Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Design and deployment of a fast neural network for measuring the properties of muons originating from displaced vertices in the CMS Endcap Muon Track Finder

Efe Yigitbasi

TL;DR

This work tackles the challenge of triggering on muons from displaced LLP decays by deploying a fast neural-network regression for muon properties within the CMS Level-1 Endcap Muon Track Finder. The proposed approach trains parallel NNs for transverse momentum $p_T$ and transverse impact parameter $d_0$, fuses them, and implements fixed-point, FPGA-friendly inference using hls4ml in a Virtex-7 device to achieve an end-to-end latency of $83$ ns. The solution improves trigger efficiency for displaced muons ($d_{\text{xy}}$ large) over the previous prompt BDT-based method and expands the CMS Run 3 LLP search phase space. The work demonstrates a practical, real-time ML deployment in an L1 trigger, increasing CMS sensitivity to LLP scenarios while respecting strict latency and resource constraints.

Abstract

We report on the development, implementation, and performance of a fast neural network used to measure the transverse momentum in the CMS Level-1 Endcap Muon Track Finder. The network aims to improve the triggering efficiency of muons produced in the decays of long-lived particles (LLPs). We implemented it in firmware for a Xilinx Virtex-7 FPGA and deployed it during the LHC Run 3 data-taking in 2023. The new displaced muon triggers that use this algorithm broaden the phase space accessible to the CMS experiment for searches that look for evidence of LLPs that decay into muons.

Design and deployment of a fast neural network for measuring the properties of muons originating from displaced vertices in the CMS Endcap Muon Track Finder

TL;DR

This work tackles the challenge of triggering on muons from displaced LLP decays by deploying a fast neural-network regression for muon properties within the CMS Level-1 Endcap Muon Track Finder. The proposed approach trains parallel NNs for transverse momentum and transverse impact parameter , fuses them, and implements fixed-point, FPGA-friendly inference using hls4ml in a Virtex-7 device to achieve an end-to-end latency of ns. The solution improves trigger efficiency for displaced muons ( large) over the previous prompt BDT-based method and expands the CMS Run 3 LLP search phase space. The work demonstrates a practical, real-time ML deployment in an L1 trigger, increasing CMS sensitivity to LLP scenarios while respecting strict latency and resource constraints.

Abstract

We report on the development, implementation, and performance of a fast neural network used to measure the transverse momentum in the CMS Level-1 Endcap Muon Track Finder. The network aims to improve the triggering efficiency of muons produced in the decays of long-lived particles (LLPs). We implemented it in firmware for a Xilinx Virtex-7 FPGA and deployed it during the LHC Run 3 data-taking in 2023. The new displaced muon triggers that use this algorithm broaden the phase space accessible to the CMS experiment for searches that look for evidence of LLPs that decay into muons.

Paper Structure

This paper contains 5 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Design and Hardware Implementation
  3. Deployment for Operations and Expected Performance
  4. Conclusion
  5. Acknowledgments

Figures (3)

  • Figure 1: The overall design of the complete displaced NN model. The complete model is obtained by stitching together two identical models with three dense layers and total of 19 nodes with one model targeting $p_T$ and the other targeting $d_0$.
  • Figure 2: The validation plots comparing measured and true $p_T$ (left) and $d_0$ (right).
  • Figure 3: The rate vs PU plot for one of the monitoring triggers recorded in LHC proton-proton collisions from July 2023 (left) which requires one muon in EMTF acceptance with $p_T$ assigned by the displaced NN greater than 10 GeV and no requirement on the $p_T$ assigned by the prompt BDT algorithm. The EMTF trigger efficiencies for prompt and displaced-muon algorithms for L1 $p_T$ > 10 GeV with respect to muon track $d_{\text{xy}}$ obtained using a displaced-muon simulation sample (right) Hayrapetyan_2024.