Researchers at the LHCb experiment at CERN have developed and implemented a new algorithm called Downstream, designed to reconstruct and select highly displaced vertices in real time at the first level of the High Level Trigger (HLT1). This software-based trigger system operates at a rate of 30 MHz, a constraint that previously limited the experiment’s ability to detect long-lived particles (LLPs) whose decay products do not pass through the innermost tracking detector, the VELO. The new algorithm utilizes the Upstream Tracker (UT) and the Scintillator Fiber detector (SciFi) to identify tracks that originate from decays occurring further from the proton-proton interaction point, thereby widening the physics reach of the experiment during Run 3.
The LHCb forward spectrometer, one of the main detectors at the Large Hadron Collider, has historically focused on searching for new physics through studies of CP-violation and heavy-flavour hadron decays. During previous runs, the first stage of the trigger relied on fast partial particle tracking using Intel Xeon CPUs, which suffered from significant throughput decreases due to large hit occupancy. With the upgrade to a full software trigger executed on NVIDIA GPUs, LHCb has been able to collect 10 fb⁻¹ at 13 TeV during the 2023–2024 period. However, the timing constraints of the HLT1 stage meant that the trigger focused solely on long tracks with hits in the VELO, potentially missing particles with lifetimes that result in decay vertices outside the VELO acceptance.
The Downstream algorithm addresses this limitation by reconstructing tracks that have information from the UT and SciFi but not the VELO. These downstream tracks typically correspond to decay products of Standard Model particles such as K⁰ₛ and Λ hadrons, but they are also crucial for detecting hypothetical long-lived particles predicted by new physics models. The algorithm is based on the extrapolation of SciFi seeds, or tracklets, to the UT detector, including the effect of the magnetic field in the x direction. Only seeds that have not been used for long track reconstruction are considered as downstream candidates, ensuring that the algorithm complements rather than duplicates existing tracking capabilities.
To manage the computational complexity of tracking in the dense environment of the LHCb detector, the algorithm employs a streamlined feed-forward neural network (FFNN) known as the Ghost-Killer. This neural network is designed to suppress ghost tracks, which are fake tracks originated by spurious hits in the detector or reconstruction ambiguities. The FFNN architecture includes eight input variables, comprising information on the downstream track state and the SciFi track properties. A hidden layer with 14 neurons and a ReLu activation function processes these inputs, while a Sigmoid function at the output defines a ghost probability score. The network was trained using 10,000 simulated Bs → φφ events, and the training process ensured that the search windows were not biased by the final track selection strategy.
The implementation of the Downstream algorithm is heavily optimized for GPU execution within the Allen software framework. The algorithm runs on approximately 500 NVIDIA A5000 GPUs, hosted in servers with AMD EPYC processors. To maximize throughput, data is processed in slices of events, with each CUDA thread block mapped to a single event. The use of static structures in the neural network architecture allows for compile-time optimizations, such as utilizing registers instead of global memory. Additionally, techniques such as loop unrolling and the use of fast mathematical functions from the CUDA library are employed to accelerate floating-point operations and reduce data dependencies. These optimizations are critical for maintaining the required 60 kHz throughput per GPU card for the LHCb HLT1 system.
Performance evaluations indicate that the inclusion of the Downstream algorithm in the HLT1 sequence slightly reduces the overall throughput by about 5 kHz when executed on an RTX A5000 card. Despite this minor reduction, the algorithm achieves a track momentum resolution of less than 2% in simulated physics samples. The physics efficiency for reconstructing downstream tracks is high, with more than 75% of downstream tracks successfully reconstructed in simulated samples of Standard Model particles. This efficiency has been verified for K⁰ₛ and Λ particles from various decay channels and minimum bias events. The ghost rejection performance is also robust, with a ghost rate that can be controlled by adjusting the threshold of the Ghost-Killer neural network. A threshold of 0.5 provides a balance between ghost suppression and algorithm efficiency.
