The upgraded LHCb detector, which will be operating during Run 3, detailed by Bediaga et al. (2012), is summarized below. The most important detector improvements as compared to the previous detector (Alves et al., 2008) are the new tracking system and the readout architecture. The readout architecture uses a purely software-based trigger that enables data to be recorded at five times the previous rate, and increases the acceptance of purely hadronic b decays by up to a factor of 2. LHCb is comprised of a tracking system (Vertex Locator, Upstream Tracker, and SciFi tracker), as well as two Ring Imaging Cherenkov detectors, a hadronic calorimeter, and an electromagnetic and muon chambers. These are summarized below.

The VErtex LOcator (VELO) is based on pixelated-silicon sensors and is critical to determine the decay vertices of b and c flavored hadrons. The Upstream Tracker, UT, contains vertically-segmented silicon strips and continues the tracking upstream of the VELO. It is also used to determine the momentum of charged particles with Δp/p ≈ 10% and is useful for removing low-momentum tracks from being extrapolated downstream, thus speeding up the software trigger by about a factor of 3. Tracking after the magnet is handled by the new scintillating fiber-based Scintillating Fiber detector (SciFi). Two Ring Imaging CHerenkov (RICH) detectors supply particle identification. RICH1 is mainly for lower momentum particles and RICH2 is for higher momentum ones. The Electromagnetic Calorimeter (ECAL) identifies electrons and reconstructs photons and neutral pions. The Hadronic Calorimeter (HCAL) measures the energy deposits of hadrons, and the muon chambers M2-M5 are mostly used for muon identification. Figure 1 shows the LHCb upgrade detector.

The main goal of the upgraded LHCb data-acquisition system is to enable the event filter to efficiently select events based on the full data readout from the detectors. The heart of the LHCb online system is the event builder, which is capable of collecting and aggregating data fragments coming from the FPGA front-ends of the various sub-detectors at the full collision rate of 30 MHz (LHCb, 2014b). The data size of a full event during the Run 3 is expected to be as high as 150 kB. Therefore, taking into account the data packed and the transfer protocol overheads, the new event builder has to sustain a throughput of up to 40 Tb/s. The event builder is essentially made up of three components: the readout units, the builder units, and the network interconnecting them. Each readout unit receives the event fragments from the detector point-to-point links and makes them available on the EB network. For each event, one of the builder units gathers all the event fragments from all the readout units, assembles them into a complete event, and writes it into HLT1 input buffer for event filtering. The present event-buider architecture relies on 163 EB servers interconnected through a 200 Gbits/s bi-directional network, as shown in Figure 2. The EB nodes have free PCIe slots that host GPU (Graphic Processing Units) cards to process the HLT1. This has the advantage of reducing the amount of data sent to the Event Filter Farm (EFF) and therefore reduces network costs.

In LHCb, the tracking system for charged particles consists of three subsystems: VELO, UT, and SciFi. The magnet is also necessary to bend particles to estimate their momentum. Several track types are defined depending on the subdetectors involved in the reconstruction, as shown in Figure 3.

The main track types considered for physics analyses are as follows:

When simulating collision data, tracks meeting certain thresholds are defined to be reconstructible and have an assigned type based on the subdetector reconstructibility. This is based on the existence of reconstructed detector digits or clusters in the emulated detector, which are matched to the simulated particles if the detector hits they originated from are properly linked (Li et al., 2021). One can distinguish reconstructibility in the following subdetectors:

Requirements for reconstructible Long tracks implies VELO and SciFi reconstructibility, Downstream tracks must satisfy the UT and SciFi reconstructibility, and T-tracks only requires SciFi reconstrucibility.

The trigger of the LHCb detector readout in Run 3 and later is fully software based and comprises two levels: HLT1 and HLT2, described in LHCb (2020). Most notably, the HLT1 level has to be executed at a 30 MHz rate, which, thus, suffers from heavy constraints on timing for event reconstruction.

The HLT1 is the first trigger stage and performs partial event reconstruction to reduce the data rate. Tracking algorithms play a key role on fast event decisions, and the fact that they are inherently parallelizable processes suggests a way to increase the trigger performance. Thus, the HLT1 has been implemented on a number of GPUs through the Allen project (Aaij et al., 2020), which manages 4 TB/s and reduces the data rate by a factor of 30. After this initial selection, data are passed to a buffer system that allows real-time calibration and alignment of the detector. This is used for the full and improved event reconstruction carried out by HLT2. Figure 4 shows the dataflow at LHCb in the first trigger level.

