Modern jet flavour tagging in hadronic Z decays with archived ALEPH data

This paper presents a reanalysis of archived ALEPH LEP data using modern deep learning-based jet flavour tagging techniques, which significantly improve the separation of b-, c-, and s-quark jets compared to legacy algorithms and demonstrate successful data calibration, thereby enabling more precise electroweak measurements and guiding future collider detector development.

The Gist

Imagine you are a detective trying to solve a mystery that happened 30 years ago. The crime scene is a massive particle collider called LEP (Large Electron-Positron collider), which was shut down in 2000. The "suspects" are tiny particles called quarks that were created when electrons and positrons smashed into each other.

The problem? The evidence was stored in a dusty, old archive. The tools the original detectives used back then were like magnifying glasses—they worked okay, but they missed a lot of details.

This paper is about a team of modern detectives who went back into that archive. They brought out super-powered AI tools (specifically, a type of deep learning called a "Transformer," similar to the technology behind advanced chatbots) to re-examine the old evidence.

Here is the story of what they found, explained simply:

1. The Crime Scene: The "Z" Boson Party

When electrons and positrons collide, they sometimes create a heavy particle called a Z boson. This particle is unstable and immediately splits apart into two jets of particles.

• Sometimes it splits into bottom quarks (b).
• Sometimes charm quarks (c).
• Sometimes strange quarks (s).
• And sometimes light quarks (u or d).

The goal of the experiment is to figure out: "Which type of quark did this Z boson turn into?" This is crucial because knowing the answer helps physicists test the laws of the universe.

2. The Old Detective Work vs. The New AI

In the 1990s, the ALEPH experiment (one of the detectors at LEP) tried to tell these quarks apart. They looked for clues like:

• How far a particle traveled before decaying (like a suspect running away from the scene).
• How much energy a particle lost as it moved through gas (like a runner getting tired).

Their methods were decent, but they often confused the suspects. For example, they might mistake a "charm" quark for a "light" quark about 10% of the time.

The New Approach:
The authors took the old data and fed it into a modern AI. Instead of just looking at one clue, the AI looked at the entire party:

• The speed and direction of every single particle.
• The exact path they took (did they leave the main group early?).
• The specific "fingerprint" of energy loss for each particle.
• The presence of specific "guests" like K-mesons (a type of particle common in strange quark jets).

Think of the old method as trying to identify a person by looking at their shoes. The new method is like using facial recognition, voice analysis, and gait analysis all at once.

3. The Big Wins

The results were shocking (in a good way):

• The "B" Quark (Bottom): The new AI is 10 times better at spotting bottom quarks than the old methods, without missing any of the real ones. It's like upgrading from a blurry security camera to a 4K HD camera with night vision.
• The "C" Quark (Charm): Similarly, it's much harder to trick the AI into thinking a charm quark is something else.
• The "S" Quark (Strange): This is the biggest breakthrough. No one had ever successfully built a tool to identify "strange" quarks in this specific type of data before. The AI learned to spot the unique "fingerprint" of strange quarks (which tend to carry more kaons, a specific type of particle). It's like finding a new suspect in the lineup that everyone previously ignored.

4. The "Tag-and-Probe" Calibration

There was one catch: The AI was trained on computer simulations, and sometimes simulations don't perfectly match reality (just like a flight simulator isn't exactly the same as flying a real plane).

To fix this, the team used a clever trick called "Tag-and-Probe."

• Imagine you have a bag of mixed marbles. You pick a few that you are 100% sure are red (the "Tag").
• You use those to calibrate your color sensor.
• Then you check the rest of the bag (the "Probe") with your now-calibrated sensor.

They did this with real data to make sure their AI wasn't "hallucinating" results. After calibration, the AI's predictions matched the real-world data almost perfectly.

5. Why Does This Matter?

You might ask, "Why bother with old data from 1994?"

• Precision Physics: By cleaning up the data, physicists can measure the properties of the universe with much higher precision. It's like re-measuring a room with a laser tape measure instead of a ruler; you might find a tiny error in the blueprint of the universe that we missed before.
• Future Colliders: There are plans to build new, even bigger electron colliders (like the FCC-ee). This paper proves that archiving data is incredibly valuable. If we save our data today, future scientists with even smarter AI can dig it up and find new treasures we can't see yet.
• The "Strange" Discovery: Since they can now identify strange quarks, they can finally measure things that were never measured before, opening up new chapters in the book of physics.

The Bottom Line

This paper is a success story of digital archaeology. The authors didn't just dig up old bones; they used modern technology to make those bones talk, revealing secrets about the universe that were hidden in plain sight for decades. They showed that with the right tools, even "old" data can be as valuable as new data.

Technical Summary

Here is a detailed technical summary of the paper "Modern jet flavour tagging in hadronic Z decays with archived ALEPH data."

1. Problem Statement

The Large Electron–Positron (LEP) collider at CERN produced high-precision electroweak measurements, many of which remain unsurpassed today. However, the analysis of hadronic $Z$ boson decays ($Z \to q\bar{q}$) at LEP was historically limited by the jet flavour tagging algorithms available at the time (mid-1990s).

• Limitations: Legacy algorithms relied primarily on impact parameter measurements or simple neural networks, resulting in significant misidentification rates between heavy flavour jets ($b$, $c$) and light flavour jets ($u, d, s$).
• Untapped Potential: The strange quark ($s$) sector was never effectively tagged at LEP, preventing the measurement of observables like $R_s$ and $A_{FB}^s$.
• Goal: To re-analyze archived ALEPH data using modern deep-learning techniques to significantly improve flavour separation, reduce background contamination, and enable the first-ever strange quark jet tagging at LEP.

