Flavour Anomalies: A comparative analysis using a machine learning algorithm

This paper presents a global fit of flavor anomalies in semileptonic rare $B$-meson decays using a machine learning Monte Carlo algorithm to emulate non-Gaussian likelihood landscapes, demonstrating that a specific scenario involving second-third generation quark mixing and independent four-fermion operator coefficients provides the best agreement with experimental data.

The Gist

Imagine the Standard Model of particle physics as a perfectly tuned orchestra. For decades, every instrument (every particle and force) has played in perfect harmony, and the music (experimental data) has matched the sheet music (theoretical predictions) almost exactly.

However, recently, the musicians in the "B-meson section" of the orchestra started playing a few notes that sounded slightly off-key. These are the "Flavour Anomalies."

This paper is like a team of detective musicologists trying to figure out: Is the sheet music wrong? Is the orchestra playing a new, hidden song? Or is there a new instrument we haven't noticed yet?

Here is a breakdown of their investigation, using simple analogies.

1. The Mystery: Two Different Kinds of "Off-Notes"

The scientists noticed two specific types of weirdness in the music:

• The Heavy Hitters (Tau Leptons): In certain decays, heavy particles (called "tau leptons") were appearing more often than the sheet music predicted. It's like if a conductor asked for a soft violin note, but the violinist kept playing a loud trumpet blast. This happens in the $R_{D^{(*)}}$ and $R_{J/\psi}$ ratios.
• The Missing Neutrinos: In another type of decay, the orchestra seemed to be producing too many "invisible" particles (neutrinos) in a specific channel ($B \to K \nu \bar{\nu}$), while the other channels looked normal.

2. The Old Theory vs. The New Clue

Previously, the detectives thought the problem was caused by a single "magic knob" that turned up the volume for all these weird notes equally. They assumed that if you fixed the "tau" problem, you automatically fixed the "neutrino" problem.

But the plot thickened:

• New measurements showed that some other "off-notes" (involving electrons and muons) have actually gone back to being perfect.
• Meanwhile, the "neutrino" problem got worse.

This meant the old "single magic knob" theory was broken. You can't use one knob to fix a trumpet blast and a missing drumbeat at the same time. You need two different knobs.

3. The Solution: Scenario III (The "Independent Knobs")

The authors proposed a new theory called Scenario III.

• The Old Way: Imagine a remote control with one button that controls both the volume and the bass. If you press it, both change together. This didn't work anymore.
• The New Way (Scenario III): They realized they need a remote with two independent buttons.
  • Button 1 (Singlet): Controls the "invisible neutrino" channel.
  • Button 2 (Triplet): Controls the "heavy tau" channel.
  • The Twist: They also realized that the "music" only gets weird when the second and third generations of particles mix together, but the leptons (the heavy particles) stay pure and don't mix with the lighter ones.

This new setup fits the data perfectly. It explains why the heavy taus are loud, why the neutrinos are missing, and why the light electrons are still playing perfectly in tune.

4. The Secret Weapon: The Machine Learning "Simulator"

Usually, to test a theory like this, physicists have to run millions of complex calculations, like trying to find a needle in a haystack by checking every single piece of hay one by one. This takes forever and is prone to errors.

Instead, this team used a Machine Learning (ML) algorithm. Think of this as a super-smart simulator:

• The Training: They fed the computer a small sample of the "haystack" (about 10,000 data points) and taught it what the "needle" (the best fit) looks like.
• The Emulator: Once trained, the computer didn't just guess; it learned the shape of the haystack. It could instantly generate millions of new scenarios and tell the scientists, "If you turn the knobs this way, the music sounds like this."
• The Benefit: It was fast, accurate, and could handle the "weird shapes" of the data that normal math struggles with. It's like having a GPS that doesn't just show you the road, but predicts traffic jams and shortcuts before you even leave the driveway.

5. The Verdict

The team concluded that:

1. New Physics is likely real: There is definitely something beyond our current "sheet music" (Standard Model) happening.
2. It's specific: This new physics seems to interact mostly with the third generation of particles (the heavy ones) and mixes the second and third generations of quarks, but leaves the lighter particles alone.
3. The "Neutrino" Puzzle: While they fixed the "tau" problem, the "neutrino" problem in the neutral channel ($B^0 \to K^{*0}$) is still a bit tricky. It suggests that while their new theory is the best fit so far, there might be even more complexity (like "right-handed" currents) that they haven't fully unlocked yet.

The Big Picture

This paper is a victory for flexibility. Instead of forcing the data to fit an old, rigid theory, the scientists used a flexible, modern approach (Machine Learning) to let the data tell them the story.

They found that the universe isn't playing a simple song with one instrument; it's playing a complex jazz improvisation where the heavy notes and the invisible notes need different rules to make sense. And thanks to their new "simulator," they can finally hear the melody clearly.

Technical Summary

Here is a detailed technical summary of the paper "Flavour Anomalies: A comparative analysis using a machine learning algorithm" by Peñaranda, Alda, and Mira.

1. Problem Statement

The paper addresses the persistent discrepancies between Standard Model (SM) predictions and experimental measurements in flavour physics, specifically within B-meson decays.

