ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM

ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM August 2025 AUTHOR(S): Kalle Pakarinen NextGen Triggers, Task 2.1 SUPERVISOR(S): Nikolaos Konstantinidis Ioannis Xiotidis Mario Campanelli CERN openlab Report x/2025 PROJECT SPECIFICATION All major CERN experiments along with the accelerator will enter soon in the phase of a big upgrade cycle, called the High-Luminosity LHC, in order to further broaden the physics reach. ATLAS has a plethora of upgrades concerning the HL-LHC era which will equip the detector with many exciting new opportunities. One of the core upgrades in ATLAS concerns the way of reading out the calorimeter sub-detector. In contrast to previous runs ATLAS will be able to read the full granularity of the calorimeter at the level of the hardware trigger system. Having this information available in such a challenging environment provides a unique opportunity to explore Machine Learning ideas on the edge within the context of the ATLAS Global Trigger. The selected student will be able to work on ultra-fast algorithms that use the calorimeter information and be able to reconstruct physics properties of interest against noise that is generated from the secondary collisions present in every bunch crossing (pileup). The project will explore definitions of more sophisticated physics quantities applied on jet reconstruction, like mass or substructure variables, that can aid specific selection strategies of ATLAS and expand into the area of anomaly detection where Machine Learning has been proved to enormously assist. With this work we aim to further expand the physics reach of ATLAS and ensure that the potential biases introduced during the event selection stages do not affect the quality of physics recorded. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 1 CERN openlab Report x/2025 ABSTRACT This work investigates the use of unsupervised anomaly detection for the ATLAS trigger system in the High-Luminosity LHC (HL-LHC) environment. To address the challenge of identifying non-standard events without relying on predefined signatures, a sparse convolutional autoencoder is trained on pile-up suppressed calorimeter tower images containing only QCD dijet background. The network learns to reconstruct typical energy deposition patterns, and the reconstruction error is used as an anomaly score. Submanifold sparse convolutions are employed to preserve the sparsity pattern and ensure accurate, localized reconstruction without introducing artificial energy deposits. Evaluation shows that the model can distinguish Higgs pair production events (HH →b¯ bb¯ b) from background, with higher reconstruction errors corresponding to anomalous spatial features. The results demonstrate the viability of sparse autoencoders for model-independent event filtering at the trigger level and motivate further analysis into the features driving anomaly detection performance. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 2 CERN openlab Report x/2025 TABLE OF CONTENTS 1 Introduction 4 2 Background 4 2.1 ATLASexperiment.................................. 4 2.2 Triggersystem .................................... 5 2.3 Anomalydetection .................................. 6 3 Implementation 8 3.1 Autoencoders ..................................... 9 3.2 Submanifold sparse convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 3.3 Autoencoder structure and training . . . . . . . . . . . . . . . . . . . . . . . . . 10 4 Results 13 4.1 Reconstruction error distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 13 4.2 Featuredistributions ................................. 15 5 Discussion 17 ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 3 CERN openlab Report x/2025 1 Introduction At the HL-LHC, the ATLAS experiment will face a challenging data environment with significantly increased event complexity and pile-up conditions. Although the collision rate from the machine remains at 40 MHz, the upgraded trigger and data acquisition system will permit a higher event acceptance, with rates of about 1 MHz at Level-0 and 10 kHz recorded to tape by the Event Filter. In this context, real-time event selection becomes more demanding, requiring both high efficiency for standard physics processes and flexibility to retain sensitivity to rare or unforeseen signals. Traditional trigger systems rely on fixed selection rules or supervised models trained on known signal topologies. However, this approach can be limited in scope and may fail to capture new or unexpected phenomena. To address this, anomaly detection techniques—particularly those based on unsupervised learning—are being explored as a complementary tool within the upgraded ATLAS trigger architecture. This work investigates the use of a sparse convolutional autoencoder as an anomaly detection model operating on pile-up suppressed calorimeter tower images. The approach is unsupervised: the network is trained only on QCD dijet background events and learns to reconstruct typical spatial energy patterns. Anomalous events, such as those from non-QCD processes, are expected to deviate from this pattern and yield higher reconstruction error. To support efficient learning on sparse calorimeter data, the network employs submanifold sparse convolutional layers, which preserve the sparsity pattern throughout the network and avoid introducing artificial activations in inactive regions. This allows accurate reconstruction of energy deposits while maintaining computational efficiency. 