![HEP2026: 42nd Conference on Recent Developments in High Energy Physics and Cosmology [NTUA] Jul 8 – 11, 2026](/event/1654737/logo-1929213524.png){kind=logo}
Speaker
Description
The unprecedented computational challenges of the High-Luminosity Large Hadron Collider, driven by the large number of simultaneous proton-proton collisions per bunch crossing, demand a fundamental change in reconstruction strategy. To address this, innovative algorithms are proposed for track seeding, building and fitting in Phase-2 CMS. For seeding, the Patatrack and Line Segment Tracking (LST) algorithms are optimized for parallel execution on GPUs. For building, the mkFit algorithm is vectorized for efficient CPU performance, and its score is further extended to track fitting, with initial measurements demonstrating a ~4× speedup for this task. Together, this new strategy can achieve up to a factor of 6× speedup of the full Phase-2 High Level Trigger (HLT) tracking sequence. Beyond computational efficiency, the combination also broadens the physics reach of CMS. LST, designed to build track candidates using outer tracker hits alone, naturally accommodates displaced track signatures and extends the acceptance for displaced tracks by ~60 cm. This talk presents the new baseline strategy for Phase-2 tracking in the CMS experiment, targeting the online HLT reconstruction and serving as the basis for the main iteration of offline tracking. Novel ML architectures based on Transformers are also discussed. These operate on LST segments rather than individual hits and use object condensation to create full tracks in a single inference step. They can achieve high reconstruction efficiency with low fake and duplicate rates, demonstrating promising preliminary results on a realistic CMS Phase-2 detector simulation.