The case for the upgrade rests on what the High-Luminosity LHC itself will deliver. When the accelerator completes its current upgrade programme, it will increase the integrated luminosity delivered to experiments by roughly an order of magnitude compared with Run 3, flooding the forward region with an unprecedented flux of neutrinos. These particles, born in the decays of heavy-flavour hadrons such as those containing charm quarks, are produced at extreme angles relative to the beamline, at pseudorapidities between about 7 and 8, and stream hundreds of metres along the tunnel. No general-purpose detector can catch them. SND@LHC, sitting off the collision axis, was conceived precisely to exploit this blind spot, and its first observations of both muon neutrinos and electron neutrinos from LHC collisions confirmed that collider neutrino physics has moved from theory to reality.

The upgraded detector, referred to in the paper as SND@HL-LHC while the collaboration keeps its original name, must contend with a fundamental technological problem: the emulsion-based tracking system that served the experiment during Run 3 cannot cope with the particle rates of the high-luminosity era. Emulsion films integrate everything that passes through them and offer no event-by-event electronic readout. The upgrade therefore replaces them with a fully electronic, silicon-based tracking system capable of real-time operation while preserving the spatial resolution needed to reconstruct neutrino interactions in a high-occupancy environment. The design also adds two capabilities the original detector never had: a precision timing system for rejecting background, and a magnetic spectrometer that measures the momentum and charge of outgoing muons, allowing, for the first time, a clean distinction between neutrino and antineutrino interactions.

The detector itself remains remarkably compact, roughly three metres long with a cross-section of about one square metre, a footprint dictated by the tight confines of the TI18 tunnel. Its heart is the neutrino target region, a stack of 58 tungsten plates, each seven millimetres thick, interleaved with silicon microstrip layers salvaged and repurposed from the CMS Microstrip Tracker. The stack weighs about 1.3 tons and corresponds to 116 radiation lengths and 4 nuclear interaction lengths, dense enough to stop and measure the showers produced when neutrinos strike tungsten nuclei. The silicon modules resolve positions to better than 30 micrometres, and consecutive layers are rotated by 90 degrees to provide precision tracking in both spatial coordinates. To keep the irradiated silicon stable, the target operates at approximately minus 20 degrees Celsius inside an insulated enclosure cooled by a closed-loop dry-air system with full redundancy and safety monitoring.

Downstream of the target sits the second major component: a magnetised hadronic calorimeter built from 34 layers of 50-millimetre-thick iron slabs interleaved with more recycled CMS silicon modules. The iron serves double duty, absorbing hadronic showers so their energy can be measured and providing the medium for a magnetic field of approximately 1.7 tesla, generated by a single-piece copper coil. This field bends the trajectories of muons leaving the target, allowing momentum measurements up to about 1 teraelectronvolt with a resolution of roughly 20 percent. Finite-element simulations confirm a homogeneous field within the active volume, with non-uniformities of only a few percent and negligible stray fields outside. Upstream, three planes of plastic scintillator bars read out by silicon photomultipliers form a veto system with a rejection efficiency at the level of one part in a billion against the few-kilohertz flux of beam muons, while a dedicated timing system, based on either fast scintillator tiles or thin-gap resistive plate chambers, aims at sub-50-picosecond resolution to tie every event to the correct LHC bunch crossing.

The most consequential part of the paper is its comparison of two installation scenarios, both using the identical detector. The Baseline configuration, which is the approved version of the upgrade, places the detector directly on the existing tunnel floor with no civil-engineering work, leaving its centre about 57 centimetres above and 39 centimetres to one side of the collision axis. That position yields an average pseudorapidity coverage of 6.9 to 7.7 but an azimuthal acceptance of only about 66 milliradians, squeezed by the clearance between detector and ceiling. The Extended configuration instead lowers the detector by 43 centimetres and shifts it 29 centimetres horizontally, requiring the removal of 4.5 cubic metres of concrete. The reward is dramatic: the transverse distance to the collision axis shrinks to roughly 14 by 10 centimetres, pseudorapidity coverage extends beyond 7.8, and the azimuthal acceptance grows roughly fourfold to about 250 milliradians, which the authors identify as the dominant contribution to the overall acceptance gain.

Simulations based on the DPMJET, FLUKA, GENIE and POWHEG generators, assuming the full 3000 inverse femtobarns of integrated luminosity expected over roughly a decade of HL-LHC operation, quantify what that geometric gain buys. In the Extended configuration the total neutrino interaction rate rises by approximately a factor of five, and the energy spectra shift toward higher values because the detector intercepts a more energetic component of the forward flux. The consequences ripple through the entire physics programme. For muon neutrino and antineutrino deep-inelastic scattering cross-sections above 350 gigaelectronvolts, a region never previously explored that bridges fixed-target experiments and IceCube, the statistical uncertainty falls from about 3 percent in the Baseline layout to below 1 percent in the Extended one, while the dominant flux systematic can be constrained to about 5 percent thanks to overlapping measurements by the LHCf experiment during Run 3.

