Enhancing Sensitivity for Di-Higgs Boson Searches Using Anomaly Detection and Supervised Machine Learning Techniques

TL;DR

This work addresses improving sensitivity to heavy resonances decaying to di-Higgs final states at the LHC. It compares a model-agnostic unsupervised autoencoder for anomaly detection with a supervised binary classifier, both fed by identical input representations of event kinematics. Across two signal topologies, X→HH and X→SH, the study demonstrates that both methods can boost discovery significance, with the classifier performing best at lower masses for HH and anomaly detection offering robust performance at higher masses and in the SH channel, highlighting the complementary nature of the approaches. The findings advocate incorporating anomaly-detection techniques as a general, BSM-agnostic tool in di-Higgs searches, potentially reducing dependence on specific BSM models, and show practical applicability with public data and code resources.

Abstract

This paper explores different strategies for enhancing sensitivity to new heavy resonances that decay into two or more Higgs bosons. This is achieved using two neural network architectures: an unsupervised autoencoder for anomaly detection and a supervised classifier. The autoencoder is trained on a small fraction of Standard Model (SM) Monte Carlo simulated events to calculate the loss distribution for input events, aiding in determining the extent to which events can be considered anomalous. The supervised classifier uses the same inputs but is trained on events simulated using both beyond Standard Model (BSM) and SM processes. By applying selection cuts to the output scores, we compare the sensitivities of the two approaches.

Figures (7)

• Figure 1: Representative Feynman diagrams for: (a) the direct decay $gg \!\to\! X\!\to\! H\,H$; and (b) the cascade decay $gg \!\to\! X\!\to\! S\,H \!\to\! H\,H\,H$.
• Figure 2: The $M_{bb}$ invariant mass for $X\rightarrow HH$ and $X\rightarrow SH$ processes before applying ML algorithms. The cross section for the latter process was set to that of $X\rightarrow HH$. Different shades of the gray color show the BSM signals with different masses of $X$, i.e. 0.5, 0.7, 1, 1.5 and 2 TeV. The larger the mass of $X$, the smaller area under the shaded histograms is expected.
• Figure 3: Scores from the SC for the processes $X\rightarrow HH$ (a) and $X\rightarrow SH$ (b). The vertical lines show the threshold values used to select signal events.
• Figure 4: Loss values for the anomaly detection used in the processes $X\rightarrow HH$ (a) and $X\rightarrow SH$ (b). The vertical red lines show the threshold values used to select anomalous events using the $3\sigma$ rule based on the background events (dashed black lines).
• Figure 5: Distributions for $X\rightarrow HH$ after applying the SC (a) and AD (b) selections.