PaperlayerView exact revision on arXiv

Source-linked AI summary

Testing a Neural Network for Anomaly Detection in the CMS Global Trigger Test Crate during Run 3

Noah Zipper

arXiv:2312.10009v1hep-exphysics.ins-det

TL;DR

CMS Level-1 triggering must reduce detector data by more than 99% under stringent latency and stability constraints, while conventional strategies may miss unforeseen signatures. The paper deploys AXOL1TL, a VAE-based anomaly trigger, in the Global Trigger Test Crate and validates it with synthesis, emulation, and collision data. The implementation meets the stated latency requirement, reproduces trigger decisions exactly in ModelSim tests, and shows minimal hardware–emulation mismatches during data taking.

  • Problem

    CMS must reduce detector data by more than 99% in real time, while conventional triggers rely on prior signal knowledge or generic kinematic selections.

    Source: §1 · p. 2

  • Method

    AXOL1TL is a VAE-based anomaly trigger trained on unbiased CMS collision data and integrated into the Global Trigger Test Crate for hardware testing and monitoring.

    Source: §2 · pp. 2–3, §4 · p. 5

  • Results

    The firmware meets the 50 ns Level-1 latency requirement, achieves perfect trigger-bit agreement in ModelSim validation, and shows minimal hardware–emulation mismatches during collisions.

    Source: §3 · pp. 4–5, §4 · p. 5

  • Takeaways & Limitations

    The Test Crate enabled consistent live-data operation and validation of a signal-agnostic anomaly trigger without using its output for CMS event readout.

    Source: §4 · p. 5, §5 · p. 5

Abstract

from arXiv · show

We present the deployment and testing of an autoencoder trained for unbiased detection of new physics signatures in the CMS Level-1 Global Trigger (GT) test crate during LHC Run 3. 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 integration of the Neural Network into the GT test crate, and the monitoring, testing, and validation of the algorithm during proton collisions.

1 Introduction

CMS must reduce detector data by more than 99% in real time while preserving interesting physics events and avoiding dead time. AXOL1TL addresses limitations of conventional signal-specific or kinematic triggers by detecting anomalous events with a VAE trained on unbiased CMS data.

1 Introduction

  • More than 99% of CMS detector data must be reduced in real time by the FPGA-based Level-1 trigger.The system operates every 25 ns and must meet microsecond latency constraints without missing interesting events.Source: §1 · p. 2
  • Traditional trigger strategies depend on prior signal knowledge or generic kinematic selections.Source: §1 · p. 2
  • A VAE trained on unbiased CMS data provides a signal-agnostic approach to triggering on anomalous events.The passage describes this approach as applicable to signatures not specifically anticipated in advance.Source: §1 · p. 2

2 Anomaly Detection Trigger Algorithm

AXOL1TL uses a resource-conscious VAE to assign anomaly scores to standard Level-1 objects, with thresholds tested on unbiased 2023 collision data. The paper also reports improved BSM-signal efficiency and evaluates trigger-rate choices for deployment.

2 Anomaly Detection Trigger Algorithm

  • AXOL1TL uses a VAE with a small latent space to learn anomalous event structure from standard Level-1 objects.Inputs include missing transverse momentum, electrons/photons, muons, and jets represented by pT, η, and ϕ vectors.Source: §2 · p. 2
  • The design removes the decoder and simplifies the latent-space loss to reduce latency and resource utilization.At inference, the KL-divergence contribution is approximated by a mean-squared term, with no reported performance impact.Source: §2 · pp. 2–3
  • 10.5 million unbiased 2023 collision events were split evenly between training and anomaly-score threshold setting.Quantization-aware training emulated inference-time quantization with QKeras for the hardware implementation.Source: §2 · p. 3
  • AXOL1TL thresholds were chosen to estimate Level-1 trigger rates across operating points from 1 kHz to 10 kHz.The distributions compare QKeras Python scores with standalone HLS-emulator scores, while the table concerns efficiency gains for a specified BSM Higgs-decay signal.Source: Figure 2 · p. 3, §2 · pp. 3–4, Table 1 · p. 4
  • Adding AXOL1TL to the 2023 trigger menu was evaluated for efficiency improvement on a BSM Higgs-to-two-pseudoscalars signal.The model was trained on Run 3 ZeroBias events from Run 367883.Source: §2 · p. 3, Table 1 · p. 4

3 CMS Global Trigger Firmware

The AXOL1TL neural network was synthesized into CMS Global Trigger firmware and integrated into the μGT board architecture. Hardware tests show that the implementation meets latency requirements and reproduces reference trigger decisions exactly in ModelSim validation.

3 CMS Global Trigger Firmware

  • AXOL1TL firmware calculates the anomaly score alongside other global-trigger quantities and sends trigger decisions through standard links.The integration was designed to meet timing within the existing μGT board structure.Source: §3 · p. 4
  • High-level synthesis with hls4ml generated VHDL for the μGT FPGA and supported bitfile production for trigger-board configuration.Source: §3 · p. 4
  • The synthesized firmware meets the 50 ns Level-1 latency requirement and uses only a small fraction of available Virtex-7 resources.Vivado co-simulation evaluated latency and LUT, FF, DSP, and BRAM utilization.Source: §3 · p. 4, Table 2 · p. 4
  • ModelSim validation found perfect trigger-bit agreement between standalone emulation and hardware across four anomaly-score thresholds.The thresholds were >1250, >250, >25, and >5, using events from Run 368566.Source: §3 · p. 5, Table 3 · p. 5

4 Test Crate Implementation

The validated firmware was deployed on the CMS Global Trigger Test Crate, which receives the same inputs as production without controlling event readout. Real-time monitoring and collision-event comparisons showed consistent hardware behavior with minimal mismatches against emulation.

4 Test Crate Implementation

  • The Test Crate runs the μGT firmware on backup MP7 boards using production-like inputs without triggering CMS readout.This configuration enables testing new trigger strategies during data taking without interrupting event collection.Source: §4 · p. 5
  • Prometheus monitoring provides real-time trigger-rate measurements for anomaly-trigger paths during data taking.Source: §4 · p. 5
  • Collision events containing Test Crate information showed minimal mismatches and reasonable agreement between hardware and emulation.The comparison used trigger bits recorded in events saved by the production system.Source: §4 · p. 5

5 Summary

AXOL1TL was integrated into the CMS Level-1 trigger architecture and operated consistently in the Global Trigger Test Crate during 2023 collisions, with hardware validation against emulation.

  • AXOL1TL operated consistently during 2023 collisions in the CMS Global Trigger Test Crate, with thresholds probing an accessible trigger-rate range.The monitored time series covers four anomaly-detection thresholds, while the thresholds were not intended to model realistic trigger rates.Source: §5 · p. 5, Figure 4 · p. 6
  • The test crate enabled live monitoring and validation without using its output to trigger CMS readout.It received the same inputs as the production system while allowing anomaly-trigger paths to be monitored during data taking.Source: §5 · p. 5, Figure 4 · p. 6
  • Hardware and emulator trigger counts showed minimal mismatches, below 1%, in 2023 Ephemeral ZeroBias data.The comparison used hardware bits recorded from the configured test crate and standalone emulator counts evaluated offline.Source: §5 · p. 5, Figure 5 · p. 6, Table 4 · p. 6
Loading 2312.10009v1…