New Physics Searches at the LHC through Event-based Anomaly Detection and Development of ADFilter Web-tool

TL;DR

The paper advocates model-agnostic, event-level anomaly detection at the LHC using unsupervised autoencoders trained on SM-like data to identify unusual events, thereby complementing targeted BSM searches. It details ATLAS Run 2 results showing that anomaly regions can substantially boost sensitivity in the sub-TeV regime and presents ADFilter, a web tool that reproduces the ATLAS pipeline for rapid reinterpretation via the acceptance $A = \frac{N_{\mathrm{AR}}}{N_{\mathrm{tot}}}$. The work compares anomaly detection with supervised approaches in di-Higgs searches, highlighting where each method excels depending on signal hypothesis alignment. Overall, the ADFilter framework provides a practical pathway to reinterpret existing limits and pursue broader, model-agnostic explorations of new physics at the LHC.

Abstract

This work presents advancements in model-agnostic searches for new physics at the Large Hadron Collider (LHC) through the application of event-based anomaly detection techniques utilizing unsupervised machine learning. We discuss the advantages of the anomaly detection approach, as demonstrated in a recent ATLAS analysis, and introduce ADFilter, a web-based tool designed to process collision events using autoencoders based on deep unsupervised neural networks. ADFilter calculates loss distributions for input events, aiding in determining the degree to which events can be considered anomalous. Real-life examples are provided to demonstrate how the tool can be used to reinterpret existing LHC results, with the goal of significantly improving exclusion limits. Furthermore, we present a comparative study between anomaly detection and supervised machine learning techniques, using the search for heavy resonances decaying into two or more Higgs bosons as a representative case to demonstrate the application and effectiveness of these methods.

Figures (3)

• Figure 1: A schematic diagram showing the workflow of the ADFilter.
• Figure 2: (a) Observed 95% credibility-level limits ATLAS2020 on $\sigma \times B$ for $tbH^{+}$ production as a function of $M_X$, using dijet masses in events with at least one isolated lepton with $p_{T}>60$ GeV. (b) ATLAS 95% limits for Gaussian signals from Ref. ATLAS:2023ixc, rescaled using $(1/A_{\rm sel}\times\varepsilon)$ following Ref. ATLAS2020. Magenta points show the cross section after the ADFilter correction. The anomaly-detection interpretation excludes the $\tan\beta=1$ benchmark up to 1.35 TeV, whereas the original selection could not.
• Figure 3: Comparison of significance values for (a) $X\rightarrow HH$ and (b) $X\rightarrow SH$ with $S\rightarrow HH$ for both supervised cluster method (SC) and unsupervised machine learning based anomaly detection(AD) method. The definitions of signal and background yields, as well as the treatment of uncertainties, follow Ref. Chekanov:2025xpk.