Track Finding in the LHCb Vertex Locator using Graph Neural Networks

Overview & Synopsis

Since 2022, during Run 3, LHCb has relied on Allen, a fully GPU-based trigger capable of reconstructing about 30 million collisions per second. For Run 5 of the HL-LHC, planned for the next decade, the increase in luminosity and granularity will make events heavier and will tighten throughput requirements even further. In this context, my thesis adapts the Exa.TrkX pipeline, based on graph neural networks (GNNs), to LHCb's first tracking detector, the VELO, and then compares it with Allen's reference GPU algorithm, “Search by Triplets”. The comparison covers both physics performance (accuracy) and speed performance (throughput), on exactly the same hardware, which is a major advantage. The resulting pipeline is dubbed ETX4VELO.Specifically, I:(i) assembled and adapted training datasets with the Python libraries XDIGI2CSV and then digout;(ii) developed a robust evaluation framework with the Python library MonteTracko;(iii) tuned and evolved Exa.TrkX to fit the VELO, a detector with well-defined vertical planes where tracks sometimes intersect;(iv) ported the C++/CUDA inference into Allen via inference engines such as TensorRT and ONNX Runtime, with combinatorial steps written in CUDA. The physics performance matches that of “Search by Triplets”, but the throughput remains roughly a thousand times lower. Nevertheless, this result provides a solid and innovative starting point, with several opportunities for optimization, notably model quantization.