Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interference Effects in Resonant Standard Model di-Higgs Production and Decay into $4b$ Final States: the Role of Machine Learning Analysis

A. Hammad, S. Moretti, A. P. Przybyl, H. Waltari

Abstract

The final state with four $b$-quarks has generally the largest event rate in Standard Model (SM)-like Higgs ($h_{\rm SM}$) pair production, but also the largest backgrounds. We study such a final state using the $gg\to h_{\rm SM}h_{\rm SM}$ production mechanism and Benchmarks Points (BPs) derived from the Next-to-Minimal Supersymmetric SM (NMSSM) in the boosted case, leading to two (fat) 'Higgs jets'. To suppress the backgrounds we use a combination of both kinematical cuts and jet substructure features exploiting Machine Learning (ML) analysis. We simulate the signal BPs both with and without the interference of the resonant $s$-channel diagram with the non-resonant topologies emerging from both the SM and NMSSM. The ML architecture of choice here is based on a multi-modal Transformer, which performs significantly better than traditional ML algorithms, in two respects: firstly, it enables to achieve higher significances and, secondly, it adapts better to the analysis dataset with interferences even if it was trained on one without these. However, neglecting the effect of the latter in experimental searches could lead to grossly mistaken results.

Interference Effects in Resonant Standard Model di-Higgs Production and Decay into $4b$ Final States: the Role of Machine Learning Analysis

Abstract

The final state with four -quarks has generally the largest event rate in Standard Model (SM)-like Higgs () pair production, but also the largest backgrounds. We study such a final state using the production mechanism and Benchmarks Points (BPs) derived from the Next-to-Minimal Supersymmetric SM (NMSSM) in the boosted case, leading to two (fat) 'Higgs jets'. To suppress the backgrounds we use a combination of both kinematical cuts and jet substructure features exploiting Machine Learning (ML) analysis. We simulate the signal BPs both with and without the interference of the resonant -channel diagram with the non-resonant topologies emerging from both the SM and NMSSM. The ML architecture of choice here is based on a multi-modal Transformer, which performs significantly better than traditional ML algorithms, in two respects: firstly, it enables to achieve higher significances and, secondly, it adapts better to the analysis dataset with interferences even if it was trained on one without these. However, neglecting the effect of the latter in experimental searches could lead to grossly mistaken results.

Paper Structure

This paper contains 8 sections, 7 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. The NMSSM Signal Scenarios
  3. ML Analysis
  4. Data Preprocessing
  5. Signal and Background Events Generation
  6. Cut-based Analysis
  7. Results
  8. Conclusions

Figures (7)

  • Figure 1: Left: Invariant mass distribution of the heavy mediator reconstructed at parton level, normalized to the expected number of events at an integrated luminosity of $1000~\mathrm{fb}^{-1}$. Middle: Invariant mass distribution of the heavy mediator reconstructed from the fat jet final state, normalized to the expected number of events at an integrated luminosity of $1000~\mathrm{fb}^{-1}$. Right: Poisson likelihood per bin quantifying the difference between the two distributions shown on the left, including systematic uncertainties of $5\%$ (black) and $20\%$ (red).
  • Figure 2: Accumulated average transverse momentum of hadrons for the leading jet, after preprocessing, computed over 50,000 events.
  • Figure 3: Kinematic distributions used for training.
  • Figure 4: Kinematic variables used in the preselection cuts.
  • Figure 5: The $4b$ invariant mass before (left) and after (right) the preselection cuts.
  • ...and 2 more figures