Signal-Aware Contrastive Latent Spaces for Anomaly Detection

Abstract

High-dimensional feature spaces in particle physics events pose a fundamental challenge to density-estimation-based weakly supervised anomaly detection, whose fidelity degrades rapidly with an increasing number of dimensions. We propose a signal-aware latent space construction using supervised contrastive learning trained on simulated Standard Model backgrounds and a diverse set of hypothesized Beyond the Standard Model (BSM) signals. The resulting latent space is low-dimensional, regularized, and signal-sensitive, enabling high-fidelity density estimation for downstream weakly supervised anomaly detection. We demonstrate the approach in a diphoton final state, testing sensitivity across a broad range of BSM scenarios including supersymmetry models, extended Higgs sectors, heavy neutral resonances, and flavor-changing neutral currents. For signals represented in the contrastive training data, the method can elevate discovery sensitivity from previously inaccessible levels to the discovery regime. Critically, the approach retains sensitivity to BSM models not present during training: interpolation and extrapolation to unseen signal topologies yield substantial improvements in expected significance compared to a background-only baseline. By bridging supervised latent space embedding with weakly supervised anomaly detection, this strategy offers a viable path toward anomaly detection in high-dimensional feature spaces at the LHC and beyond.

Figures (8)

• Figure 1: One-dimensional projections of the six latent space features for the configuration using the pseudo-data in the . The distribution of the background from the pseudo-data (blue outline), the generated background from the (blue filled), and $(\chi_1^{\pm}\chi_2^{0})_{200}$ signal (red) are compared. Background refers to both the non-resonant and resonant Higgs components.
• Figure 2: $t$-SNE visualizations of the contrastive latent space for three embedding configurations: \ref{['fig:tsne_all']} trained with all signal and background processes (), \ref{['fig:tsne_holdout_zp']} with the $Y_{400} \to H\gamma$ mass point held out for interpolation (), and \ref{['fig:tsne_bkg']} trained with background processes only. In \ref{['fig:tsne_all']} and \ref{['fig:tsne_bkg']} "Background" refers to SM non-resonant background and SM Higgs processes, in \ref{['fig:tsne_holdout_zp']} "Other Processes" includes all physics processes other than $Y_m \to H\gamma$.
• Figure 3: Sculpting tests for the configuration without signal injection. \ref{['fig:sculpting_roc']} curve of a classifier trained to distinguish -generated background from true background in the , yielding an of 0.504 $\pm$ 0.002. \ref{['fig:myy_shape']} Normalized $m_{\gamma\gamma}$ spectrum before (blue) and after (black) a cut on the classifier output. The uncertainties include statistical effects and uncertainties from the initialization and training of the ML models.
• Figure 4: as a function of the injected signal strength for the configuration. Two classifier working points are shown: $\varepsilon_B = 0.5\%$ (matching the working point of Ref. haxadv1) and $\varepsilon_B = 0.1\%$. \ref{['fig:sic_haxad_wnh150']} Chargino--neutralino signal $(\chi_1^{\pm}\chi_2^{0})_{150}$. \ref{['fig:sic_haxad_xsh']} Extended Higgs sector signal $X \to S(\ell\ell)H$. The uncertainty bands include statistical effects and uncertainties from the initialization and training of the ML models.
• Figure 5: as a function of the injected signal strength for signals not included in the embedding training. Results are shown for the , , and background-only configurations. \ref{['fig:sic_ip_zphyy']} Heavy neutral resonance $Y_{400} \to H\gamma$. \ref{['fig:sic_ip_wnh200']} Chargino--neutralino $(\chi_1^{\pm}\chi_2^{0})_{200}$. The uncertainty bands include statistical effects and uncertainties from the initialization and training of the ML models.