2 Background 2.1 ATLAS experiment The ATLAS detector at CERN’s Large Hadron Collider (LHC) is one of the most complex scientific instruments ever built. Located 100 m underground, it measures 46 m in length, 25 m in diameter, and weighs about 7,000 tonnes. The LHC accelerates two counter-rotating proton beams to nearly the speed of light and collides them at center-of-mass energies up to 14 TeV, producing up to a billion interactions per second. To handle this volume, ATLAS employs a sophisticated trigger and data acquisition system to reduce the raw 40 MHz collision rate to a manageable stream for offline analysis. [3] ATLAS consists of three primary subsystems: the Inner Detector (ID), the calorimeters, and the muon spectrometer, all enclosed within a powerful magnet system. The ID, immersed in a 2 T solenoidal magnetic field, measures charged-particle trajectories with micrometer precision using three technologies: the Pixel Detector (∼100 M channels), the SemiConductor Tracker (SCT, ∼6 M channels), and the Transition Radiation Tracker (TRT, ∼350 000 straws). Surrounding the ID, the calorimeter system measures the energies of electrons, photons, hadrons, and jets across a pseudorapidity range of |η|<4.9. It comprises two main components: •Liquid Argon (LAr) calorimeter: operates at −183◦C, providing high-precision, finegranularity measurements of electromagnetic showers. •Tile Calorimeter: built from nearly 500,000 scintillator tiles, it measures hadronic energy deposits. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 4 CERN openlab Report x/2025 The calorimeters are segmented both longitudinally and transversely, enabling accurate reconstruction of shower shapes, efficient particle identification, and precise measurements of missing transverse momentum. They provide both coarse-granularity inputs for fast trigger decisions and full-resolution data for offline reconstruction. The outermost muon spectrometer, extending to |η|<2.7, uses drift tubes, cathode strip chambers, and resistive plate chambers within a system of large superconducting toroids to deliver precise muon momentum measurements. Overall, ATLAS integrates nearly 100 M readout channels and over 3,000 km of cabling, producing several petabytes of data annually, analyzed worldwide via the LHC Computing Grid by a collaboration of more than 3,000 scientists from 38 countries. A schematic image of the ATLAS detector is presented in Figure 1. Figure 1: Schematic cross-sectional view of the ATLAS detector, adapted from: [4]. 2.2 Trigger system The ATLAS experiment employs a multi-level trigger system to reduce the proton–proton collision rate from the LHC’s 40 MHz bunch crossing frequency to approximately 1 kHz of events that are stored for offline analysis. Given the volume of data and the limited bandwidth for storage and processing, the trigger system plays a critical role in ensuring that only events with potential physics interest are retained. [8] The trigger system consists of two primary levels: the Level-1 (L1) trigger and the HighLevel Trigger (HLT). The L1 trigger is hardware-based and uses coarse-granularity information from the calorimeter and muon systems to make a decision within a latency of a few microseconds, reducing the event rate to around 100 kHz. Events accepted by L1 are sent to the HLT, a software-based system running on a computing farm. The HLT applies more refined algorithms using full detector granularity and detailed reconstruction techniques to further reduce the rate to approximately 1 kHz. