Physics Objects in CMS Run 3

This paper reviews the performance, calibration, and resilience of key physics objects in CMS Run 3 at 13.6 TeV, highlighting the deployment of transformer-based algorithms for heavy-flavor jet identification and boosted resonance reconstruction under high pileup conditions.

The Gist

Imagine the Large Hadron Collider (LHC) as the world's most powerful particle smasher. Inside the CMS detector, it smashes protons together at nearly the speed of light. The goal is to see what tiny pieces fly out of the crash, hoping to find new physics or measure known particles with extreme precision.

This paper is a progress report from the CMS team about how they are handling the data from Run 3 (the current phase of experiments). Here is the breakdown of their work, explained simply:

1. The "Crowded Room" Problem (Pileup)

Imagine trying to hear a single person whisper in a quiet room. Now, imagine that same room is suddenly packed with 60 other people all talking at once. That is what the LHC is like right now. Every time the machine fires a beam, it creates about 60 collisions at the exact same time. This is called "pileup."

• The Challenge: It's very hard to tell which particle came from the "main" collision you are interested in and which ones are just noise from the other 59 collisions.
• The Solution: The team has built new, smarter software algorithms that act like a super-powered noise-canceling headset. They can filter out the "background chatter" (the pileup) so the physicists can clearly hear the "whisper" (the interesting physics event).

2. The "Detective" Tools (Physics Objects)

To understand the collisions, the team needs to identify specific "clues" or physics objects. They have upgraded their toolkit for this new, crowded environment:

• Leptons (Electrons & Muons): These are like the "clean" messengers of the crash. The team has refined how they spot them, ensuring they don't get confused by the crowd. They use a "tag-and-probe" method (like checking a known ID card against a suspect) to make sure their measurements are accurate.
• Photons: These are flashes of light. The team has improved how they measure these flashes, making sure the "brightness" (energy) is calculated correctly even when the room is noisy.
• Jets: When quarks (tiny building blocks) fly out, they don't travel alone; they burst into a spray of other particles, forming a "jet." In the past, the team had to manually subtract the noise. Now, they use a new tool called PUPPI.
  • The Analogy: Imagine trying to count apples in a basket that also has a lot of confetti. Old methods tried to pick out every apple and ignore the confetti. PUPPI is like a smart scale that instantly knows which items are heavy apples and which are light confetti, adjusting the weight of the apples based on how much confetti is touching them. This makes the measurement of the apples much more accurate.

3. The "AI Brain" Upgrade (Machine Learning)

The biggest news in this paper is that the team is now using Transformer-based AI (the same type of technology behind modern chatbots) to identify complex patterns.

• Heavy Flavor Tagging: Sometimes, a jet comes from a heavy particle (like a "bottom" or "charm" quark). Identifying these is like finding a specific type of grain in a pile of sand. The old AI (DeepJet) was good, but the new AI models (ParticleNet and UParT) are like having a team of expert detectives who can look at the entire "cloud" of particles in a jet and instantly recognize the heavy ones with much higher accuracy.
• Boosted Objects: Sometimes particles are moving so fast they squash together. The new AI can spot these "squashed" particles (like a boosted top quark) much better than before, rejecting background noise 10 times more effectively.

4. The "Invisible" Clue (Missing Momentum)

Sometimes, particles fly out of the detector that we can't see (like neutrinos). We know they are there because the total balance of energy doesn't add up.

• The team has upgraded how they calculate this "missing money" (missing momentum). By using the new PUPPI system and a new deep learning tool called DeepMET, they can calculate exactly how much "invisible" energy is missing, even in the noisy, crowded environment.

5. The "Simulation" (The Practice Run)

Before they analyze real data, they run millions of "practice collisions" on computers (Monte Carlo modeling).

• The paper notes that their computer simulations of top quarks (heavy particles) have been fine-tuned to match reality much better than before. They have adjusted the "rules" of the simulation (like how particles bounce off each other) so that the virtual data looks exactly like the real data.

The Bottom Line

The CMS team has successfully upgraded their software to handle a much noisier, more crowded environment than ever before. By switching to PUPPI for cleaning up the data and using Transformer AI to identify complex particles, they are getting clearer, more precise results. This sets the stage for them to continue making world-class discoveries about the fundamental building blocks of the universe in the coming years.

Technical Summary

Technical Summary: Physics Objects in CMS Run 3

Problem Statement
The CMS experiment is currently operating in Run 3 of the Large Hadron Collider (LHC) at a center-of-mass energy of $\sqrt{s} = 13.6$ TeV. A defining characteristic of this data-taking period is high instantaneous luminosity, resulting in an average of approximately 60 proton-proton interactions per bunch crossing (pileup). This high-pileup environment poses significant challenges for the reconstruction and calibration of physics objects (charged leptons, photons, jets, and missing transverse momentum). The primary problem addressed is maintaining sensitivity to rare processes and precision measurements under these harsh conditions while managing the computational demands of reconstruction. Additionally, there is a need to improve the identification of heavy-flavor jets and boosted resonances, which are critical for top quark physics and beyond-Standard-Model searches.

