Extension and Optimization of GPU-Accelerated Cellular Automaton Track Reconstruction for the CMS Experiment at the HL-LHC

The High-Luminosity LHC will significantly increase the collision rate and pileup delivered to the CMS experiment, with up to about 200 simultaneous proton–proton interactions per bunch crossing. These conditions make charged particle track reconstruction at the High-Level Trigger increasingly challenging, both in terms of physics performance and computing resources. To satisfy the Phase-2 trigger requirements, CMS is developing heterogeneous reconstruction workflows in which computationally demanding algorithms are offloaded to GPUs. This thesis presents extensions and optimizations of the GPU-accelerated Patatrack pixel-tracking reconstruction for the CMS Phase-2 High-Level Trigger. First, the Patatrack cellular-automaton algorithm is extended to include macro-pixel hits from the innermost rings of the Tracker Endcap Double Disks. The additional measurements improve the transverse-momentum resolution in the forward transition region, with gains of up to about 25%, while leaving the tracking efficiency essentially unchanged and introducing only a modest increase in fake rate. Production-like benchmarks show that this extension is compatible with the GPU-offloaded reconstruction workflow. Second, a deep neural network (DNN)-based High-Purity selector is developed to replace the current cut-based pixel-track selection. The classifier uses track-level quantities available after pixel-track reconstruction and does not rely on reconstructed vertex information. It recovers most of the efficiency loss introduced by the cut-based selection while reducing the fake rate in the nominal tt validation sample, and it is further tested on high-pT QCD and B0s → μ+μ− samples. Finally, the DNN selector is integrated into the CMS Software heterogeneous reconstruction framework using the PyTorch-Alpaka interface. After dedicated developments to stabilize GPU memory usage, including fixed queue assignment, mini-batched inference, and half-precision execution, the optimized selector runs in about 3 ms on GPU, with a memory overhead of about 10 MiB per stream. The results demonstrate that improved pixel-track selection can be integrated into the GPU-resident Patatrack workflow with an acceptable timing and memory cost.