Due to timing constraints, the current LHCb implementation in this stage is based on partial reconstruction and focuses solely on long tracks, i.e., tracks that have hits in the VELO. This trigger thus significantly affects the identification of particles with long lifetimes, in particular LLP searches in LHCb, where some of the final-state particles are created further than roughly a meter away from the IP and thus outside of the VELO acceptance.

Implementing new fast tracking algorithms in HLT1, allowing direct access to Downstream and T tracks from HLT1 selections, could solve this actual LHCb inability in LLPs reconstruction.

Hybrid seeding, detailed in Aiola et al. (2021), is an iterative reconstruction algorithm that uses information from the SciFi subdetector of LHCb to reconstruct track segments, also called T-tracks. It exploits the four-layered structure of the SciFi stations, as shown in Figure 5, which have two vertical outer (X) layers and two inclined inner (U, V) layers, which are inclined by ±5° to determine the y-position of the trajectory.

This algorithm operates in iterations or cases, starting from different layers to cover for hit inefficiency. Between each iteration, hits that were used to form a good-quality track are removed from the container, which allows the loosening of tolerance windows and track quality requirements as the environment is cleaned up by the algorithm.

Each iteration can be separated into two parts: the x − z track finding and the U/V hit addition. The trajectory of a track, as shown in Figure 5, can be approximated as a parabola with a fixed cubic correction in the x − z plane and a straight line in the y − z plane. Due to the geometry of the SciFi detector, y information can only be accessed using the x(z) equation.

Figure 6 shows the principle of the triplet finding in the x − z plane, which is the most computationally expensive part of the algorithm. Each hit in the first considered X layer is propagated onto a second X layer, assuming the track has infinite momentum and comes from the origin. A tolerance window is opened around that projection, assuming a minimum momentum, and all hits within this window are collected. The difference between the position of the second hit and the infinite momentum projection is a measurement of the momentum, which allows to estimate the bending within SciFi and to look for the third hit. Each triplet then provides the full track equation, and remaining hits are searched for. Tracklets are then selected according to quality criteria such as χ² of the fit to the trajectory and the number of hits.

Given the x(z) track equation, hits on tilted U and V layers can be converted to a y measurement. Considering that we expect tracks to be a straight line coming from the origin, these y measurements are converted to a slope measurement, and the algorithm looks for the accumulation of similar slope measurements in different U/V layers. This is called Hough transformation and clustering. Each Hough cluster which has enough hits is then tested by adding it to the x − z track and performing a global fit to the trajectory. Tracks that do not pass quality requirements on that fit are rejected. The seeds reconstructed by the Hybrid seeding are then used to form more complex tracks. They are matched with VELO segments to form Long tracks or with hits in the UT to form Downstream tracks. The remaining tracklets are then called T-tracks.

Recently, an optimization and high-throughput configuration of this algorithm has been implemented in the new GPU architecture of the LHCb trigger system. The main features of this new seeding is the high level of parallelism and the reimplementation of the U/V hit addition as a sequential hit search, similar to that performed in the x − z part. Additionally, some quality criteria have been tightened. Figure 7 shows the preliminary physics performance of the Hybrid seeding algorithm as implemented on GPUs on Long tracks coming from the B_s → ϕϕ decay channel, where the ϕ meson is reconstructed from two kaons of opposite signs that traverse the SciFi detector.

Some models of physics beyond the Standard Model (BSM) predict the Standard Model (SM) Higgs field serving as the portal to a dark sector, which could accommodate dark matter candidates (Kachanovich, Aliaksei et al., 2020). One of the simplest and well-known models in this regard predicts the existence of a mixed state between a new scalar low-mass boson (H') and the SM Higgs (H), regulated by the mixing strength θ:

In this model, H' can be interpreted as a mediator to a dark sector of unknown mass and lifetime. This model could be validated through the experimental signature of the decay B→H′K, with the H' decaying into π⁺π⁻, K⁺K⁻, μ⁺μ⁻, orτ⁺τ⁻, depending on its mass. A displaced vertex could be determined, allowing to reconstruct the H' mass from the kinematics and the identification of the two decay particles. The sensitivity to this model depends nevertheless on the H' mass and lifetime, which could lead the new scalar to decay outside the detector fiducial volume. In particular, if the H' has long lifetime, the two final decay particles would not be selected by the first level of the LHCb trigger, thus escaping detection. Figure 8 shows the parameter regions and the sensitivity of different (present and future) experiments to the H' (Kachanovich, Aliaksei et al., 2020). One should note that the mixing angle is directly related to the decay width, and thus to the lifetime, with small mixing angles implying long lifetimes.