Methodology
The paper outlines a comprehensive upgrade in reconstruction algorithms and calibration strategies for Run 3:

• Reconstruction Framework: The experiment continues to rely on the Particle Flow (PF) algorithm, which combines tracker, calorimeter, and muon system information. A novel development is the Machine Learning Particle Flow (MLPF) algorithm, which utilizes a transformer model to infer particle content directly from tracks and clusters.
• Leptons and Photons:
  • Muons: Reconstruction links silicon tracker tracks with muon system segments. Identification employs both cut-based and multivariate techniques, with specific algorithms for low and high $p_T$. Momentum corrections address magnetic field mismodeling and alignment biases.
  • Electrons/Photons: Electrons use the Gaussian Sum Filter (GSF) for bremsstrahlung and "mustache" superclusters. Energy regression based on shower shape and pileup density is applied. A new tag-and-probe method using unbinned likelihood fits and normalizing flows is introduced for efficiency mapping.
• Jets and Pileup Mitigation: The collaboration has fully adopted the Pileup Per Particle Identification (PUPPI) algorithm as the default for pileup mitigation, replacing the previous Charged Hadron Subtraction (CHS). PUPPI uses local shape information to rescale particle four-momenta, minimizing pileup contributions. Jet calibration involves L1 offset corrections (largely unnecessary for PUPPI jets), simulation-based response corrections, and residual data/simulation corrections derived from dijet and $\gamma/Z$+jet events.
• Tagging Algorithms: Heavy-flavor and boosted object identification has transitioned from Recurrent Neural Networks (e.g., DeepJet) to transformer-based architectures:
  • ParticleNet: A graph neural network treating jet constituents as an unordered "particle cloud."
  • UParT (Unified Particle Transformer): Leverages pairwise interaction features with adversarial training for robustness, enabling strange-jet tagging and quark/gluon discrimination.
  • DeepTau: An upgraded algorithm for tau lepton identification with significantly lower jet mis-identification probabilities.
  • Boosted Tops: ParticleNet and ParT are applied to large-radius AK8 jets for top quark identification.
• Missing Transverse Momentum ($p_T^{miss}$): The default has shifted to PUPPI-based $p_T^{miss}$ due to reduced pileup dependence. A new DeepMET algorithm, based on deep neural networks, learns optimal weights for PF candidates to improve resolution.
• Monte Carlo Modeling: Signal and background simulations utilize the Powheg-hvq generator interfaced with Pythia 8 (CP5 tune). Improvements include better modeling of initial/final state radiation, top $p_T$ spectra, and Color Reconnection (CR) models tuned to CMS data.

Key Contributions and Results

• Performance under High Pileup: The adoption of PUPPI has rendered boosted object identification more resilient against pileup and strongly suppressed pileup jets within tracker coverage. The pileup energy offset for PUPPI jets is near zero, reducing the need for traditional L1 corrections.
• Computational Efficiency: The MLPF algorithm achieves a factor of 2 speedup on modern heterogeneous computing infrastructure (CPU+GPU) compared to standard PF, while demonstrating comparable physics performance.
• Identification Improvements:
  • Heavy Flavor: The new transformer-based taggers (ParticleNet, UParT) have been successfully commissioned.
  • Tau Leptons: DeepTau v2.5 shows a 30–50% lower jet mis-identification probability compared to the previous v2.1.
  • Boosted Tops: The new taggers provide up to 10 times better background rejection than the mass-decorrelated DeepAK8 tagger used in previous runs and show improved top quark mass resolution.
  • Jet Energy: Auxiliary jet $p_T$ regression tasks in the new architectures improve jet energy resolution by up to 20% after Monte Carlo corrections.
• Precision Metrics:
  • Luminosity: Preliminary uncertainty for 2023 data is $\pm 1.3\%$, with a refined Run 2 analysis yielding a total integrated luminosity of $137.88 \pm 1.01 \text{ fb}^{-1}$ (0.73% uncertainty).
  • Muon Momentum: Standard calibration achieves $\approx 0.05\%$ precision; dedicated analyses (e.g., W mass) reach $0.006\%$.
  • Jet Energy: Run 2 total uncertainty is at the 1% level; Run 3 calibrations for 2022–2024 datasets show improvements, particularly in the endcap due to better calorimeter calibration.

Significance
The paper asserts that the physics objects used for CMS analysis have been successfully commissioned and calibrated for the high-pileup environment of LHC Run 3. The widespread adoption of the PUPPI algorithm and the integration of transformer-based tagging represent significant performance improvements despite the harsher collision conditions. Combined with precise luminosity determination and refined Monte Carlo modeling, these advancements position the CMS collaboration to continue delivering precision top quark physics results in the coming years. The transition to machine learning-driven reconstruction and identification (MLPF, ParticleNet, UParT, DeepMET) marks a critical evolution in handling the complexity of Run 3 data.