عبدالله سعيد

Artificial Intelligence Specialization · Cohort 2026

Linking Tracks and Tracksters in the TICL Reconstruction Pipeline

The Phase-2 upgrade of the CMS detector at the CERN Large Hadron Collider will require a new generation of reconstruction algorithms capable of handling the much higher event complexity and particle densities expected in future running conditions. In the high-granularity endcap calorimeter, the reconstruction of particle showers and the association between tracks and calorimeter tracksters are especially challenging, since the same physical particle may be represented by several partially overlapping calorimeter clusters, while nearby particles may share detector activity in a highly non-trivial way. To address this problem, we are developing a many-to-many linking strategy between tracks and tracksters. The goal is to move beyond a simple one-to-one assignment and instead allow a more flexible and physically meaningful reconstruction of complex showers. This approach is based on a minimum-cost flow (MCF) formulation, where the linking problem is cast as an optimization over a graph of possible connections. In this framework, the algorithm can naturally represent charged and neutral components, branching structures, and ambiguous associations in a globally optimized way. In parallel, a graph neural network (GNN) has already been developed to operate on the same linking graph and to assign scores to possible connections. These learned scores encode the local compatibility between tracks and tracksters, and they can capture subtle patterns that are difficult to express with hand-crafted rules alone. The central idea of the project is therefore to combine the strengths of the two approaches: use the GNN output as input to the MCF formulation, so that the learned scores define the edge costs of the optimization problem. In this way, the GNN provides a flexible, data-driven estimate of link quality, while the MCF enforces a globally consistent solution that respects exclusivity, flow conservation, and the overall structure of the event. In this way, the final linking decision is not made by the neural network alone, but by an optimization procedure informed by machine learning. This combination is expected to improve reconstruction performance in dense environments, especially for close-by electrons and photons, and to provide a more robust and interpretable linking algorithm.

Read the final report (PDF)