High-level hadronic tau lepton triggers of the CMS experiment in proton-proton collisions at $\sqrt{s}$ = 13.6 TeV

This paper summarizes the implementation and performance of new machine-learning-based algorithms in the CMS high-level trigger for identifying hadronically decaying tau leptons, utilizing proton-proton collision data at 13.6 TeV collected during 2022–2023.

The Gist

Imagine the Large Hadron Collider (LHC) at CERN as the world's most chaotic, high-speed particle collision factory. Every second, it smashes protons together billions of times, creating a blizzard of new particles. The CMS detector is the massive, ultra-sensitive camera trying to take a picture of this explosion.

But here's the problem: The camera is so fast and the explosions are so frequent that it would fill up its memory card in a split second if it tried to save everything. It needs a bouncer (the "Trigger System") to stand at the door and decide, in a microsecond, which collisions are worth saving and which are just boring background noise.

This paper is about upgrading that bouncer to be much smarter at spotting a very specific, elusive guest: the Tau Lepton.

The Mystery Guest: The Tau Lepton

Think of the Tau lepton as a "ghost" in the machine. It's a heavy cousin of the electron, but it lives for such a tiny fraction of a second (like a firefly that flashes once and vanishes) that it never reaches the camera. Instead, it decays instantly into other particles.

Sometimes, it decays into a "hadronic" spray—a small, tight bundle of particles that looks suspiciously like a messy pile of junk (called a "jet") created by ordinary quarks and gluons. Distinguishing a genuine Tau decay from a pile of junk is like trying to find a specific, rare type of origami swan in a pile of crumpled trash.

The Challenge: Too Much Noise

In recent years, the LHC has been cranked up to higher speeds. This means more collisions happen at the exact same time (a phenomenon called "pileup"). It's like trying to hear a whisper in a stadium that has suddenly doubled its crowd. The old bouncer (the system used in 2016–2018) was getting overwhelmed and starting to miss the real Taus or get confused by the noise.

The Solution: AI Bouncers

The CMS team realized they needed to upgrade their bouncer with Artificial Intelligence (AI). They didn't just want a bouncer who checks a list of rules (like "Is the object heavy? Is it in the right spot?"); they wanted a bouncer who could learn and intuitively recognize the pattern of a Tau lepton.

They introduced two new AI algorithms into the "High-Level Trigger" (the second, smarter layer of the bouncer system):

1. L2TAUNNTAG (The "Pattern Recognizer"):

   • What it does: This AI looks at the raw data from the detector's cameras and sensors. It uses a type of AI called a "Convolutional Neural Network" (similar to the tech in your phone that recognizes faces in photos).
   • The Analogy: Imagine looking at a blurry photo of a crowd. A human might struggle to pick out one person. But if you train an AI on thousands of photos of that specific person, it can spot them instantly, even if they are partially hidden or the photo is grainy. L2TAUNNTAG does this for Tau particles, filtering out the "junk" jets so only the promising candidates move to the next stage.
   • Result: It's faster and more accurate than the old rule-based system, allowing the experiment to keep the data flow manageable without losing the good stuff.

2. DEEPTAU (The "Expert Detective"):

   • What it does: This is a more advanced AI that acts as the final judge. It takes the candidates that passed the first filter and analyzes them in extreme detail, looking at tracks, energy deposits, and how the particles are arranged.
   • The Analogy: If L2TAUNNTAG is the security guard checking IDs at the door, DEEPTAU is the private investigator inside the club. It looks at the guest's behavior, their history, and their outfit to confirm, "Yes, this is definitely a real Tau lepton, not an imposter."
   • Result: It is incredibly good at telling the difference between a real Tau and a fake one, even when the detector is old and the crowd is huge.

The Results: A Smoother Party

The paper reports on data collected in 2022 and 2023 (the "early Run 3" era). The results show that:

• Efficiency went up: The new AI bouncers are catching more real Tau leptons than the old system did.
• Speed stayed the same: Despite being smarter, the AI doesn't slow down the process. It runs on special computer chips (GPUs) designed for this exact kind of math.
• Noise went down: They are better at ignoring the background junk, meaning physicists get cleaner data to study.

Why Does This Matter?

Why do we care about spotting these ghostly Taus? Because they are key to unlocking the secrets of the universe.

• The Higgs Boson: The Higgs particle loves to decay into Taus. To understand the Higgs better, we need to catch as many of these decays as possible.
• New Physics: If there are new, undiscovered particles (Beyond the Standard Model), they might also decay into Taus. By improving our ability to spot Taus, we increase our chances of discovering something completely new.

In summary: The CMS experiment upgraded its "bouncer" system with smart AI. Instead of just checking a checklist, the new system uses deep learning to recognize the unique "fingerprint" of a Tau lepton amidst a chaotic crowd of particles. This allows scientists to capture more rare events, leading to better discoveries about the fundamental laws of nature.

Technical Summary

1. Problem Statement

The CMS experiment at the Large Hadron Collider (LHC) faces increasing challenges in data acquisition due to higher instantaneous luminosities and a greater number of proton-proton interactions per bunch crossing (pileup). Specifically, identifying hadronically decaying tau leptons ($\tau_h$) at the trigger level is difficult because:

• Background Contamination: Genuine $\tau_h$ decays must be distinguished from the overwhelming background of quark- and gluon-initiated jets (QCD jets), which have similar topologies.
• Computational Constraints: The High-Level Trigger (HLT) has strict latency and processing time limits. Traditional "cut-based" algorithms (relying on fixed thresholds of simple observables) struggle to maintain high efficiency for genuine taus while suppressing the high QCD background rates in Run 3 conditions.
• Detector Aging: Despite an aging detector and increased pileup (averaging 46–52 interactions per crossing in 2022–2023), the trigger system must maintain or improve efficiency compared to Run 2.

