Review of Machine Learning for Real-Time Analysis at the Large Hadron Collider experiments ALICE, ATLAS, CMS and LHCb

• 30 MHz collisions produce 1 petabyte per second in LHC Run 3, requiring real-time processing to decide which fraction of data to keep.Source: §1 · p. 3
• LHC real-time systems must read out sub-detectors, reconstruct particle candidates, and store selected detector information under strict timing constraints.Source: §1 · p. 3
• Trigger systems determine which events are retained or discarded, and their latency and bandwidth constraints can limit the experiments’ physics prospects.Source: §1 · p. 3
• Modern ML models can perform physics-object reconstruction and detector alignment or calibration on complex, high-dimensional data within real-time latency requirements.Source: §1 · p. 3
• Academia–industry collaboration has improved real-time ML viability, while dedicated hardware is sometimes required to achieve the necessary inference latency.Source: §1 · p. 3

2.1 The ALICE Detector

• ALICE studies heavy-ion and proton collisions using tracking and particle-identification subsystems including the ITS, TPC, TRD, and TOF.Ahead of Run 3, ALICE adopted streaming readout; its CTS synchronises sub-detectors before the HLT reconstructs, calibrates, aligns, and reduces data.Source: §2.1 · p. 4

2.2 The ATLAS Detector

• ATLAS combines inner tracking, electromagnetic and hadronic calorimetry, and muon spectroscopy across nearly the full solid angle.Source: §2.2 · p. 4

2.3 The CMS Detector

• CMS uses a 3.8 T superconducting solenoid containing silicon tracking, ECAL, and HCAL systems, with forward calorimeters and external muon detectors extending coverage.Source: §2.3 · pp. 4–5

2.4 The LHCb Detector

• LHCb is a forward spectrometer for heavy-flavour and broader physics studies, with tracking, RICH detectors, calorimetry, and muon stations spanning 2 < η < 5.During Run 3, its luminosity is levelled at 2 × 10^33 cm−2 s−1, corresponding to approximately six proton–proton interactions per bunch crossing.Source: §2.4 · p. 5
• LHCb removed its hardware trigger before Run 3 and uses HLT1 and HLT2 software stages on heterogeneous GPU-based computing.Source: §2.4 · p. 5

3.1 The Real-Time Environment

• 99.75% of collisions are rejected by the ATLAS and CMS hardware trigger, while accepted events proceed to software triggering with fuller detector information and more complex algorithms.Hardware triggers read only selected regions of interest or sub-detectors, reducing input fidelity; software triggers can reconstruct physics objects using the full detector.Source: §3.1 · p. 6
• Online reconstruction performance now frequently approaches offline performance as CPU capability, algorithmic efficiency and ML models improve.Online performance still often falls short because real-time readout has lower detector fidelity and must handle immense data volumes.Source: §3.1 · p. 6

3.2 Integration of ML into Existing Workflows

• ML applications in real-time processing range from simple hardware-deployed models to neural-network reconstruction and fast approximations of more complex algorithms.Model placement depends on computational constraints and the detector readout available within the trigger sequence.Source: §3.1 · p. 6, §3.2 · p. 7

3.2.1 Machine Learning on Dedicated Accelerators

• FPGAs support low-latency, high-throughput ML by parallelising operations such as matrix multiplication, enabling more complex models in HEP triggers.Open-source and vendor tools translate models into FPGA implementations, while abstraction layers shorten development cycles.Source: §3.2.1 · p. 7
• Quantisation-aware training, knowledge distillation and pruning adapt ML models to resource-constrained hardware with only small performance degradation.Frameworks including qKeras, Brevitas and ONNX Runtime support quantisation-aware or mixed-precision conversion.Source: §3.2.1 · pp. 7–8
• Maturing hardware-deployment tools and FPGA capabilities are expected to move ML models closer to LHC detector front-end electronics.The paper links this progression to increasing use of ML in hardware-based triggers.Source: §3.2.1 · p. 8

