Potential of di-Higgs observation via a calibratable jet-free $HH\to 4b$ framework

Abstract

We present a calibratable, jet-free framework that enhances the search significance of the flagship LHC channel $HH \to 4b$ by more than a factor of five compared to existing approaches. The method employs a mass-decorrelated discriminant to identify $h_1 h_2 \to 4b$ with variable $h_{1,2}$ masses and a simultaneous estimator of $(m_{h_1}, m_{h_2})$, both derived from multiclass classification on all-particle inputs. The $HH$ signal response can be calibrated using $ZZ \to 4b$. Using a highly realistic simulation framework validated through multiple tests, we demonstrate the method's robustness and identify two prerequisites essential for achieving this level of sensitivity. Results indicate that with LHC Run 2 and 3 data, observation-level sensitivity to $HH$ appears within reach, enabling constraints on $κ_λ$ comparable to HL-LHC projections and offering an accelerated path to precision measurements of the Higgs trilinear coupling.

Figures (18)

• Figure 1: Receiver operating characteristic (ROC) curves for the $HH \to 4b$ signal versus the QCD background, scaled to the original event yields under 450 $\mathrm{fb^{-1}}$. Shown are the reproductions of the CMS resolved-channel CMS:2022dwd (green) and boosted-channel CMS:2023yay (purple) strategies, enhanced resolved strategy probing the limit of jet-based approaches (blue), and our jet-free classifiers trained on the full $1.4\times 10^8$ events (red) or one-tenth of that sample (orange). True signal and background yields read from the signal-region figures in Refs. CMS:2022dwdCMS:2023yay are marked by stars. Except for the boosted case, all strategies use resolved-channel triggers. Uncertainty bands indicate Poisson statistical limits. Our approach shows significant gains over established methods.
• Figure 2: Reconstructed $(m_{h_1},\,m_{h_2})$ distributions, normalized to unity, after applying a reference discriminant threshold $D_{h_1 h_2 \to 4b} > 0.9$. Six representative processes are shown: $gg\text{F}$$HH\to 4b$, $ZZ$, $ZH$, QCD multijet, $t\overline{t}\xspace$, and $Z$+jets. The upper panels display signal-like processes, with reconstructed points forming the peak structure localized near $m_Z$ and $m_H$ along the two axes.
• Figure 3: Left: Validations of signal ($gg\text{F}$$HH \to 4b$) versus background (QCD events with $\geq$4 $b$-hadrons) separation using ROC curves at different Lorentz boosts, characterized by maximum $p_{\text{T}}$ among all $b$-hadron pair systems. Uncertainty bands show Poisson statistical limits. Right: Comparison of three generators for hadronization modeling: Pythia 8.3 (green), Herwig 7.2 (orange) and Vincia (blue) showing the $D_{h_1 h_2 \to 4b}$ selection efficiencies for $gg\text{F}$$HH \to 4b$, $ZH \to 4b$ and $ZZ \to 4b$. The upper panel uses a threshold defining our signal region, while the lower panel applies a looser threshold.
• Figure S1: Performance of the Sophon model $H\to bb$ versus QCD jets, shown as ROC curves with AUC values. The performance is directly comparable to established CMS large-$R$$X\rightarrow bb$ jet taggers under identical phase-space selection CMS:2025kje.
• Figure S2: Performance of the SophonAK4 model (v2) for (left) $b$ versus light/charm jet tagging and (right) $c$ versus light/$b$ jet tagging, shown as ROC curves with AUC values. Signal and background jets originate from $t\overline{t}$ events with jets satisfying $p_{\text{T}}\xspace > 30$ GeV and $|\eta| < 2.5$. The performance is directly comparable to established CMS flavor taggers under identical phase-space selections.