Design and FPGA Implementation of WOMBAT: A Deep Neural Network Level-1 Trigger System for Jet Substructure Identification and Boosted $H\rightarrow b\bar{b}$ Tagging at the CMS Experiment

TL;DR

This paper presents WOMBAT, a novel FPGA-implemented deep neural network trigger system for identifying boosted Higgs to bottom quark pairs at CMS, demonstrating improved efficiency and resource management for real-time particle detection.

Contribution

It introduces WOMBAT, a deep neural network-based trigger system with FPGA implementation, optimized for CMS Level-1 Trigger, and benchmarks its performance against existing methods.

Findings

WOMBAT achieves a 1 kHz trigger rate at 146.8 GeV jet pT threshold.

W-AM reduces the pT threshold to 140.4 GeV with acceptable efficiency.

FPGA implementation confirms resource feasibility with low latency.

Abstract

This thesis investigates the physics performance, trigger efficiency, and Field Programmable Gate Array (FPGA) implementation of machine learning (ML)-based algorithms for Lorentz-boosted $H\rightarrow b\bar{b}$ tagging within the CMS Level-1 Trigger (L1T) under Phase-1 conditions. The proposed algorithm, WOMBAT (Wide Object ML Boosted Algorithm Trigger), comprises a high-performance Master Model (W-MM) and a quantized, FPGA-synthesizable Apprentice Model (W-AM), benchmarked against the standard Single Jet 180 and the custom rule-based JEDI (Jet Event Deterministic Identifier) triggers. All algorithms process calorimeter trigger primitive data to localize boosted $H\rightarrow b\bar{b}$ jets. Outputs are post-processed minimally to yield real-valued $(\eta, \phi)$ jet coordinates at trigger tower granularity. Trigger rates are evaluated using 2023 CMS ZeroBias data (0.64 fb$^{-1}$), with efficiency assessed via a Monte Carlo sample of $H\rightarrow b\bar{b}$ offline reconstructed AK8 jets. W-MM achieves a 1 kHz rate at an offline jet $p_T$ threshold of 146.8 GeV, 40.6 GeV lower than Single Jet 180, while maintaining comparable signal efficiency. W-AM reduces the threshold further to 140.4 GeV, with reduced efficiency due to fixed-output constraints and limited multi-jet handling. FPGA implementation targeting the Xilinx Virtex-7 XC7VX690T confirms that W-AM meets resource constraints with a pre-place-and-route latency of 22 clock cycles (137.5 ns). In contrast, JEDI requires excessive resource usage and a 56-cycle latency, surpassing the 14-cycle L1T budget. Originally developed for Run-3 CMS L1T, WOMBAT serves as a proof-of-concept for Phase-2 triggers, where hardware advances will enable online deployment of more sophisticated ML-based L1T systems.