3.2.2 ML Models and Software

• User-friendly ML libraries and growing HEP expertise reduce barriers to complex models and duplicated MLOps work, but integration with experiment-specific trigger code remains crucial.The interface between existing trigger software stacks and evolving external libraries is identified as central to successful deployment.Source: §3.2.2 · p. 8

3.3 Associated Challenges

• Real-time ML trades latency or throughput against accuracy, while interpretability, out-of-distribution behavior, simulation mismodelling and unrecoverable trigger errors constrain reliable deployment.Models may need to fit on-chip or within the trigger network, and incorrect selections can permanently discard collision events.Source: §3.3 · pp. 8–9

4.1.1 Track Reconstruction in LHCb

• 15% more electrons that only leave tracks in the VELO are detected by ETX4VELO, while fake tracks fall below 1% from previously more than 2%.ETX4VELO uses sequential embedding, edge-classification, triplet-classification, and track-construction stages; its performance rivals Allen across track types.Source: §4.1.1 · p. 10

4.1.2 Electron and Photon Identification in ATLAS

• 50% lower overall CPU demand for single-electron triggers is achieved by Ringer’s greater rejection efficiency despite being marginally slower than cut-based selection.Ringer processes normalized energy sums from O(100) concentric calorimeter rings with region-specific MLP ensembles.Source: §4.1.2 · p. 10

4.1.3 Tau Reconstruction in CMS

• DeepTau improves efficiency by 10%–30% at fixed jet rejection, while loose-identification efficiency increases by 14% over baseline.Its convolutional and dense subnetworks jointly process local calorimeter information and high-level variables to classify tau, jet, muon, and electron candidates.Source: §4.1.3 · p. 11

4.1.4 Jet Flavour Tagging in ATLAS

• ATLAS Run 3 flavour taggers replace engineered high-level observables with permutation-invariant Deep Set and Transformer models using low-level vertex, track, and jet features.Their outputs form multi-class flavour probabilities and can provide early approximate preselections to reduce light-jet rates and downstream CPU cost.Source: §4.1.4 · pp. 11–12

4.1.5 Pile-up Mitigation

• PUMA outperforms classical pile-up suppression methods, improving missing-transverse-momentum resolution per event by 20%.PUMA uses sparse self-attention to reconstruct each particle’s hard-energy fraction while controlling the complexity of full self-attention.Source: §4.1.5 · pp. 12–13

4.1.6 Identifying Heavy-flavour Decays in LHCb

• The monotonic Lipschitz neural network improves on Run 2 BDTs while satisfying Run 3 LHCb trigger bandwidth and latency requirements.Bounding the Lipschitz constant limits classifier-response variation for nearby inputs, improving robustness to detector resolution and experimental instabilities.Source: §4.1.6 · p. 13

4.2.1 TPC Calibration in ALICE

• ALICE’s Run 3 streaming readout continuously processes and records every collision, making real-time calibration important for detector operation and rapid malfunction feedback.Source: §4.2.1 · p. 13
• The TPC supports charged-particle tracking and identification, while its space-charge density depends on pile-up ions accumulated during one ion-drift time.Source: §4.2.1 · p. 13
• TPC calibration corrects average space-charge distortions and their fluctuations during synchronous and asynchronous reconstruction.Source: §4.2.1 · p. 13
• A fast U-net CNN predicts distortion-fluctuation corrections from 3D space-charge maps on 90×17×17 or 180 × 33 × 33 grids.The target readout timing for fluctuation calibration is O (10 ms).Source: §4.2.1 · pp. 13–14
• Initial CNN performance appears to depend on the distance from the TPC’s fiducial boundaries, and development remains ongoing.Source: §4.2.1 · p. 14

4.3 Anomaly Detection

• Trigger-level anomaly detection enables less biased model-agnostic new-physics searches and broader data-quality monitoring earlier in the event-selection chain.The two anomaly-detection branches address noisy detector events and rare events inconsistent with the Standard Model, respectively.Source: §4.3 · p. 14