The upgrade also promises precision tests of fundamental symmetries. Because electron neutrinos in the detector’s acceptance originate predominantly from charm decays, their measurement probes forward charm production and the gluon distribution of the proton at Bjorken-x values below ten to the minus five, a regime relevant both for future colliders such as the FCC and for predictions of ultra-high-energy atmospheric neutrinos. Per-bin statistical uncertainties on these charm-related measurements improve from about 10 percent in the Baseline to 5 percent in the Extended configuration. Tests of lepton flavour universality sharpen considerably: the ratio of electron-neutrino to tau-neutrino interactions, sensitive to the cross-sections of the two species, would be measured with a statistical uncertainty of about 6 percent in the Baseline and 2 percent in the Extended layout, compared with 30 percent in Run 3. Perhaps most tantalising, the Extended configuration could deliver the first direct observation of tau antineutrinos, which have never been seen experimentally: simulations predict roughly 120 tau antineutrino charged-current interactions in the target, about 9 observable signal events against a background of 2, yielding an observation significance of about 5 sigma, where the Baseline manages only around 3 sigma.

Beyond the Standard Model, the detector doubles as a hunting ground for feebly interacting particles such as dark photons, heavy neutral leptons and axion-like particles produced in the far-forward region, and the paper evaluates sensitivity to light dark matter scattering off protons via a leptophobic baryonic mediator. Here too the Extended geometry wins, since the larger acceptance and closer approach to the collision axis increase the dark matter flux intersecting the detector and extend the reach in mediator coupling into territory not excluded by other experiments. The collaboration also foresees roughly 600 to 1300 neutrino interactions coincident with events recorded by ATLAS, enabled by sub-50-picosecond timing synchronisation, which would directly link a neutrino interaction to its production mechanism. Whatever installation scenario is ultimately realised, the experience gained will feed directly into the SND detector at the proposed SHiP facility, for which the upgraded instrument serves as a full-scale prototype, making this modest tunnel experiment a testing ground for the next generation of collider neutrino physics.

Subject of Research: Upgrade of the SND@LHC forward neutrino detector for the High-Luminosity LHC, comparing installation scenarios and physics reach

Article Title: SND@LHC upgrade for the high-luminosity LHC: physics reach and installation scenarios

Article References: SND@LHC Collaboration, Abbaneo, D., Ahmad, S., Albanese, R., Alexandrov, A., Alicante, F., Aloschi, F., Androsov, K., Anokhina, A., Arellano, L. G., Asawatangtrakuldee, C., Torres, M. A. A., Bangaru, N., Battilana, C., Bay, A., Bertocco, A., Betancourt, C., Bick, D., Biswas, R., … Saá, J. Z. (2026). SND@LHC upgrade for the high-luminosity LHC: physics reach and installation scenarios. The European Physical Journal C, 86(9), Article 1030. https://doi.org/10.1140/epjc/s10052-026-15924-5

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-15924-5

Keywords: SND@LHC, High-Luminosity LHC, neutrinos, CERN, particle physics, forward physics, heavy-flavour decays, dark matter, lepton flavour universality, tau antineutrino, silicon trackers, magnetic spectrometer

The motivation is stark. The Large Hadron Collider’s physics programme depends on Monte Carlo simulation, in which the fates of billions of proton collisions are computed particle by particle, interaction by interaction. The workflow has four stages: generating the physics events, propagating particles through a virtual detector, digitising the deposited energy into signals that mimic real detector readout, and reconstructing tracks with the same algorithms used on actual data. The detector response step, traditionally handled by the Geant4 toolkit, is the most computationally punishing, and the problem is about to get dramatically worse. By the end of the High-Luminosity LHC programme, the ATLAS and CMS experiments will have collected up to ten times more data than in their first three runs combined, and many precision analyses will find their sensitivity limited not by the data themselves but by the statistical size of the simulated samples needed to model backgrounds.

Generative machine learning has already made inroads on the calorimeter problem, where the goal is to reproduce the showers of energy deposited when particles slam into dense material. Generative adversarial networks, variational autoencoders, normalising flows, diffusion models and continuous flow models have all been applied, and the ATLAS experiment already uses GANs in production for fast photon shower simulation. But tracking detectors present a fundamentally different challenge. A calorimeter response can be interpreted as an image with correlated entries. A tracking detector, by contrast, records a sparse sequence of discrete interactions: a particle crossing thin layers of silicon widely separated in space. Rendering that as an image would produce a picture that is almost entirely empty, which is precisely why the Slovenian team turned to sequence-based models instead.