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 5 CERN openlab Report x/2025 With the upcoming High-Luminosity LHC (HL-LHC), the ATLAS experiment faces a significant increase in pile-up, requiring substantial upgrades to the trigger system. The NextGeneration Trigger (NextGen) program addresses this by exploring the deployment of advanced techniques, such as machine learning algorithms, within the existing Phase-II trigger architecture to enhance its capability and flexibility in handling the increased data rates and event complexity expected at the HL-LHC. The Level-0 trigger operates with an extended latency and enhanced input bandwidth, allowing for the incorporation of higher-granularity information from the calorimeters and precision data from the muon spectrometer at the earliest stage. It includes several key subsystems: the L0 Calorimeter Trigger (L0Calo), the L0 Muon Trigger (L0Muon), the L0 Topological Processor (L0Topo), and the L0 Global Trigger (L0Global), which combines and oversees information from the other L0 subsystems to make final trigger decisions. The L0 system outputs events at a rate of up to 1 MHz, significantly higher than the current L1 system. [8] Additional upgrades include new front-end electronics for the calorimeters and muon systems to support high-rate readout, and a revamped data acquisition (DAQ) system capable of handling the increased throughput. These enhancements allow for more sophisticated selection criteria and improved efficiency for a wide range of physics processes. Overall, the NextGen trigger system ensures that ATLAS maintains high performance and flexibility in the challenging conditions of the HL-LHC, enabling the experiment to pursue its physics goals with optimal sensitivity. 2.3 Anomaly detection Anomaly detection refers to the identification of patterns or events in data that deviate from the expected behavior. It is commonly applied in contexts where large volumes of data are produced, and rare or unknown events are of particular interest. Anomalies may indicate critical occurrences such as faults, outliers, or novel phenomena. When real-time decision making is required, efficient and accurate anomaly detection methods become essential to support timely and informed actions. In particle physics, anomaly detection is employed to identify rare or previously unobserved physical processes, potential signals of new physics, or detector malfunctions. Given the data volume and complexity of modern collider experiments, combined with the lack of clear indications of what beyond-the-Standard-Model phenomena might look like—especially after the discovery of the Higgs boson—traditional rule-based or supervised approaches are insufficient for exhaustive coverage of all possible scenarios. As such, machine learning-based unsupervised or weakly supervised methods have become increasingly relevant, allowing the system to learn standard behaviors from data and flag deviations without requiring explicit labeling. In the context of the HL-LHC, the trigger system faces increasing challenges due to, increased pile-up, and the need to retain sensitivity to rare or unexpected phenomena. To address these issues, anomaly detection is being considered as a complementary strategy to conventional trigger algorithms. Rather than relying solely on predefined signal signatures, anomaly detection methods aim to identify deviations from learned patterns of Standard Model processes, offering sensitivity to unforeseen or non-standard events that might otherwise be missed. The NextGen project is a collaborative initiative aimed at exploring the integration of machine learning techniques into the ATLAS Trigger and Data Acquisition (TDAQ) system for the HL-LHC. The focus is on investigating real-time, low-latency algorithms deployable in both hardware and software trigger stages, particularly those capable of adapting to evolving detector conditions. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 6 CERN openlab Report x/2025 Figure 2: Schematic overview of the ATLAS Phase-II Trigger and Data Acquisition (TDAQ) architecture. The figure illustrates the upgraded data flow from detector subsystems to the Global Trigger system in the HL-LHC era. Adapted from [1]. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 7 CERN openlab Report x/2025 3 Implementation At the HL-LHC, the experimental environment will be characterized by an average of about ⟨µ⟩= 200 additional proton–proton interactions per bunch crossing. These simultaneous interactions, referred to as pile-up, introduce large numbers of additional energy deposits in the calorimeters that are uncorrelated with the hard-scatter event of interest, i.e., the primary highmomentum interaction producing the physics process under study. If left untreated, pile-up distorts the reconstructed objects, obscures the true event topology, and complicates the identification of rare processes. In particular, calorimeter-based observables become dominated by these contributions, reducing the efficiency of triggering and degrading the accuracy of physics measurements. For this reason, effective pile-up suppression is essential when working with HL-LHC simulation data. The datasets used in this work correspond to HL-LHC simulation samples with an average pile-up of ⟨µ⟩= 200. Prior to analysis, the calorimeter tower images are preprocessed using a dedicated pile-up suppression method based on convolutional neural networks (CNNs), as described in [2]. This approach exploits spatial correlations in the η–ϕplane and across calorimeter layers to distinguish true energy deposits from pile-up contributions. The CNN model, trained on simulated data, learns to identify the structured signatures of hard-scatter interactions while removing diffuse pile-up noise. As a result, the output images retain the localized energy patterns relevant for physics analysis, enabling more robust input for trigger-level anomaly detection in the HL-LHC regime. 