Weakly supervised anomaly detection for resonant new physics in the dijet final state using proton-proton collisions at $\sqrt{s}=13$ TeV with the ATLAS detector

TL;DR

This ATLAS study develops a model-agnostic anomaly-detection framework for narrow dijet resonances in $pp$ collisions at $13$ TeV, using $139\,\mathrm{fb}^{-1}$ of data. It employs two background strategies, SALAD and CURTAINs, within a weakly supervised (CWoLa) classification framework to identify overdensities in the $m_{\mathrm{JJ}}$ spectrum, while leveraging jet-substructure observables. A four-step pipeline—region definition, background templating, weakly supervised classification, and likelihood-based inference—yields model-agnostic results showing no significant local excesses, along with injections of representative signal models to quantify sensitivity. The work extends prior ATLAS searches by enabling higher-dimensional feature usage, providing complementary signal-model coverage, and delivering competitive upper limits across a broad set of resonance scenarios.

Abstract

An anomaly detection search for narrow-width resonances beyond the Standard Model that decay into a pair of jets is presented. The search is based on 139 fb$^{-1}$ of proton-proton collisions at $\sqrt{s}=13$ TeV recorded during 2015-2018 with the ATLAS detector at the Large Hadron Collider. The analysis is optimized without a particular signal model and aims to be sensitive to a broad range of new physics. It uses two different machine learning strategies to estimate the background in different signal regions. In each region, a weakly supervised classifier is trained to distinguish this background estimate from data. The analysis focuses on events with high transverse momentum jets reconstructed as large-radius jets. The mass and substructure of these jets are used as inputs to the classifiers. After a classifier-based selection, the distribution of the invariant mass of the two jets is used to search for potential local excesses. The model-independent results of both the anomaly detection methods show no signs of significant local excesses. In addition to model-independent results, a representative set of signal models is injected into the data, and the sensitivity of the methods to these scenarios is reported.

Figures (11)

• Figure 1: A schematic drawing of the analysis strategy with the consecutive steps shown in the different panels. The distribution shows the data which consists of a background and a potential signal which is localized on the $m_{\mathrm{JJ}}$ spectrum and (potentially) different from the background in other features, here collectively labeled as $\mathcal{T}$. The first panel (a) shows how the SR and SB regions are defined on the $m_{\mathrm{JJ}}$ spectrum. The second panel (b) shows how the background is estimated in the SR. Then a classifier is trained between the estimated background and the data in the SR. The third panel (c) shows the data after a cut is applied on the classifier output of each event, thereby increasing the relative contribution of a potential signal. After this step, a bump hunt is performed in the $m_{\mathrm{JJ}}$ spectrum. The whole process is repeated with shifted regions SR and SB until the entire $m_{\mathrm{JJ}}$ spectrum is covered.
• Figure 2: Histograms of $m_{\mathrm{JJ}}$ in the first set of non-overlapping $m_{\mathrm{JJ}}$ SRs for the CURTAINs method on all feature sets at the $\epsilon=0.1$ classifier selection using the $|\Delta Y|\xspace < 1.2$ data. The figures show the resulting distributions when the analysis is evaluated on the first, third, fifth, and seventh $m_{\mathrm{JJ}}$ SRs. Shown are different feature sets: (a) is the result of $\mathcal{T}=M\xspace$, (b) is the result of $\mathcal{T}=M,\tau_{21}\xspace$ and (c) is the result of $\mathcal{T}=M,\tau_{21},\tau_{32}\xspace$. The fit is derived from the background-only fit interpolated from the SBs. The uncertainties in the observed counts include the Poisson statistical uncertainty of the bin counts. The uncertainties in the fit are represented by the dashed histograms and include the uncertainties in the fit parameters and the uncertainty from the classifier ensemble on the data. The vertical dashed lines mark the edges of each SR in $m_{\mathrm{JJ}}$. The lower panel in each plot shows the Gaussian-equivalent significance of the deviation between the fit and data.
• Figure 3: Observed significances ($Z$) for SALAD at the two different selections, (a) $\epsilon=0.1$ and (b) $\epsilon=0.02$. The significances are shown for all feature sets and $m_{\mathrm{JJ}}$ SRs.
• Figure 4: Observed significances ($Z$) for CURTAINs at the two different selections, (a) $\epsilon=0.1$ and (b) $\epsilon=0.02$. The significances are shown for all feature sets and $m_{\mathrm{JJ}}$ SRs.
• Figure 5: Signal injection tests with $\mathcal{T}=M,\tau_{21}\xspace$ at $\epsilon=0.1$ in (a) and $\epsilon=0.02$ in (b) for both the SALAD and CURTAINs for all simulated signal models. The signal models are described in detail in \ref{['sec:data']}. The local observed significance ($Z$) is shown as reported by the analysis pipeline after running the analysis with $3\sigma$ of signal injected into the data in the $m_{\mathrm{JJ}}$ SR centered on the different signals. The errors show one standard deviation on the reported significance as calculated by bootstrapping the signal injection procedure.