This thesis presents a novel methodology within the ATLAS experiment that exploits the large number of recorded untriggered pp interactions, commonly referred to as pile-up, to construct a dedicated low-p_T dataset. By treating each additional inelastic collision in a recorded bunch crossing as an independent low-energy interaction, the work establishes a trigger-unbiased framework to compute the jet energy resolution and apply heavy-flavour identification in the low-pT regime, where traditional trigger-based datasets are statistically limited. A full reconstruction strategy is developed, including the removal of the triggered interaction, multi-vertex jet reconstruction, vertex dependent jet selections, and a detailed validation of the resulting dataset using ATLAS Run 2 data. The extracted jet energy resolution is shown to be consistent with established ATLAS measurements while providing significantly improved statistical precision at low transverse momentum.

Building on this dataset, this thesis extends ATLAS flavour tagging to operate in a multi-vertex environment for the first time. The GN2 algorithm, a transformer-based, end-to-end neural network, is adapted to incorporate vertex-dependent information arising uniquely from pile-up reconstruction. New track-vertex and jet-track association techniques are introduced, allowing GN2 to process inputs from multiple primary vertices simultaneously. To validate these developments, new b-filtered JZ samples are produced, providing a controlled environment in which to assess displaced-track behaviour and background rejection. Performance studies in both tt̄ and dijet samples demonstrate a first step towards flavour tagging applied to pile-up, revealing the sensitivity of GN2 to vertex misassociation and truth-labelling ambiguities.

The results establish a foundation for precision heavy-flavour measurements using the pile-up reconstruction approach and demonstrate that modern flavour tagging algorithms, traditionally designed for single-vertex topologies, can be successfully generalised to encompass multi-vertex classification. The methods developed here open the door to future low-energy QCD and flavour physics studies in a region of phase space previously statistically limited to ATLAS.