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 0 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 1 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 2 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 3 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 4 21012 3 2 1 0 1 2 3 ATLAS Simulation Internal p s =14TeV Layer 5 0 1 2 3 4 5 6 ET [GeV] Figure 3: Example of a pile-up suppressed calorimeter image representing energy deposits across six calorimeter layers. Each subplot corresponds to a distinct layer in a 50 ×64 ×6 image, where suppression has been applied using a convolutional neural network (CNN). ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 8 CERN openlab Report x/2025 and accurately reconstruct various QCD-like patterns. Notably, even though the number of JZ1 events in the training set was not increased, the reconstruction error for JZ1 improves significantly, indicating that additional diversity in training improves performance even on regions already represented in the dataset. Most importantly, the HH4b distribution remains clearly distinct. While many signal events still yield low reconstruction error (i.e., in the first few bins), a much larger fraction of HH4b events extend into the high-error region compared to the background. This separation enables effective anomaly detection. For example, applying a cut on the reconstruction error at a moderate value (e.g., error > 0.5) would retain a substantial portion of signal events while selecting only a negligible fraction of background. These results demonstrate that the reconstruction error is sensitive both to training coverage and to physical differences between background and signal events. Broader background training improves reconstruction quality and sharpens separation from signal-like outliers, enhancing the effectiveness of unsupervised anomaly detection. 4.2 Feature distributions To better understand what distinguishes the signal (HH4b) from the background (JZ1, JZ2, JZ3), we examine the distributions of basic jet features—transverse momentum (pT), pseudorapidity (η), and azimuthal angle (ϕ)—for both datasets. These distributions offer insight into whether the autoencoder’s anomaly detection capability correlates with physical characteristics of the jets. Figure 7shows the pT,η, and ϕdistributions for all jets in the HH4b (left) and combined JZ1, JZ2, JZ3 (right) samples. Both sets span the same range in ηand ϕ, with very similar distributions for ϕ(the slight excess at the edges of the ϕrange arises from binning near the 2πwrap-around, where some entries may be counted twice). The HH4b ηdistribution, however, exhibits a preference for central jets (|η| ≈ 0), whereas the background is more uniform across the detector acceptance. For pT, the HH4b sample covers a slightly broader spectrum, as expected from Higgs boson decay products, while the background distributions are more sharply peaked and fall off faster at high pT. These subtle differences, particularly in pTshape and ηlocalization, may contribute to the autoencoder’s ability to distinguish between signal and background. The signal jets appear to be more central and slightly higher in energy, potentially resulting in localized energy patterns not seen in the background training samples. To further study what kinds of events are identified as anomalous, Figure 8shows the same feature distributions, but restricted to events with a reconstruction error greater than 0.5. These represent the subset of events where the autoencoder fails to accurately reproduce the input, and are thus flagged as anomalous. Due to the model’s strong reconstruction ability on background, very few JZ events exceed this error threshold, leading to limited statistics in the right column. Nonetheless, a comparison is still informative. In ϕ, both signal and background remain fairly uniform, indicating no strong ϕ-dependence in the anomaly score. In η, however, a same difference is before can be observed: the HH4b jets selected as anomalous are more concentrated near the center of the detector (|η|<0), while the few background events passing the error cut are more broadly distributed. This suggests that the model learns a mild sensitivity to η, especially after training on a diverse set of background events. The HH4b jets are more central and may present more localized energy patterns that differ from the more dispersed QCD jets. For pT, the anomalous HH4b jets span a broader range than the anomalous background jets, ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 15 CERN openlab Report x/2025 HH4b, whole sample JZ1, JZ2, JZ3, whole sample Figure 7: Feature distributions for jets in HH4b (left) and JZ1, JZ2, JZ3 (right), using the full sample. Rows correspond to pT,η, and ϕ. