Inclusive Flavour Tagging at LHCb

✨

This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a mystery at a massive, chaotic train station (the Large Hadron Collider). Two special trains, made of invisible "B-mesons," have just arrived. One train is the "good" version (a particle), and the other is the "evil" twin (an antiparticle). To solve the mystery of how these trains behave and change over time, you need to know exactly which one arrived first.

The problem? The station is so crowded with thousands of other people (other particles) that it's incredibly hard to tell which train is which just by looking at the immediate surroundings.

The Old Way: The "Specialist Detectives"

For years, the LHCb experiment used two types of "detectives" to solve this:

The Opposite-Side Detective: This detective looks at the other train that arrived at the same time. If that train left a specific clue, they guess what the first train was.
The Same-Side Detective: This detective looks at the people walking right next to the train as it arrived. If they see a specific type of person, they make a guess.

The Flaw: These detectives are very picky. They only look at specific people or specific clues. If the clue is hidden, or if they pick the wrong person to follow, they give up or make a mistake. They ignore the rest of the crowd, even though that crowd might hold the answer.

The New Way: The "Super-Intelligent Crowd Analyst" (DeepSets)

The paper introduces a new, revolutionary detective: The Inclusive Flavour Tagger (IFT).

Instead of picking just one or two clues, this new detective uses a super-smart AI (a Deep Neural Network called DeepSets) to look at everyone in the station at once.

The Metaphor: Imagine the old detectives were like people trying to find a friend in a crowd by only looking at the person's hat or shoes. The new AI is like a super-observer who looks at the hat, shoes, gait, voice, who they are talking to, and the general vibe of the entire crowd simultaneously.
How it works: The AI doesn't care how many people are in the crowd (some events have 10 tracks, others have 100). It processes every single person (track) individually, then combines all that information into one giant "gut feeling" to decide: "Is this the good train or the evil twin?"

The Results: A Massive Upgrade

The paper tested this new AI against the old detectives using real data from 2016–2018. The results were like upgrading from a flip phone to a supercomputer:

For the B0 mesons: The new AI improved the ability to correctly identify the train by 35%.
For the B0s mesons: It improved by 20%.

Why does this matter?
In physics, being more accurate means you need less data to prove a theory. Think of it like taking a photo in the dark. The old detectives were like a camera with a slow shutter speed; you had to stand still for a long time (collect years of data) to get a clear picture. The new AI is like a camera with a super-fast shutter; it captures a crystal-clear image instantly.

This means scientists can now measure CP violation (a fundamental difference between matter and antimatter) with much higher precision. This helps us understand why the universe is made of matter and not just empty space.

The "Portability" Check

A smart reader might ask: "What if the AI was trained on one type of crowd but has to solve a mystery in a totally different type of crowd?"

The scientists tested this. They trained the AI on one type of decay and asked it to solve a different one. They found that the AI didn't get confused. It performed just as well as the old specialists, proving it's a robust tool that can be used for many different physics experiments without needing to be retrained from scratch.

The Future

The paper concludes that this "inclusive" approach (looking at everything) is the future. While the current AI is already working great for the LHCb experiment's recent data, the team is already working on an even smarter version (using "Transformers," a newer type of AI) to handle the even more crowded and chaotic environment of the future High-Luminosity LHC.

In short: They replaced a picky, limited detective with a super-observant AI that looks at the whole picture, resulting in a 20–35% boost in accuracy for some of the most important measurements in particle physics.

1. Problem Statement

Precise measurements of time-dependent $B^0_{(s)}$ – $\bar{B}^0_{(s)}$ oscillations and CP-violating observables at the LHCb experiment rely heavily on knowing the production flavour of neutral B mesons (i.e., whether the meson was produced as a $B^0$ or $\bar{B}^0$ ).

Current Limitations: The existing LHCb tagging strategy employs two complementary methods:
- Opposite-Side (OS) Tagger: Uses decay products of the other $b$ -quark from the $b\bar{b}$ pair.
- Same-Side (SS) Tagger: Uses particles produced in association with the signal $B$ meson during hadronization.
- The Bottleneck: These methods rely on predefined, specialized subsets of tracks. This selection process discards valuable correlations present in the full event and introduces ambiguities in track selection.
Future Challenges: With the LHCb Upgrade I (purely software trigger) and future High-Luminosity LHC (HL-LHC) plans, instantaneous luminosity will increase significantly. The current reliance on strict track selection will become a limiting factor, necessitating a more inclusive approach that utilizes all available information to mitigate efficiency losses and maintain precision.

2. Methodology: The Inclusive Flavour Tagger (IFT)

The paper introduces a novel Inclusive Flavour Tagger (IFT) based on a Deep Neural Network (DNN) architecture known as DeepSets.

