White, MartinJackson, PaulMullin, Anna Jane2021-11-262021-11-262021https://hdl.handle.net/2440/133440Beyond-Standard Model (BSM) physics searches at the LHC are limited by the amount of information available to distinguish a new physics process from its backgrounds. Analyses apply a range of classification algorithms to obtain sensitivity to rare signals, but are challenged to obtain enough information in a broad parameter space without relying on heavy optimisation in narrow search regions. LHC event classification techniques become more powerful when they can be applied broadly to diverse models, requiring a large number of independent variables sensitive to anomalous signals. In our prototype ATLAS search, we create new variables that target information not used in current methods. Whereas typical variables treat events in isolation, we obtain further discrimination from the “similarity” between event pairs by evaluating “distances” in a kinematic space. A map of event similarities forms a graph network, which provides a convenient range of network variables able to quantify local topologies. In networks constructed from nodes of LHC events, we aim to use network variables to increase sensitivity to anomalous topologies local to BSM events. Our proof-of-principle analysis reveals that BSM physics events may populate distinct distributions compared with Standard Model events in several types of network variables, including measures of local centrality and clustering, using supersymmetry searches as examples. Graph network analysis may contribute power to existing methods of event classification and increase sensitivity to anomalous signatures.engraphgraph networkATLASParticle physicsLHCHEPThesis