Source-linked AI summary
Real-time Anomaly Detection at the L1 Trigger of CMS Experiment
Abhijith Gandrakota
TL;DR
CMS must select a tiny fraction of 40 MHz collisions without losing unexpected new-physics signatures to predefined trigger criteria. The paper develops and deploys real-time autoencoder-based anomaly detection at the L1 Trigger, with AXOL1TL and CICADA using distinct inputs and architectures. AXOL1TL showed stable performance and event selection orthogonal to traditional triggers during CMS data-taking.
Problem
Predefined L1 criteria can miss unexpected new-physics signatures, motivating unbiased anomaly detection that does not rely heavily on specific BSM predictions.
Source: §1 · p. 2Method
The paper implements AXOL1TL and CICADA as real-time autoencoder-based anomaly detectors at the CMS L1 Trigger, using L1 objects or calorimeter energy deposits as inputs.
Source: §1 · p. 2, §2 · p. 3, §2.2 · p. 4Results
Takeaways & Limitations
Abstract
from arXiv · showhide
We present the preparation, deployment, and testing of an autoencoder trained for unbiased detection of new physics signatures in the CMS experiment Global Trigger (GT) test crate FPGAs during LHC Run 3. The GT makes the final decision whether to readout or discard the data from each LHC collision, which occur at a rate of 40 MHz, within a 50 ns latency. The Neural Network makes a prediction for each event within these constraints, which can be used to select anomalous events for further analysis. The GT test crate is a copy of the main GT system, receiving the same input data, but whose output is not used to trigger the readout of CMS, providing a platform for thorough testing of new trigger algorithms on live data, but without interrupting data taking. We describe the methodology to achieve ultra low latency anomaly detection, and present the integration of the DNN into the GT test crate, as well as the monitoring, testing, and validation of the algorithm during proton collisions.
1. Introduction
CMS must reduce 40 MHz proton-proton collisions to about 1,000 stored events per second, but predefined L1 criteria can miss unexpected new-physics signatures. The work motivates unbiased anomaly detection at the L1 Trigger using machine learning.
1. Introduction
- The CMS trigger reduces 40 MHz collisions through an FPGA-based L1 stage at about 100 kHz and an HLT stage at about 1 kHz before storage.L1 operates within a fixed 4 μs latency and discards the largest fraction of events.Source: §1 · p. 2
- Predefined selection criteria risk missing unexpected new-physics signatures that are not anticipated by the trigger menu.This limitation arises because only around 1,000 events per second can be stored for detailed offline analysis.Source: §1 · p. 2
- Unsupervised anomaly detection is pursued to identify events deviating from expected Standard Model patterns without relying heavily on specific BSM predictions.The motivation is to reduce selection bias while retaining sensitivity to a wide range of possible signatures.Source: §1 · p. 2
- The paper implements machine-learning-based real-time anomaly detection at the CMS L1 Trigger to identify potentially interesting events while maintaining high background rejection.It covers development, implementation, and preliminary deployment results.Source: §1 · p. 2
2. Anomaly detection L1 triggers in CMS
CMS develops AXOL1TL and CICADA as complementary model-independent anomaly-detection triggers. They use autoencoder-based scores, distinct L1 inputs, and FPGA-oriented implementations for real-time inference.
2. Anomaly detection L1 triggers in CMS
- AXOL1TL and CICADA provide complementary machine-learning approaches for identifying atypical events in real time without predefined new-physics models.Both are intended to capture potential BSM signatures through anomaly detection at L1.Source: §2 · p. 2
- Autoencoders learn dominant Standard Model patterns, then use reconstruction error to assign higher anomaly scores to poorly reconstructed rare or BSM-like events.The encoder compresses inputs into a latent representation and the decoder reconstructs the original input.Source: §2 · pp. 2–3
- The algorithms train on 2023 ZeroBias collisions at √s=13.6 TeV with pile-up 62, splitting the data equally between training and validation/performance testing.The passage reports an estimated 46% efficiency gain for AXOL1TL versus the rest of L1 at 1 kHz for exotic Higgs decays to four b quarks.Source: §2 · p. 3
2.1 Inputs to L1 anomaly detection
- AXOL1TL uses ten jets, four electron/photon objects, four muons, and MET, whereas CICADA uses an 18x14 calorimeter-energy image input.AXOL1TL uses raw hardware integer values for object pT, η, and ϕ.Source: §2.1 · p. 3
2.2 Network Architecture and FPGA implementations
- AXOL1TL uses a variational autoencoder with an eight-dimensional latent vector, while CICADA uses convolutional layers and MSE-based reconstruction for calorimeter deposits.CICADA is knowledge-distilled to mimic the MSE loss for ultra-low-latency inference.Source: §2.2 · pp. 3–4
- Quantization-aware training, hls4ml-generated HLS, and Xilinx Virtex-7 deployment support bit-exact FPGA inference for both algorithms.The implementations are tuned for ultra-fast inference and checked for bit-exactness.Source: §2.2 · p. 4
3. Performance of AXOL1TL
AXOL1TL operated stably during CMS data-taking across five anomaly-score thresholds. Its selected events were largely orthogonal to regular-trigger events, favored higher multiplicities, and showed no apparent invariant-mass sculpting in preliminary checks.
3. Performance of AXOL1TL
- AXOL1TL operated stably during data-taking across very tight, tight, nominal, loose, and very loose trigger thresholds.Nominal-seed events were sent to HLT scouting, while very-tight events received offline full-event reconstruction.Source: §3 · p. 4, Figure 3 · p. 5
- AXOL1TL-selected events tend to be orthogonal to the regular L1 menu, indicating that the anomaly trigger contributes a distinct event sample.The passage connects this sample to multijet final states and BSM signatures such as SUEPs that rule-based algorithms can overlook.Source: §3 · p. 4, Figure 4 · p. 5
- AXOL1TL and CICADA show a preference for assigning higher anomaly scores to events with higher L1 object multiplicity.This multiplicity dependence is visible in the AXOL1TL score distribution and is relevant to multijet selections.Source: §3 · p. 4, Figure 4 · p. 5
- Preliminary invariant-mass distributions for jet, muon, and photon pairs appear free of potential sculpting from the AXOL1TL trigger selection.The check uses objects reconstructed from data scouting through the nominal trigger path.Source: §3 · p. 4, Figure 5 · p. 5
4. Summary
The paper presents AXOL1TL and CICADA for real-time, model-independent anomaly selection at the CMS L1 Trigger. AXOL1TL uses a variational autoencoder and showed stable, orthogonal event selection during data-taking, while CICADA uses a convolutional autoencoder.
4. Summary
- The paper introduces AXOL1TL and CICADA as machine-learning anomaly detectors for real-time event selection at the CMS L1 Trigger.They target anomalous events without predefined selection criteria.Source: §4 · p. 5
- AXOL1TL analyzes L1 trigger objects with a variational autoencoder, whereas CICADA processes calorimeter energy deposits with a convolutional autoencoder.The two methods therefore use different input representations and network architectures.Source: §4 · pp. 5–6
- During CMS data-taking, AXOL1TL demonstrated stable performance and event selection orthogonal to traditional triggers.The summary identifies it as a promising tool for exploring novel physics signatures.Source: §4 · p. 6