Track Finding in the LHCb Vertex Locator using Graph Neural Networks

The start of Run 3 of the LHCb experiment in 2022 introduced Allen, a fully GPU-based first-level trigger system. This framework performs a near-complete reconstruction of proton-proton collisions at a rate of 30 MHz. Looking toward Run~5 and the High-Luminosity LHC (HL-LHC), increased luminosity and detector occupancy necessitate exploring alternative reconstruction methods, such as Artificial Intelligence, within this existing GPU infrastructure. This thesis presents the adaptation of the Exa.TrkX Graph Neural Network (GNN) pipeline, originally designed for central detectors like ATLAS and CMS, to the LHCb Vertex Locator (Velo). A key advantage of this study is the ability to benchmark the GNN pipeline against the standard "Search by triplet" algorithm on identical hardware. The adaptation and implementation rely on four main contributions: First, the preparation of large training and testing datasets by converting grid simulation files using the developed digout library. Second, an evaluation framework, MonteTracko, designed to compare the physics performance of the ETX4VELO pipeline against the baseline.Third, architectural adaptations to address the specific challenges of a forward detector, such as well-defined sensor layers and crossing tracks. Fourth, the deployment of an inference pipeline in C++/CUDA within Allen. While training is performed in Python (PyTorch), production inference leverages TensorRT and custom CUDA kernels to handle combinatorial operations. The implemented pipeline achieves physics performance comparable to the Search by triplet algorithm. Although the current throughput is three orders of magnitude lower than the baseline, this work establishes a functional GNN tracking implementation on GPUs, identifying specific optimisations, such as quantisation, to address the speed gap.