Considering the leptonic decay mode H' → μ⁺μ⁻, the decay rate can be expressed by:

where G_F is the Fermi constant and m_l the lepton mass. The Higgs lifetime can be computed:

The following sections discuss the sensitivity of the LHCb experiment at the trigger level to the H' decay into the μ⁺μ⁻ final state. This decay channel binds the Higgs mass to be m_H > 2m_μ ≈ 212 MeV/c². The H' mass is also constrained to mH′<mB+-mK+≈4700 MeV/c².

Using the upgraded LHCb simulation, 99 MC samples of 7,000 events each are simulated from proton-proton collisions. The samples are generated using Pythia8 and assuming Run 3 beam conditions. The B→H′(→ μ⁺μ⁻)K decay channel is generated considering H' masses in the range of 500–4,500 MeV and lifetimes from 1 to 2,000 ps.

The decay vertex of the H' is expected to be displaced with a dependence on these variables, and they will be labeled in the following according to the track type of the two muons (two Long tracks = LL, two Downstream tracks = DD and two T-tracks = TT). LL vertices are thus expected to be produced in the Velo detector, DD between the Velo and the UT, and TT vertices between the UT and SciFi. The reconstructibility of these vertices is defined according to the track reconstructibility criteria defined in Section 2.3, and imposing that the two muons are coming from the same vertex. Figure 9 shows the reconstructibility of the decay vertex of the H' particle as a function of its mass and lifetime. For lifetimes below 10 ps, a large proportion of LL vertex topologies is found, as expected, where the H' decays in the VELO acceptance and both muons can be reconstructed as Long tracks. Nevertheless, for H' lifetimes larger than 100ps (and small-mixing angle), most of the decays are produced downstream the VELO, resulting in a large proportion of the DD and TT topologies. These fractions are very similar in the case that the H' decays into two hadrons¹.

Figure 10 shows the LHCb HLT1 effect when triggering on the H' decay products (Trigger on Signal, TOS). Since only Long tracks are reconstructed at the HLT1 level, a high inefficiency can be observed for large H' lifetimes, going down 10% for lifetimes larger than 500ps. A loss in sensitivity for small H' masses is also observed, and since the H' is experiencing larger boosts, muons, are escaping from detection in the VELO.

Some strange SM particles, such as KS0 and Λ⁰ baryons, also have long lifetimes of the order of 100ps. These particles are reconstructed in many physics analysis to measure observables which can be sensitive to new physics models. One example is the rare Λb→Λ0γ decay, where the Λ⁰ baryon decays into a proton and a pion. Measurements of the branching fraction and the angular distribution of the decay particles are sensitive to models with non-standard right-handed currents (Aaij et al., 2019, 2022; Garćıa Mart́ın et al., 2019). Another example is the very rare KS0 → μ⁺μ⁻ decay, very suppressed in SM, not observed experimentally at present, which is sensitive to different BSM scenarios such as SUSY (Zhu, 2012) or presence of leptoquarks (Bobeth and Buras, 2018). Figure 11 shows the sensitivity to the KS0 → μ⁺μ⁻ decay channel (in terms of branching fraction limit) as a function of the product of the trigger efficiency and luminosity.

Using 10,000 simulated events, the HLT1 trigger effect has been studied on these two decay channels involving Λ⁰ and KS0 particles. Figure 12 shows the normalized number of reconstructible events (LL+DD+TT) as a function of the end decay vertex of the Λ⁰ and of the KS0.

The relative proportions of track types are 12% (LL), 51% (DD), and 37% (TT) for the Λb→Λ0γ decay channel. For the case of prompt KS0, the amount of reconstructible tracks is 46% (LL), 38% (DD), and 16% (TT), One should notice that a large amount of decays occurs at the end of the VELO and they do not let enough hits to be reconstructed as Long tracks, being reconstructed thus in the DD category. That is, the amount of DD+TT tracks is about 88% for the Λb→Λ0γ decay channel and 54% for prompt KS0.

One can apply the HLT1 conditions on the reconstructible events, and in particular, to check how many events are selected by inclusive trigger lines, such as the OneTrackMVA or TwoTrackMVA. They require tracks with minimum transverse momentum and minimum Impact Parameter (IP) significance with respect to the Primary Vertex, and for the latter, they form a vertex with minimum requirements. In the case of Λb→Λ0γ, the HLT1 signal efficiency for the proton and pion coming from the Λ⁰ is found to be less than 10%. In the case of KS0, the HLT1 efficiency on the muons, adding some inclusive muon lines and dedicated lines for KS0 selection, is less than 25%. We normalize the efficiency to the sum of LL+DD+TT reconstructible events. The KS0 candidates in this study are prompt, produced at the interaction point. In the case that KS0 candidates are coming from the decays of b or c-hadrons, the amount of reconstructible LL candidates is expected to decrease, thus increasing the HLT1 inefficiency.