Core Concept: Unlike classical taggers that select specific tracks, the IFT utilizes all reconstructed tracks in the event. This preserves correlations that are otherwise lost.
DeepSets Architecture: To handle the variable number of tracks per event (a challenge for standard DNNs), the model uses two sub-networks:
1. $\phi$ (Feature Extraction): Processes per-track feature vectors independently.
2. Aggregation: Combines the resulting latent vectors into a single permutation-invariant representation (order of tracks does not matter).
3. $\rho$ (Decision): A standard DNN (input layer, 3 hidden layers, output layer) that processes the aggregated vector to produce the tagging output.
Training Configuration:
- Architecture: Uses Rectified Linear Unit (ReLU) activations, except for a final sigmoid output ensuring a probabilistic prediction $y \in [0, 1]$ .
- Loss Function: Binary Cross-Entropy.
- Parameters: The model contains 3,980 trainable parameters.
- Separate Models: Distinct classifiers are trained for $B^0$ and $B^0_s$ (due to different fragmentation of $d$ and $s$ quarks) and for specific data-taking periods (2016–2018) to account for trigger configuration changes.
Input Features:
- Low-level: All reconstructed track kinematics.
- High-level: Outputs from a pre-trained multiclass Boosted Decision Tree (BDT) that categorizes tracks (SS fragmentation, OS fragmentation, OS decay, or unrelated).
- Optional: Predictions from classical SS and OS taggers.
- Note: The IFT outperforms classical taggers even with low-level inputs alone, but including high-level features yields further gains.
Calibration Strategy:
- Training: Simulated samples ( $B^0 \to J/\psi K^{*0}$ and $B^0_s \to J/\psi \phi$ ).
- Calibration: Background-subtracted real data from flavour-specific decays ( $B^0 \to J/\psi K^+\pi^-$ and $B^0_s \to D^-_s \pi^+$ ).
- Linear Fit: The mistag probability $\eta$ is calibrated using a linear fit to correct for residual differences between simulation and data: $\omega(\eta) = p_0 \pm p_1 \cdot (\eta - \bar{\eta})$ .

3. Key Contributions

Inclusive Approach: The first application of a DeepSets architecture to flavour tagging in high-energy physics, moving from selective track analysis to full-event inclusivity.
Performance Metric: The paper defines performance via Effective Tagging Power ( $\varepsilon_{eff} = \varepsilon_{tag} \langle D^2 \rangle$ $ε_{e f f} = ε_{t a g} ⟨ D^{2} ⟩$ ), where $\varepsilon_{tag}$ $ε_{t a g}$ is efficiency and $\langle D^2 \rangle$ $⟨ D^{2} ⟩$ is average squared dilution.
- The IFT achieves much higher $\varepsilon_{tag}$ because it only excludes events with a mistag probability of exactly $\eta = 0.5$ , whereas classical taggers must discard events where the "correct" track cannot be identified.
Portability Validation: The model was tested on "golden modes" not used in training ( $B^0 \to J/\psi K_S$ and $B^0_s \to J/\psi K^+K^-$ ) to ensure the algorithm generalizes across different kinematic and multiplicity distributions.

4. Results

The IFT was evaluated on LHCb Run 2 data (13 TeV) across three years (2016–2018).

For $B^0$ Mesons:

Calibration Channel ( $B^0 \to J/\psi K^+\pi^-$ ):
- The IFT achieved an effective tagging power of ~5.6–5.9%, compared to ~4.0–4.3% for the combined classical SS+OS taggers.
- Gain: A consistent improvement of ~35% in tagging power across all datasets.
- Efficiency ( $\varepsilon_{tag}$ ) increased from ~87% to ~95%.
Validation Channel ( $B^0 \to J/\psi K_S$ ):
- The gain was even more pronounced (~65%), attributed to the classical taggers' suboptimal selection efficiency in this specific channel (classical $\varepsilon_{tag} \approx 69\%$ vs. IFT $\varepsilon_{tag} \approx 95\%$ ).

For $B^0_s$ Mesons:

Calibration Channel ( $B^0_s \to D^-_s \pi^+$ ):
- Effective tagging power increased from ~6.0–6.5% (classical) to ~7.2–7.8% (IFT).
- Gain: A consistent improvement of ~20%.
Validation Channel ( $B^0_s \to J/\psi K^+K^-$ ):
- Similar ~20% improvement observed.
- Impact on Physics: A blinded fit to this channel demonstrated a ~10% reduction in statistical uncertainty for the CP-violating phase $\phi_s$ , consistent with the square-root dependence of uncertainty on tagging power.

Systematic Uncertainties:

Studies confirmed that porting the IFT to decay modes with different kinematics/multiplicities does not introduce systematic uncertainties larger than those of current taggers. The differences in calibration parameters were found to be subdominant to statistical uncertainties.

5. Significance and Outlook

Immediate Impact: The IFT provides a significant boost in statistical precision for time-dependent CP violation measurements without requiring additional data. It is already integrated into the LHCb framework for ongoing Run 2 analyses.
Future Readiness: As LHCb moves toward Run 3 and the HL-LHC era, where pile-up and trigger rates will be extreme, the inclusive nature of the IFT is crucial. It mitigates the limitations of track selection that will become more severe at higher luminosities.
Technological Evolution: The success of DeepSets paves the way for even more advanced architectures (e.g., Transformers) to be explored for future inclusive tagging in the more challenging environments of Run 3 and beyond.

In conclusion, the Inclusive Flavour Tagger represents a paradigm shift in flavour tagging at LHCb, leveraging deep learning to utilize the full event topology, resulting in substantial gains in tagging power (35% for $B^0$ , 20% for $B^0_s$ ) and enabling more precise tests of the Standard Model.