4.3.1 Data Quality Monitoring in ATLAS and CMS

• CMS’s online ECAL CNN autoencoder identifies approximately 99% of injected anomalies in validation data while reducing reported false alarms and simplifying expert decisions.The model reconstructs occupancy-histogram images using residual encoder and decoder networks, optimized with mean squared error on nominal detector operation.Source: §4.3.1 · p. 15
• ATLAS’s LSTM autoencoder models contiguous calorimeter time-series noise bursts, requires few modifications for real-time triggering, and extends to heavy-ion collisions without retraining.Most LAr calorimeter noise bursts last less than 60 s, motivating models that capture long-term temporal dependencies.Source: §4.3.1 · p. 15

4.3.2 New Physics Searches in CMS

• CMS deploys CICADA and AXOL1TL in the 4 µs hardware trigger, with approximate latencies of O(100) ns and 50 ns respectively.CICADA uses calorimeter images, while AXOL1TL uses up to 19 coarse L1 physics objects; both are trained on zero-bias data and synthesized with HLS4ML on Virtex-7 FPGAs.Source: §4.3.2 · p. 16
• These CMS deployments demonstrate that model-independent anomaly detection is feasible in severely resource-constrained real-time environments with extreme latency requirements.Faster inference allows more events, up to the 30 MHz collision rate, to be examined before model-dependent trigger and acquisition algorithms discard them.Source: §4.3.2 · p. 16

5.1 Common Challenges of Real-Time Analysis

• 30 MHz collision rates and approximately 1 PB/s streaming bandwidth make LHC triggers subject to exceptional throughput and latency constraints.The trigger accepts on average only 1 in 30,000 collision events because readout and storage cannot accommodate all events.Source: §5.1 · p. 17
• Shared HEP–industry challenges support common solutions including open-source frameworks, high-performance-computing tools, and cross-platform hardware-accelerator support.Both fields must reduce huge data volumes into manageable representations while meeting real-time constraints.Source: §5.1 · p. 17

5.2 Applications in Industry: Time-Series Anomaly Detection

• Less than 1% of credit-card transactions are fraudulent, creating severe class imbalance and scarce trustworthy labels for evaluating financial anomaly-detection models.Privacy regulations further constrain dataset size and completeness, while fraud labels are difficult to produce.Source: §5.2 · p. 17
• Transformer-based autoencoders encode sequential financial time series and use reconstruction error for anomaly scoring, comparing favourably with MERLIN, Isolation Forest, fully connected NNs, and LSTMs.The approach evaluates metrics, anomaly thresholds, and warning protocols because these choices affect model performance.Source: §5.2 · p. 18
• Synthetic financial data generation combines Markov-chain simulations of legitimate and fraudulent behaviour with multi-step LLM generation to address unreliable labelled data.The work remains ongoing but reports promising initial results.Source: §5.2 · p. 18

5.3 Applications in Industry: Computer Vision

• Less than 5 ms per image and under 30% FPGA resources were achieved by the best updated ENet model on a Zynq Ultrascale+ MPSoC, faster than previous FPGA implementations.The model used HLS4ML, quantisation-aware training, and a filter-kernel constraint balancing performance against resources.Source: §5.3 · pp. 18–19
• HLS4ML-based deployment of a distilled, quantised U-net on a ZCU102 achieved more than an eightfold performance improvement over Vitis AI while respecting resource constraints.The task segmented 256×256×3 satellite images to identify cloud-covered pixels.Source: §5.3 · p. 19

6 Conclusion

• ML has markedly improved real-time trigger and data-acquisition performance, in many cases approaching the state of the art achieved offline.Trigger systems have undergone significant changes since LHC Run 1.Source: §6 · p. 19
• Technological and algorithmic advances have enabled ML deployment in LHC trigger systems despite real-time constraints.The review presents examples of ML models operating during real-time data taking across the large LHC experiments.Source: §6 · p. 19
• Real-time ML may extend LHC physics programmes through model-agnostic anomaly detection that searches for discrepancies relative to the Standard Model.Such searches could enlarge the phase space explored for new physics.Source: §6 · p. 19
• CERN–industry collaboration has advanced open, accessible, and supported libraries for hardware deployment of ML.Continued knowledge exchange is described as important for further progress in both LHC triggers and industrial real-time applications.Source: §6 · pp. 19–20