2. Methodology

Data and Simulation

• Dataset: Archived $e^-e^+$ collision data from 1994 at the $Z$ pole (91.2 GeV), corresponding to $\approx 58 \text{ pb}^{-1}$ of integrated luminosity ($\approx 9 \times 10^6$ events).
• Simulation: Generated using JETSET 7.4 (tuned to ALEPH data) and passed through the GALEPH detector simulation. The data and simulation were converted to the EDM4HEP format, a standard event model for future $e^-e^+$ colliders (e.g., FCC-ee).
• Selection: Events were selected as hadronic dijets using the Durham $k_T$ algorithm, requiring exactly two jets with $p > 10$ GeV and $|\cos\theta| < 0.65$.

Feature Engineering

The algorithm utilizes a comprehensive set of inputs derived from the ALEPH detector components:

1. Kinematics: Momentum four-vectors, jet mass, and energy.
2. Vertexing:
   • Primary Vertices: Fitted per event to define the interaction point.
   • Secondary Vertices: Reconstructed inclusively and explicitly for $K^0_S \to \pi^+\pi^-$ and $\Lambda^0 \to p\pi^\mp$ decays ($V^0$ candidates).
   • Impact Parameters: Signed 2D and 3D impact parameter significances of tracks relative to the primary vertex.
3. Particle Identification (PID):
   • dE/dx: Energy loss measurements from the Time Projection Chamber (TPC) wire and pad subsystems.
   • Probabilities: Explicit $p$-values for electron, muon, pion, kaon, and proton hypotheses based on the Bethe-Bloch parametrization. This is crucial for distinguishing kaon-enriched strange jets.

Algorithm Architecture

• Model: A ParticleTransformer (ParT) architecture, similar to those used at the LHC (CMS).
• Mechanism: It processes the set of particles in a jet using self-attention mechanisms to learn correlations between constituents without a predefined graph structure.
• Training: Trained on $\approx 1$ million simulated jets, balanced across $b, c, s,$ and light ($u,d$) classes. The output is a set of four normalized probabilities (softmax) for each jet flavour.

3. Key Contributions

1. Multi-Class Flavour Tagger: The first implementation of a simultaneous multi-class tagger for $b, c, s,$ and light quarks using LEP data.
2. Strange Quark Tagging: The first successful implementation and performance study of strange quark jet tagging at LEP, enabling the selection of $Z \to s\bar{s}$ enriched samples.
3. Data Preservation & Format: Demonstrated the viability of converting legacy ALEPH data into the modern EDM4HEP format and applying state-of-the-art machine learning directly to real experimental data (not just simulation).
4. Calibration Strategy: Developed a tag-and-probe calibration method to correct for residual mismodeling between data and simulation, particularly for the $s$-jet score.

4. Results

Performance Improvements

• $b$-jet Tagging: Compared to legacy ALEPH algorithms (impact parameter based and early neural networks), the new ParT tagger achieves a reduction in misidentification rates (background) by up to one order of magnitude for the same $b$-jet efficiency (e.g., 20%).
  • Example: At 20% $b$-efficiency, the $c$-jet mistag rate drops from $\sim 10^{-3}$ (legacy) to $\sim 10^{-4}$ (ParT).
• $s$-jet Tagging:
  • The inclusion of dE/dx information reduces the light ($u,d$) background contamination by 25–45% relative to a baseline without dE/dx.
  • The inclusion of $V^0$ candidates ($K^0_S, \Lambda^0$) provides an additional 3–5% improvement in background rejection.
  • The Area Under the Curve (AUC) for separating $s$-jets from $u,d$-jets is 0.66.

Data vs. Simulation Agreement

• Input Features: Most kinematic and vertexing features show good agreement between data and simulation.
• Discrepancies: Initial comparisons showed mismodeling in the $s$-jet score distribution, attributed to imperfect simulation of neutral particle content and dE/dx resolution.
• Calibration: Applying the tag-and-probe reweighting procedure successfully corrected these discrepancies, resulting in excellent agreement across the full range of classifier scores for $b, c,$ and $s$ channels.

Physics Reach

• Purity: The method allows for the selection of highly pure samples. For example, at 20% signal efficiency, the background contamination for $bb$ events is reduced to $\sim 3.8 \times 10^{-5}$.
• New Observables: The ability to tag $s$-jets opens the door to measuring $R_s$ and $A_{FB}^s$ for the first time at LEP.

5. Significance

• Legacy Physics: This work revitalizes archived LEP data, potentially improving the precision of electroweak observables (like $R_b$ and $A_{FB}^b$) and reducing systematic uncertainties dominated by flavour misidentification.
• Future Colliders: The successful application of ParT to real LEP data serves as a critical proof-of-concept for future $e^-e^-$ colliders (like FCC-ee or CEPC). It validates the use of EDM4HEP formats and advanced ML algorithms for precision physics in clean collision environments.
• Technique Transfer: It demonstrates that modern deep learning techniques, originally developed for the high-pileup, complex environment of the LHC, can be effectively adapted to the clean, low-multiplicity environment of LEP to extract physics that was previously inaccessible.

In conclusion, this paper establishes that modern machine learning techniques can significantly enhance the physics potential of decades-old experimental data, offering a roadmap for both legacy re-analyses and the design of future high-precision collider experiments.