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 16 CERN openlab Report x/2025 with a slightly harder tail. The shape differences in pT, combined with the higher likelihood of finding dense, central jets in signal events, likely contribute to why these are reconstructed poorly and flagged as anomalous. This indicates that the autoencoder is not simply learning to reject high-pTor central jets but is sensitive to the underlying spatial and energy structure in the full 3D calorimeter image. HH4b, reconstruction error >0.5JZ1, JZ2, JZ3, reconstruction error >0.5 Figure 8: Feature distributions for jets in HH4b (left) and JZ1, JZ2, JZ3 (right), for events with reconstruction error greater than 0.5. Rows correspond to pT,η, and ϕ. 5 Discussion The results demonstrate that the autoencoder is capable of distinguishing between the HH4b signal and the QCD background represented by the JZ1, JZ2, and JZ3 samples. A key enabler of this performance is the use of submanifold sparse convolutions, which preserve the original sparsity pattern of the calorimeter data throughout the network and allow for efficient learning of localized spatial structures. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 17 CERN openlab Report x/2025 From the reconstruction error distributions, it is clear that the model assigns low reconstruction errors to background-like events and significantly higher errors to signal-like events, particularly when trained on a broad set of background processes. This indicates that the network learns to model standard QCD jet topologies effectively and can flag non-standard structures as anomalous. The jet feature distributions provide further context. While signal and background jets have broadly similar kinematic properties, slight differences are visible—most notably, signal jets tend to be more central (closer to η= 0) and exhibit a broader pTdistribution. These features likely contribute to the model’s ability to separate the two classes, but the differences alone do not fully explain the observed reconstruction error patterns. This points to an important direction for further research. Since the input feature distributions are only slightly different between signal and background, it is not yet fully understood what specific aspects of the input cause the model to assign high anomaly scores. A more detailed analysis of the latent space—e.g., through Principal Component Analysis (PCA), t-SNE, or layer-wise feature correlation studies—could provide insight into what representations the network has learned and what features it considers anomalous. Another important next step is to evaluate the model in a broader context by including a wider variety of physics processes, both in training and evaluation. This would help clarify the model’s generalization behavior and its robustness in distinguishing between typical and non-standard events under different conditions. Overall, the use of sparse autoencoders with submanifold convolutions shows strong potential for anomaly detection in the trigger context, and further interpretability studies will be important for building confidence and improving performance. ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 18 CERN openlab Report x/2025 REFERENCES [1] A. Koulouris et al. “Phase-II Upgrade of the ATLAS L1 Central Trigger”. In: (2024). ATLDAQ-PROC-2024-007. url:https://cds.cern.ch/record/XXXXXXX. [2] ATLAS Collaboration. “Convolutional Neural Networks for Pile-Up Suppression in the Global Trigger”. In: (2025). ATL-UPGRADE-PUB-2025-002. url:https://cds.cern. ch/record/2887853. [3] ATLAS Collaboration. “The ATLAS Experiment at the CERN Large Hadron Collider”. In: (2008). doi:10.1088/1748-0221/3/08/S08003. [4] ATLAS Experiment. Schematic Diagram of the ATLAS Detector. 2014. [5] Benjamin Graham. SparseConvNet.https://github.com/facebookresearch/SparseConvNet. Accessed: 2025-08-29. 2017. [6] Benjamin Graham and Laurens van der Maaten. Submanifold Sparse Convolutional Networks. 2017. eprint: 1706.01307.url:https://arxiv.org/abs/1706.01307. [7] Benjamin Kreis. “Particle Flow and PUPPI in the Level-1 Trigger at CMS for the HLLHC”. In: (Aug. 2018). doi:10.48550/arXiv.1808.02094. [8] Nuno dos Santos Fernandes and The ATLAS Collaboration. “ATLAS Trigger and Data Acquisition Upgrades for the High-Luminosity LHC”. In: (2024). ANOMALY DETECTION FOR THE ATLAS TRIGGER SYSTEM 19