2. Methodology

To address these challenges, the CMS collaboration deployed Machine Learning (ML) algorithms into the HLT for the first time specifically for $\tau_h$ candidates. The methodology involves a two-tiered ML approach replacing previous cut-based sequences:

A. L2TAUNNTAG (L2 Step)

• Architecture: A Convolutional Neural Network (CNN) designed to process data structured in a $5\times5$ grid in the $\eta-\phi$ plane.
• Inputs:
  • Global: Number of vertices (reconstructed via GPU-accelerated PATATRACKS).
  • Object Level: Properties of Level-1 (L1) $\tau_h$ candidates ($p_T$, $\eta$, $\phi$, isolation).
  • Detector Level: Calorimeter energy deposits (ECAL/HCAL) and GPU-based track information from the pixel detector (track multiplicity, scalar $p_T$ sum, impact parameters, $\chi^2$).
• Training: Trained on Monte Carlo (MC) samples of Drell-Yan, $t\bar{t}$, and $W$+jet events (signal) and QCD multijet events (background).
• Function: Replaces the L2 and L2.5 cut-based filtering steps to reduce event rates early in the HLT chain while preserving signal efficiency.

B. DEEPTAU (L3 Step)

• Architecture: A simplified version of the offline Deep Neural Network (DNN) used for $\tau_h$ identification. It utilizes a dual-grid structure: a dense inner "signal cone" ($11\times11$ cells) and an outer "isolation cone" ($21\times21$ cells).
• Inputs: High-level features of the reconstructed $\tau_h$ candidate combined with low-level information from the inner tracker, calorimeters, and muon subdetectors within the isolation cone.
• Simplifications for HLT: To meet timing constraints, the online version excludes electron and muon discriminants (as leptons are not the primary background for $\tau_h$ triggers) and omits complex primary vertex association fits.
• Function: Provides the final discrimination score against QCD jets at the L3 reconstruction stage.

C. Workflow Integration

The paper details the integration of these algorithms into specific HLT paths:

• Single-$\tau_h$ and Di-$\tau_h$ paths: Utilize L2TAUNNTAG to filter events before the full L3 reconstruction.
• Cross-triggers ($e\tau_h$, $\mu\tau_h$): Use DEEPTAU at the L3 stage for final identification.
• Reconstruction: The HPS (Hadron-Plus-Strips) algorithm is used to reconstruct $\tau_h$ candidates from Particle-Flow (PF) jets, identifying charged hadrons and neutral pions (strips).

3. Key Contributions

• First Deployment of ML in CMS HLT for $\tau_h$: This paper documents the inaugural use of neural networks (L2TAUNNTAG and online DEEPTAU) in the CMS HLT for hadronic tau identification.
• GPU Acceleration: The implementation leverages the new hybrid CPU/GPU farm and GPU-based track reconstruction (PATATRACKS) to handle the computational load of ML algorithms in real-time.
• Performance Validation: Comprehensive validation using 2022–2023 data ($\sqrt{s}=13.6$ TeV, $62 \text{ fb}^{-1}$) demonstrating that ML algorithms outperform legacy cut-based methods.
• Scale Factors: Derivation of data-to-simulation scale factors to correct for residual differences in trigger efficiency, ensuring physics analyses can use these triggers accurately.

4. Results

The performance was evaluated using data corresponding to an integrated luminosity of $62 \text{ fb}^{-1}$ and compared against Run 2 baselines and simulations:

• Efficiency vs. Rate Trade-off:
  • Di-$\tau_h$ Path: L2TAUNNTAG reduced the event rate to the next HLT step (from ~6.1 kHz to ~5.4 kHz in emulation) while improving absolute efficiency across most of the visible $p_T$ spectrum compared to the cut-based approach.
  • Single-$\tau_h$ Path: The introduction of L2TAUNNTAG reduced processing time by approximately 40% while maintaining similar event rates and efficiency levels.
• Identification Performance:
  • The online DEEPTAU algorithm demonstrated superior discrimination power against QCD jets compared to Run 2 cut-based selections across the majority of the $p_T$ range, maintaining similar total event rates.
  • Efficiency plateaus were observed to be robust against pileup (NPV dependence was minimal).
• Data/MC Agreement:
  • Scale factors (Data/MC efficiency ratios) were found to be close to unity ($\approx 1.0$) for $p_T > 60$ GeV, indicating excellent simulation modeling of the trigger response.
  • Discrepancies at low $p_T$ (< 60 GeV) were attributed to limitations in HCAL energy response modeling in simulation.

5. Significance

• Physics Reach: The improved trigger efficiency allows for the collection of a larger sample of genuine hadronic tau decays without increasing the computational cost or data storage rate. This is critical for:
  • Higgs Boson Physics: Precision measurements of $H \to \tau\tau$ decays.
  • BSM Searches: Enhanced sensitivity to Beyond-the-Standard-Model processes involving tau leptons (e.g., heavy $Z'$ bosons, supersymmetry).
• Operational Milestone: This work validates the feasibility of running complex, deep-learning-based algorithms in the real-time trigger environment of a high-luminosity collider, setting a precedent for future upgrades (e.g., High-Luminosity LHC).
• Resource Efficiency: By achieving higher efficiency at the same or lower rates, the CMS experiment maximizes its physics output per unit of computing resource, a crucial factor as the LHC luminosity continues to increase.