Speaker
Description
Anomaly detection has emerged as a promising paradigm for BSM searches at the LHC, leveraging modern machine learning techniques to identify subtle deviations directly in data. In particular, out-of-distribution (OOD) approaches target signals that populate the tails of an anomaly score, offering sensitivity to a wide range of unforeseen signatures. A central challenge in such searches is the estimation of the background in the extreme tail of the distribution: by construction, events in the signal region are poorly modelled by simulation. While extrapolation from control regions is in principle possible, it is typically unreliable, and sideband interpolation is not applicable.
In this talk, I will present a novel strategy that combines OOD anomaly detection with the ABCD method for data-driven background estimation. The approach integrates normalizing flows and variational autoencoders to construct an end-to-end framework that enables robust background predictions directly from data, even in the far tail of the anomaly score distribution. I will present preliminary results demonstrating the performance of the method on a benchmark BSM scenario and discuss its implications for future anomaly-based searches at the LHC.