Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

tZ FCNC Case study: LLM Application in signal/Background discrimination analyses in Particle Physics

A. Saqlain, Z. E. Çabukoğlu, J. Smiesko, S. Kartal

TL;DR

The paper investigates using OpenAI's o3 as a process-agnostic tool to optimize signal/background discrimination for the FCNC top decay t → uZ at FCC-hh energies. It compares o3 to a traditional cut-based control group and TMVA BDT within a realistic MadGraph–Pythia–Delphes workflow, evaluating performance across FCC-hh, HE-LHC, and HL-LHC energy scales. TMVA BDT achieves the best overall signal significance, while o3 offers competitive, fast cuts and potential as a generalizable, rapid-analysis precursor for new-physics searches. The study highlights both the promise and energy-scale limitations of applying a non-HEP-specific LLM to high-energy physics analyses, suggesting further refinements for broader applicability.

Abstract

We present a case study exploring the potential of OpenAI's o3 model as a process-agnostic tool (within fixed topology) for predicting optimized selection cuts in high-energy physics analyses. Specifically, we investigate the effectiveness of the model in separating signal from relevant Standard Model backgrounds in the context of Flavour-Changing Neutral Current (FCNC) top-quark couplings, focusing on the rare decay process t \rightarrow uZ. The study is performed at the Future Circular Collider in hadron-hadron mode (FCC-hh) setup. We prompt the o3 model on detector level data for signal and background processes to predict selection cuts that enhance Signal-to-Background (S/B) discrimination. A comparative analysis is then carried out between the efficiencies resulting from o3-predicted cuts, TMVA BDT separation and those derived from traditional manually designed strategies used by a control group; all utilizing the same parameters for the sake of the integrity of the study. Our results demonstrate that the o3 model performs a degree better than the control group, suggesting its promise as a fast and generalizable tool for new physics searches. Meanwhile, BDT results were considerably higher than both the o3 and the traditional cut-based methods. Furthermore, in order to test its limitations, o3 cuts were applied to data generated via the approved High-Luminosity Large Hadron Collider (HL-LHC) and the proposed High-Energy LHC (HE-LHC) setup in order to examine its effectiveness at different energy scales when provided with data at FCC energies.

tZ FCNC Case study: LLM Application in signal/Background discrimination analyses in Particle Physics

TL;DR

The paper investigates using OpenAI's o3 as a process-agnostic tool to optimize signal/background discrimination for the FCNC top decay t → uZ at FCC-hh energies. It compares o3 to a traditional cut-based control group and TMVA BDT within a realistic MadGraph–Pythia–Delphes workflow, evaluating performance across FCC-hh, HE-LHC, and HL-LHC energy scales. TMVA BDT achieves the best overall signal significance, while o3 offers competitive, fast cuts and potential as a generalizable, rapid-analysis precursor for new-physics searches. The study highlights both the promise and energy-scale limitations of applying a non-HEP-specific LLM to high-energy physics analyses, suggesting further refinements for broader applicability.

Abstract

We present a case study exploring the potential of OpenAI's o3 model as a process-agnostic tool (within fixed topology) for predicting optimized selection cuts in high-energy physics analyses. Specifically, we investigate the effectiveness of the model in separating signal from relevant Standard Model backgrounds in the context of Flavour-Changing Neutral Current (FCNC) top-quark couplings, focusing on the rare decay process t \rightarrow uZ. The study is performed at the Future Circular Collider in hadron-hadron mode (FCC-hh) setup. We prompt the o3 model on detector level data for signal and background processes to predict selection cuts that enhance Signal-to-Background (S/B) discrimination. A comparative analysis is then carried out between the efficiencies resulting from o3-predicted cuts, TMVA BDT separation and those derived from traditional manually designed strategies used by a control group; all utilizing the same parameters for the sake of the integrity of the study. Our results demonstrate that the o3 model performs a degree better than the control group, suggesting its promise as a fast and generalizable tool for new physics searches. Meanwhile, BDT results were considerably higher than both the o3 and the traditional cut-based methods. Furthermore, in order to test its limitations, o3 cuts were applied to data generated via the approved High-Luminosity Large Hadron Collider (HL-LHC) and the proposed High-Energy LHC (HE-LHC) setup in order to examine its effectiveness at different energy scales when provided with data at FCC energies.

Paper Structure

This paper contains 13 sections, 9 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Theoretical Background
  3. Event generation
  4. Methodology
  5. Analysis
  6. Control Group
  7. TMVA BDT
  8. o3 model
  9. Comparative Analysis
  10. Dependence on Center-of-Mass Energy Scale
  11. Conclusions
  12. Appendix
  13. Appendix

Figures (5)

  • Figure 1: Feynman diagrams representing the $ug \to tZ$ production and decay via the tZq FCNC anomalous couplings
  • Figure 2: The variation of the cross section $\sigma$ (in pb) with respect to the FCNC couplings $\kappa_{tqZ}$ (top row) and $\lambda_{tqZ}$ (bottom row) is illustrated for three collider scenarios: HL-LHC (left), HE-LHC (center), and FCC-hh (right). All results are obtained via basic cuts of $p_T^j > 40$ GeV and $|\eta_j| < 2.5.$
  • Figure 3: A workflow schematic for this analysis
  • Figure 4: Left: The ROC curve for the BDT model with Background rejection and signal efficiency at its axises. Right: Training error that corresponded to said BDT model
  • Figure :