The conceptual leap is elegant. In a large language model, text is broken into tokens and the network learns to predict the next token given everything that came before. Novak and Kerševan applied the same logic to particle tracks. Each hit in a silicon sensor is described by seven features: a discrete particle identifier encoding type and charge, a geometry identifier for the detector module, two local coordinates on the sensor surface, and three components of the particle momentum after the hit. Every feature is tokenised into a shared dictionary of integers, with continuous quantities rounded to two decimal places. A track becomes a flat sequence of tokens, bookended by virtual start and end hits representing the particle’s birth at the beamspot and its exit from the tracker. The transformer’s job is then exactly what GPT does with words: predict the next element of the sequence, one feature at a time.

The technical implementation relies on a decoder-only transformer built on Karpathy’s nanoGPT codebase, with eight layers and eight attention heads each, feed-forward dimensions four times the input dimension, and dropout of 0.2. Model sizes range from 9.1 to 35 million parameters depending on the token dictionary. Because self-attention scales quadratically with sequence length, the authors employed a sliding-window training scheme, chopping full tracks into rolling windows of four hits. This is physically justified since correlations between hits weaken with distance and the dominant feature, local track curvature, is captured by nearby hits. The trick also augments the training data by an order of magnitude, letting the graphics accelerators work with much larger batches. Training used the AdamW optimiser with gradient clipping and a carefully scheduled learning rate, running for roughly 5000 epochs on Slovenian supercomputing hardware.

Validation proceeded at two levels. At the hit level, the researchers compared distributions of individual hit properties between Geant4 and the neural network, always against the rounded Geant4 data on which the model was trained. For single muons, which barely interact with the detector beyond multiple scattering, the agreement is impressive. The number of hits per track is well modelled, with the largest discrepancy for any muon model being just six hits, and directly tokenised quantities such as the momentum component p_x agree at the percent level. Derived global coordinates fluctuate up to about five percent, with deviations growing in the distribution tails. At the higher level, the simulated hits were fed into the ACTS tracking software and reconstructed using the standard Open Data Detector configuration, allowing a direct comparison of seeding and tracking efficiencies.

Those reconstruction results are the heart of the paper. The technical seeding efficiency measures how often the algorithm finds viable track seeds, while the technical tracking efficiency measures how often full tracks are successfully fitted. A benchmark 11.2-million-parameter model reaches 94.9 percent tracking efficiency with only a slight seeding efficiency drop, comparable to the rounded Geant4 reference. Scaling the model’s layer dimensions by a factor of two improves both metrics further, though they remain below the reference sample. One persistent weakness emerged: the azimuthal angle phi is poorly modelled when the full detector coverage is included, and a dedicated test with a wider phase space showed that the model learns probabilities less accurately as the training domain grows. The authors conclude that detector-aware rounding and tokenisation strategies will be essential for production-quality models.

Electrons and pions exposed the limits of the approach. Electrons undergo bremsstrahlung and pair production, processes that can abruptly change their direction and momentum, and the model’s token dictionary grew 1.4 times larger to accommodate the added variety. Although overall track quality remained comparable to muons, the momentum modelling proved insufficient, with a sizeable fraction of generated electrons carrying too much transverse momentum; the network learns the shape of the momentum tail but is biased toward smaller changes. Pions, which can decay inside the tracker, revealed a related failure: rare, low-probability processes are systematically under-learned, so machine-generated pions almost never decay, artificially inflating the tracking efficiency even though the tracks themselves remain good quality. Reweighting such rare events could help, but the authors note it would erode the statistical power that makes fast simulation worthwhile in the first place.

Computing performance adds a further caveat. GPU inference runs at roughly the same speed as Geant4 simulation on same-generation CPUs, scaling linearly with model size, so better physics currently means slower generation. CPU inference is orders of magnitude too slow to replace conventional simulation, and given the much higher cost of high-performance GPUs, the authors judge the present speed insufficient for general production use. Still, benchmarks across Nvidia hardware generations show clear gains, and half-precision bf16 computation speeds up training by about a third with no loss of physics performance. The team has released its SiliconAI training and validation code and datasets openly, and sees paths forward in extending the sequence representation to secondary particles and in exploring alternative sequence-based architectures. The message is provocative: the same technology that writes essays can, with careful tokenisation, learn to write the passage of a muon through a silicon tracker, and with further work it may help the LHC’s next generation of experiments simulate their way to new discoveries.

Subject of Research: Generative transformer models for fast simulation of silicon tracking detectors in high energy physics

Article Title: GPT-like transformer model for silicon tracking detector simulation

Article References: Novak, T., & Kerševan, B. P. (2026). GPT-like transformer model for silicon tracking detector simulation. The European Physical Journal C, 86(9), Article 1090. https://doi.org/10.1140/epjc/s10052-026-16362-z

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-16362-z

Keywords: transformer, GPT, silicon tracker, detector simulation, Geant4, machine learning, LHC, High-Luminosity LHC, particle physics, generative models, Open Data Detector, computing