Minimising Event Size, Maximising Physics: Inclusive Particle Isolation for LHCb's Run 3

To address the data volume challenges of LHCb's Run 3, this paper introduces the novel Inclusive Multivariate Isolation (IMI) algorithm, which reduces event size by 45% while maintaining 99% signal efficiency and superior background rejection compared to traditional methods.

The Gist

Imagine the Large Hadron Collider (LHC) at CERN as the world's most powerful, high-speed camera. It doesn't take photos of landscapes; it takes "snapshots" of subatomic collisions happening billions of times a second.

The LHCb experiment is a specific camera designed to study heavy particles (like "beauty" quarks) that decay into other particles. In the new "Run 3" phase, this camera is shooting at a rate five times faster than before.

The Problem: The "Digital Hoarding" Crisis
Here is the catch: Every time the camera takes a picture (an "event"), it captures everything in the frame.

• The Signal: The actual physics we care about is like a tiny, precious diamond in the center of the photo. It usually involves just a few particles (maybe 3 to 7).
• The Noise: The rest of the photo is filled with hundreds of other particles flying around from the collision. These are like background clutter, dust, and random debris.

In the past, the experiment would save the entire photo to a hard drive. But with the new speed, the data is piling up so fast that the hard drives are about to overflow. If they keep saving every single particle, they will run out of storage space and have to delete the precious diamonds just to make room for the trash.

The Old Solution: The "Cone" and the "Fence"
Previously, scientists used simple rules to decide what to keep.

• The Cone: Imagine drawing a circle around the diamond. If a particle is inside the circle, keep it. If it's outside, throw it away.
• The Fence: Imagine a fence around the diamond's starting point. If a particle looks like it came from that spot, keep it.

These rules worked okay in the past, but in the new, crowded environment, they are too clumsy. They either throw away too many diamonds (losing physics) or keep too much trash (filling up the hard drive).

The New Solution: The "Smart Bouncer" (IMI)
This paper introduces a new, super-smart algorithm called Inclusive Multivariate Isolation (IMI). Think of IMI as a highly trained, intuitive bouncer at an exclusive club.

Instead of just measuring distance (like the old "cone" method), the bouncer looks at the behavior of every single particle in the room.

• The Question: "Did you come from the same party as the diamond, or are you just a random stranger from the street?"
• The Clues: The bouncer checks the particle's speed, its angle, how far it traveled, and its history.
• The Decision:
  • If the particle is a "stranger" (background noise), the bouncer says, "You're not on the list," and deletes it immediately.
  • If the particle is a "guest" (part of the signal), even if it's a bit far away or moving strangely, the bouncer says, "You belong here," and keeps it.

How It Works in Plain English

1. The Training: The bouncer was trained on millions of simulated parties. It learned that real guests usually hang out together, move in specific patterns, and share a common history. Strangers look different.
2. The Sorting: When a new collision happens, the bouncer scans the hundreds of particles. It assigns a "score" to each one.
   • High score = "Keep this!" (It's part of the story).
   • Low score = "Delete this!" (It's just noise).
3. The Result: Instead of saving 200 particles per event, the system now saves only the top 10 or so that matter.

Why This is a Big Deal

• Saves Space: By deleting the trash, they reduced the size of the data files by 45%. That's like turning a 100-page novel into a 55-page summary without losing the plot.
• Keeps the Good Stuff: They managed to keep 99% of the actual physics signals. They didn't lose the diamonds; they just threw away the dust.
• Works in Crowds: The old methods got confused when the room was too crowded (high "pile-up"). The new bouncer stays calm and accurate even when the party is packed.

The Future
This isn't just about saving hard drive space. It's about being ready for the future. As the LHC gets even faster in the coming years, this "Smart Bouncer" will be essential. It allows scientists to be more selective, ensuring that when they finally look at the data, they are looking at the most interesting parts of the universe, not a mountain of digital junk.

In short: They taught a computer to be a better editor, cutting out the boring scenes so the movie (the physics) plays perfectly without crashing the system.

Technical Summary

1. Problem Statement

The LHCb experiment during Run 3 (2022–2026) faces a critical data volume challenge driven by a five-fold increase in instantaneous luminosity ($2 \times 10^{33} \text{ cm}^{-2}\text{s}^{-1}$) and a software-only High-Level Trigger (HLT) operating at 30 MHz.

• The Bottleneck: While signal decays (heavy-flavour physics) typically involve only a few charged particles, a single proton-proton collision produces $\sim$200 reconstructed tracks. Charged particle information dominates the event size (approx. 55% of the total), followed by neutral particles (35%) and metadata (10%).
• The Constraint: The data rate must be reduced from $\sim$5.9 GB/s (output of HLT2 for the "Full" stream) to 0.8 GB/s after the "Sprucing" stage (a secondary selection step) to fit storage bandwidth.
• The Challenge: Traditional methods of reducing data volume often rely on aggressive event rejection (lowering the event rate). However, this paper focuses on reducing the event size (information stored per event) without compromising physics reach. The core difficulty is distinguishing relevant signal particles (often from displaced vertices or excited states) from the dense background of pileup and unrelated tracks in a high-multiplicity environment, without introducing biases toward specific decay topologies.

2. Methodology

The authors developed a suite of isolation tools, culminating in a novel Inclusive Multivariate Isolation (IMI) algorithm.

