Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-time Anomaly Detection at the L1 Trigger of CMS Experiment

Abhijith Gandrakota

TL;DR

This work addresses the risk that traditional L1 triggers may miss unexpected signatures by introducing unbiased, real-time anomaly detection at the CMS Level-1 trigger. Two complementary algorithms, AXOL1TL and CICADA, are developed and trained on trigger-less ZeroBias data and deployed on the CMS GT FPGA test crate to operate under stringent latency constraints. It details the end-to-end workflow from training with quantization-aware methods to FPGA deployment via HLS4ML, including monitoring and validation during proton collisions. During CMS data-taking, AXOL1TL demonstrated stable performance with orthogonal event selection relative to conventional triggers and showed potential efficiency gains for exotic signatures, illustrating a path toward model-independent triggering for BSM searches.

Abstract

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.

Real-time Anomaly Detection at the L1 Trigger of CMS Experiment

TL;DR

This work addresses the risk that traditional L1 triggers may miss unexpected signatures by introducing unbiased, real-time anomaly detection at the CMS Level-1 trigger. Two complementary algorithms, AXOL1TL and CICADA, are developed and trained on trigger-less ZeroBias data and deployed on the CMS GT FPGA test crate to operate under stringent latency constraints. It details the end-to-end workflow from training with quantization-aware methods to FPGA deployment via HLS4ML, including monitoring and validation during proton collisions. During CMS data-taking, AXOL1TL demonstrated stable performance with orthogonal event selection relative to conventional triggers and showed potential efficiency gains for exotic signatures, illustrating a path toward model-independent triggering for BSM searches.

Abstract

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.

Paper Structure

This paper contains 6 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Anomaly detection L1 triggers in CMS
  3. Inputs to L1 anomaly detection
  4. Network Architecture and FPGA implementations
  5. Performance of AXOL1TL
  6. Summary

Figures (5)

  • Figure 1: Components of the L1 trigger, along with the input and output paths for the anomaly detection algorithms.
  • Figure 2: Anomaly score distributions from AXOL1TL (left) and CICADA (right) as output by both HLS emulation and qkeras. Higher-scoring events are flagged as anomalous for further analysis.
  • Figure 3: Global trigger rate monitoring time series over the course of data-taking in June 2024, showing the rates of AXOL1TL seeds.
  • Figure 4: Left: Scores for all live AXOL1TL seeds and all events triggered by non-AXOL1TL HLT Scouting seeds, showing where the AXOL1TL contribution lies. Right: The distribution of AXOL1TL scores as a function of L1 object multiplicity.
  • Figure 5: Invariant mass distributions of pairs of jets (left), muons (center), and photons (right) from objects reconstructed from data scouting as triggered by the AXOL1TL nominal trigger path.