Robust Anomaly Detection in Independent Searches for New Physics with  the ATLAS/LHC Experiment

This work explores a model-independent approach based on Anomaly Detection techniques to search for New Physics phenomena in the ATLAS experiment at the LHC.
Focusing on a fully-hadronic topology characterized by large missing transverse energy and at least one large-R jet with high momentum, we explore semi-supervised Machine Learning methods trained on a broad region of Standard Model simulated events, making them sensitive to New Physics signals manifesting as anomalies.

First, we assess the sensitivity of both deep (AutoEncoder and Deep SVDD) and shallow (Isolation Forest) implementations to a wide topology of benchmark signals, with varying signatures and acceptances within the search space.
The performance of the proposed models is benchmarked with relevant kinematic variables.

Second, to improve robustness against instrumentation limitations such as systematic uncertainties arising from experimental calibrations and reconstruction of missing transverse energy and large-R jets, we implement an adversarially-trained AutoEncoder.

Our findings evidence the strong potential of these techniques in the independent searches beyond the Standard Model, having achieved high sensitivity to diverse signatures and strong resilience to data uncertainties.