A. Classical Isolation Baselines

Three traditional strategies were reviewed and implemented for Run 3:

1. Track Isolation: Rejects particles based on impact parameter significance ($\chi^2_{IP}$) relative to primary vertices (PV) or requires association with the same PV as the signal.
2. Cone Isolation: Rejects particles within a geometric cone ($\Delta R$) around the signal candidate.
3. Vertex Isolation: Tests compatibility of extra tracks with the signal's secondary vertex (SV) via vertex fitting.
   Limitation: These methods struggle in high-pileup environments, often sacrificing signal efficiency for background rejection or failing to handle complex decay chains (e.g., excited states).

B. Inclusive Multivariate Isolation (IMI)

The IMI algorithm is a machine-learning-based classifier designed to score every "extra" particle in an event relative to a "base" signal candidate.

• Architecture: Uses XGBoost (Extreme Gradient Boosting) for its balance of performance, speed, and interpretability.
• Training Strategy:
  • Inclusive Training: Trained on a broad simulation sample covering diverse $b$-hadron decays ($B^0, B^+, B_s^0, \Lambda_b^0$) with various final states (electrons, muons, taus) and intermediate states (short-lived $D^*$, long-lived $D^0$, excited states).
  • Classes: Binary classification separating Signal (particles from the same decay chain, including prompt decays of excited states) from Background (pileup, other $b$-hadron decays, soft QCD).
• Input Features: The model uses 10 key variables combining geometric, topological, and kinematic information:
  • Impact parameter significances ($\chi^2_{IP}$) relative to PV and SV.
  • Angular separation ($\Delta R$, $\cos \theta$).
  • Vertex displacement metrics (SV shift, signed flight distance).
  • Distance of Closest Approach (DOCA) significance.
  • Transverse momentum ($p_T$).
• Workflow:
  1. Trigger (HLT2): Full events are reconstructed and written to tape.
  2. Sprucing Stage: The IMI algorithm runs offline on the fully reconstructed event. It evaluates all base-extra particle pairs, assigns an IMI score, and only retains particles exceeding a threshold (e.g., score > 0.05).
  3. Output: A relation table linking base particles to retained extra particles, drastically reducing the stored data volume.

3. Key Contributions

1. Novel Algorithm: Introduction of IMI, the first inclusive multivariate isolation tool for LHCb that unifies classical isolation strengths into a single classifier.
2. Data-Size Reduction: A strategy to reduce event size by $\sim$45% while preserving $\sim$99% of signal candidates, directly addressing the storage bandwidth bottleneck.
3. Physics Inclusivity: Unlike previous tools, IMI is designed to handle diverse topologies, including:
   • Excited charm/charmless states (e.g., $D^*$, $\Lambda_c^*$).
   • Prompt particles from excited $b$-hadrons (e.g., $B^*_{s2} \to B^+ K^-$).
   • Complex decay chains with varying multiplicities.
4. Validation: Comprehensive validation using Run 3 data, demonstrating the algorithm's ability to reconstruct resonances (e.g., $D^{*-}$, $\Lambda_c^*$) without channel-specific tuning.

4. Key Results

• Performance Metrics:
  • AUC (Area Under Curve): 0.9964, indicating excellent separation between signal and background.
  • Efficiency vs. Rejection: At the nominal working point (IMI score > 0.05), the algorithm achieves 99% signal candidate efficiency while rejecting >90% of background tracks. This represents a 2–5$\times$ improvement in background rejection over classical methods at the same efficiency.
  • Robustness: Performance remains stable across varying event multiplicities (from low to high pileup) and diverse decay channels ($B^0 \to D^{*-}\mu\nu$, $\Lambda_b \to \Lambda_c^*\mu\nu$, $B_s \to D_s^{(*)}\mu\nu$).
  • Kinematic Bias: The signal efficiency is flat across the $q^2$ (momentum transfer squared) spectrum, crucial for precision measurements of CKM matrix elements.
• Data-Size Impact:
  • Reduces the average event size by 45%.
  • Throughput impact is minimal (<0.1% variation in processing speed), though the overall Sprucing stage sees a ~20% throughput reduction due to the necessary vertex fits for feature extraction. This is acceptable given the lower event rate at the Sprucing stage compared to HLT2.
• Data Validation:
  • IMI rankings in real Run 3 data match simulation trends (signal-like particles cluster at low $\Delta R$ and low $\chi^2_{IP}$).
  • Successfully reconstructed $D^{*-} \to \bar{D}^0 \pi^-$ and $\Lambda_c^*$ resonances using only the highest-ranked particles selected by IMI, confirming its ability to recover non-isolated signal products.

5. Significance and Outlook

• Immediate Impact: IMI is a critical enabler for the LHCb semileptonic physics programme in Run 3, allowing the experiment to meet storage constraints while retaining the full physics potential for precision tests of Lepton Flavour Universality and CKM unitarity.
• Scalability: The lightweight nature of IMI makes it a candidate for future high-luminosity runs (HL-LHC, Run 4/5), where data rates will be even higher.
• Future Directions:
  • Extending isolation to neutral particles (photons, neutral hadrons).
  • Incorporating VELO-only tracks for better background rejection in dense environments.
  • Adopting a multiclass classifier to specifically identify excited heavy-flavour states currently handled by cut-based selections.
  • Using IMI as a "fast pruning" front-end for more computationally intensive reconstruction algorithms (e.g., Graph Neural Networks) in future runs.

In conclusion, the paper demonstrates that Inclusive Multivariate Isolation is a robust, high-performance solution to the data volume crisis in high-luminosity heavy-flavour physics, successfully balancing aggressive data reduction with the preservation of complex physics signals.