An improved version of the algorithm, designated Dv1, has been developed to enhance the reconstruction of new unknown particles with large masses. The original Dv0 version assumed that downstream tracks point to the origin of the coordinate system, which can negatively impact tracking efficiency for decays with very large opening angles. Dv1 introduces an initial tracking step with a triplet search algorithm to increase the efficiency of heavy particle decays. This new version improves the tracking efficiency for decays of heavy Beyond Standard Model (BSM) particles by up to a factor of 4 and removes any mass dependence. It also enhances the efficiency for K⁰ₛ and Λ by about 20% and 10%, respectively. The Dv1 configuration is planned for testing during the 2025 data-taking period.
The algorithm also includes a vertexing component that reconstructs secondary vertices using two downstream tracks. This process involves a Newton–Raphson method to find the point of closest approach between two extrapolated tracks. To account for the magnetic field between the UT and the vertex position, a second-order polynomial is used to determine the track coordinates, with a parameter γ that depends on the charge-to-momentum ratio (q/p) and the magnet polarity. The bias in the z and x positions of the vertex obtained using this technique are 14 cm and 1 cm, respectively.
Validation of the algorithm was performed using real data collected by LHCb during October 2024, when the UT was running in global conditions and the Downstream algorithm was included in the HLT1 sequence. Tracks of interest in physics events were required to have a momentum above 5 GeV/c and a transverse momentum pT above 0.5 GeV/c. The successful reconstruction of Λ and K⁰ₛ particles that do not interact with the VELO detector demonstrates the algorithm’s capability to extend the LHCb physics potential. The study also evaluated the power consumption of the algorithm, finding only a 6% increase in energy consumption due to the highly optimized throughput of the Downstream sequence. This minimal impact on power consumption is significant in the context of sustainability efforts for next-generation experiments.
The expected physics impact of the Downstream and vertexing algorithms is substantial, particularly for searches involving long-lived particles. Two HLT1 trigger lines have been developed to select very detached downstream secondary vertices, relying on four different fast neural networks to select K⁰ₛ and Λ decays and to distinguish detached from prompt decays. For certain charm and baryon decay channels that are only available for downstream tracks, the efficiency increase is 100%. The successful implementation of this algorithm in the real-time trigger system marks a significant step forward in the LHCb experiment’s capability to probe new physics beyond the Standard Model.
The development of the Downstream algorithm represents a major advancement in the software trigger strategy of the LHCb experiment. By leveraging GPU computing and neural network techniques, the algorithm overcomes the limitations of previous CPU-based triggers and expands the detector’s sensitivity to long-lived particles. The successful validation with real data and the planned deployment of the improved Dv1 version in 2025 highlight the ongoing evolution of LHCb’s data acquisition and processing systems. This work not only enhances the search for Standard Model particles but also provides a robust framework for the discovery of new long-lived particles, contributing to the broader goals of particle physics research at the LHC.
Subject of Research: Development and implementation of a GPU-based tracking and vertexing algorithm for long-lived particle selection in the LHCb High Level Trigger.
Article Title: A Downstream and Vertexing Algorithm for Long Lived Particles (LLP) Selection at the First High Level Trigger (HLT1) of LHCb
Article References: Kholoimov, V., Jashal, B. K., Oyanguren, A., Svintozelskyi, V., & Zhuo, J. (2025). A Downstream and Vertexing Algorithm for Long Lived Particles (LLP) Selection at the First High Level Trigger (HLT1) of LHCb. Computing and Software for Big Science, 9(1), Article 10. https://doi.org/10.1007/s41781-025-00141-8
Image Credits: AI Generated
DOI: 10.1007/s41781-025-00141-8
Keywords: LHCb, Long-Lived Particles, High Level Trigger, GPU Computing, Neural Networks, Particle Physics, CERN, Real-Time Reconstruction, Vertexing, Run 3, Downstream, Algorithm
News Source: Katie Riggs. (November 10, 2026). LHCb Algorithm Enhances Real-Time Search for Long-Lived Particles. Scienmag.
