was much better than the standard smart algorithm. It could predict exactly how much each individual particle was pushed by the wind.

• The Improvement: The old method left a "fuzziness" (uncertainty) of about 80 parts in a million. The new ASMR method reduced this to just 5 parts in a million. It's like going from measuring a table's height with a ruler to measuring it with a laser.

The Bottom Line

The paper concludes that to reach the ultra-precise measurements needed for future physics:

1. Hardware Upgrade is Mandatory: You cannot just use software to fix the "fake photon" problem; you physically need the upgraded, high-detail detector (GLIP) to see the difference.
2. Smart Software is a Game Changer: Using the new AI (ASMR) to correct the "magnetic wind" on a case-by-case basis makes the measurement much sharper than the old "average" method.

By combining the upgraded hardware with these new AI tools, the factory can finally count its collisions with the extreme precision required to unlock new secrets of the universe. Without these steps, the measurements will remain too "fuzzy" to be useful for the most advanced physics experiments.

Technical Summary

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

Problem Statement
Future $e^+e^-$ colliders operating at the Z pole require an integrated luminosity measurement precision of $\delta L/L < 10^{-4}$ to enable high-precision electroweak observables. This study focuses on the International Large Detector (ILD) design, potentially with a Granular Long Instrument for Precision (GLIP) upgrade, and considers linear collider scenarios with specific beam polarizations. Two dominant sources of uncertainty hinder this precision:

1. Event Identification and Backgrounds: Specifically, the contamination of the diphoton ($\gamma\gamma$) channel by Small-Angle Bhabha Scattering (SABS) and low-invariant-mass neutral hadrons.
2. Beam Deflection Bias: The electromagnetic deflection of outgoing $e^+e^-$ pairs by the bunch fields, which biases the polar angle measurement and consequently the luminosity calculation if not corrected on an event-by-event basis.

Methodology
The study employs two machine learning algorithms to address these uncertainties:

• Gradient Boosted Decision Tree (BDTG): Utilizing ROOT's TMVA library, trained on kinematic, spatial, and cluster observables from the forward calorimeter and tracker.
• Adaptive Symbolic Memetic Regression (ASMR): A newly developed algorithm for this study. Unlike BDTG, ASMR is a symbolic memetic regression method capable of learning analytic solutions. It utilizes a global evolutionary scheme combined with local optimization (e.g., Minuit) and adaptive data partitioning.

The analysis involves:

• Background Simulation: Using WHIZARD for event generation and Pythia for parton kinematics, with custom GEANT4 code added to handle low-invariant-mass hadronization (specifically radiative hadrons $e^+e^- \to q\bar{q}\gamma$).
• Particle ID Classification: Training classifiers to distinguish photons, electrons, positrons, and neutral hadrons.
• Beam Deflection Regression: Training models on reconstructed energies and polar angles $(E_+, E_-, \theta_+, \theta_-)$ to predict the electron's polar angle deflection. The ground truth is derived from the event generator level (BHWIDE) before beam deflection simulation (GUINEA-PIG).
• Benchmarking: Both algorithms were first tested on the Weierstrass-Mandelbrot (W-M) fractal function to evaluate Mean Absolute Error (MAE) scaling.

Key Contributions and Results

1. Background Analysis and Particle ID:

   • The study identifies SABS and low-invariant-mass neutral hadrons (from radiative hadrons) as the primary backgrounds for the diphoton channel.
   • Classification Performance: Both the existing ILD LumiCal and the upgraded GLIP LumiCal designs effectively reject neutral hadrons. However, only the GLIP LumiCal upgrade achieves sufficient rejection of SABS to meet the $\delta L/L < 10^{-4}$ requirement.
   • Impact on Precision: For the ILD LumiCal, the dominant uncertainty remains misidentifying Bhabha events as diphotons ($35 \times 10^{-4}$). For the GLIP LumiCal, SABS contamination is reduced to $\approx 1 \times 10^{-6}$, making neutral hadron contamination the leading residual effect ($6 \times 10^{-6}$).

2. Beam Deflection Correction:

   • Traditional corrections model the mean deflection, resulting in a luminosity uncertainty contribution of $\approx 8 \times 10^{-5}$.
   • Event-by-Event Prediction: Both BDTG and ASMR were trained to predict deflection on an event-by-event basis.
   • ASMR Superiority: ASMR outperformed BDTG in raw MAE and scaling with parameter count. It successfully identified $\frac{\theta_- - \theta_+}{2}$ as a leading-order predictor for deflection.
   • Precision Gain: The ASMR algorithm achieved a residual deflection uncertainty ($\sigma_{def}$) of $\approx 50$ $\mu$rad, reducing the beam deflection effect on integrated luminosity to $5 \times 10^{-6}$. This is significantly more precise than the mean-based method.

3. Updated Uncertainty Budget:

   • Combining the GLIP LumiCal design with the ASMR-based beam deflection correction yields a total integrated luminosity uncertainty for the diphoton channel of $5.7 \times 10^{-4}$ (dominated by statistics and angular resolution), with the beam deflection contribution reduced to negligible levels compared to other factors.
   • The combined measurement of SABS and diphotons allows for cross-verification of models, potentially outperforming individual measurements.

Significance and Claims
The paper claims that the combination of a granular calorimeter upgrade (GLIP) and novel machine learning techniques (specifically ASMR) provides a viable path to reducing integrated luminosity uncertainties.

• Detector Design: The results strongly motivate the upgrade to a highly granular LumiCal (GLIP) to effectively reject SABS contamination in diphoton measurements.
• Algorithmic Advantage: The event-by-event correction of beam deflection using ASMR is demonstrated to be more precise than traditional mean-based modeling, reducing the associated uncertainty by an order of magnitude.
• Limitations and Outlook: The authors remain modest regarding the ultimate goal. They note that angular resolution and beam polarization precision remain dominant outstanding uncertainties not addressed in this work. Consequently, it is currently unknown whether future colliders can surpass the $1 \times 10^{-4}$ threshold without further improvements in these areas. The paper concludes that funding for further LumiCal development and precision studies is essential to realize the planned Z pole electroweak measurements.