Seminarium "Teoria cząstek elementarnych i kosmologia"

Conventional searches for new physics derive their event selection from a specific signal hypothesis. Anomaly detection reverses this logic: the selection is learned from the data, by characterizing the Standard Model background and identifying regions of phase space where the data are inconsistent with it, without assuming what the signal is. Current approaches fall into two broad classes. Overdensity searches look for a localized excess of events over a smooth background, typically in a resonant variable, with the background estimated from sidebands. Out-of-distribution searches look for events populating regions of phase space where the background density is low, so that each such event is individually unlikely under the background hypothesis. The second class does not require a resonance, but it faces two practical problems: constructing a general anomaly score, and reliably estimating the background in the signal region. In this seminar, I will present a method that addresses both problems by combining normalizing flows with the ABCD method. ABCD provides a data-driven background estimate but requires two independent variables, whereas kinematic observables are typically strongly correlated. A normalizing flow transforms the observables into independent Gaussian variables, from which two independent anomaly scores are built and used as the axes of an ABCD plane. Using an R-parity-violating gluino benchmark, I will show that, with the flow trained on a background-only sample, the method improves the signal significance by an order of magnitude and predicts the background in the signal region accurately. Finally, I will discuss training the flow directly on data containing signal, a step towards a fully data-driven search.