• Historical Context: Previous analyses focused on anomalies in $b \to s \ell^+ \ell^-$ transitions (e.g., $R_K$ and $R_{K^*}$ ratios) and charged-current anomalies ($R_{D^{(*)}}$).
• New Experimental Landscape: Recent data (2022–2024) has fundamentally shifted the landscape:
  • The $R_K$ and $R_{K^*}$ anomalies have resolved, with measurements now consistent with the SM.
  • A new, significant excess has emerged in the rare decay $B^+ \to K^+ \nu \bar{\nu}$ (measured by Belle II).
  • Tensions remain in $R_{D^{(*)}}$ (semileptonic $b \to c \tau \nu$) and $R_{J/\psi}$.
• The Challenge: Previous Effective Field Theory (EFT) analyses assumed a specific correlation between Wilson coefficients (specifically $C_1 = C_3$) driven by the $R_K$ anomalies. With $R_K$ resolved and $B \to K \nu \bar{\nu}$ showing an excess, the authors argue that the previous operator structure is insufficient. They need a framework that can simultaneously explain the $b \to c \tau \nu$ anomalies and the $b \to s \nu \bar{\nu}$ excess while respecting the SM-like behavior of $b \to s \ell^+ \ell^-$.
• Computational Challenge: The likelihood landscape for these global fits is highly non-Gaussian, featuring curved ridges, approximate degeneracies, and mixed parameter scales, making standard sampling methods (like MCMC) inefficient for high-resolution confidence region mapping.

2. Methodology

A. Theoretical Framework: SMEFT and WET

The authors employ the Standard Model Effective Field Theory (SMEFT) matched to the Weak Effective Theory (WET) at the electroweak scale.

• Operators: They focus on dimension-6 operators involving left-handed lepton and quark doublets ($\ell q$). The Lagrangian includes singlet ($C_1$) and triplet ($C_3$) operators.
• Flavor Structure: The flavor structure is determined by rotation matrices ($\lambda$) acting on the generation indices. The authors parameterize these matrices using complex numbers $\alpha$ and $\beta$.
• Key Assumption: The New Physics (NP) affects only one generation in the interaction basis, requiring rotation to the mass basis.

B. Three Comparative Scenarios

The authors test three distinct hypotheses to determine the best fit:

1. Scenario I: Mixing only between 2nd and 3rd generations ($\alpha=0$), with $C_1 = C_3$.
2. Scenario II: Mixing with 1st and 2nd generations allowed, but still assuming $C_1 = C_3$.
3. Scenario III (Proposed): $C_1$ and $C_3$ are independent parameters. No mixing in the lepton sector ($\alpha_\ell = \beta_\ell = 0$), but mixing between 2nd and 3rd quark generations is allowed.

C. Machine Learning (ML) Approach

Instead of relying solely on traditional Markov Chain Monte Carlo (MCMC), the authors utilize a Machine Learning Monte Carlo framework:

• Algorithm: XGBoost (eXtreme Gradient Boosting), a tree-based regression algorithm.
• Rationale: Tree-based methods are superior for non-Gaussian likelihood landscapes with narrow ridges and sharp transitions. They do not assume global smoothness (unlike Neural Networks) and are computationally efficient.
• Training:
  • Generated a training set of 10,500 parameter points.
  • Trained the model to approximate the $\Delta \chi^2$ (log-likelihood) function.
  • Validated with a Pearson correlation coefficient of $r=0.96$ between predicted and actual $\Delta \chi^2$.
• Analysis Tools:
  • SHAP (Shapley Additive exPlanations): Used to decompose the model output and quantify the relative importance of each parameter ($C_1, C_3, \beta_q$) on the likelihood, ensuring the fit is driven by physical observables rather than numerical noise.
  • High-Resolution Sampling: The trained emulator generates millions of sample points instantly, allowing for detailed correlation studies and high-resolution confidence contours that would be computationally prohibitive with exact likelihood evaluations.

3. Key Contributions

1. Introduction of Scenario III: The paper demonstrates that the assumption $C_1 = C_3$ (valid when $R_K$ anomalies were present) is no longer viable. The new experimental data requires independent Wilson coefficients for singlet and triplet operators to simultaneously fit the $b \to c \tau \nu$ and $b \to s \nu \bar{\nu}$ anomalies.
2. Flavor Structure Identification: The best-fit scenario (III) implies:
   • No mixing in the lepton sector (NP affects only the 3rd generation, $\tau$).
   • Significant mixing in the quark sector between the 2nd and 3rd generations ($\beta_q \approx 0.64$).
   • This structure naturally suppresses contributions to $b \to s \ell^+ \ell^-$ (keeping $R_K$ SM-like) while enhancing $b \to s \nu \bar{\nu}$ and $b \to c \tau \nu$.
3. ML Validation for EFT: The paper provides a robust case study for using XGBoost in particle physics global fits. It validates the ML emulator against exact likelihood calculations, showing excellent agreement in confidence regions and demonstrating that ML can handle complex, non-Gaussian topologies more efficiently than standard grid sampling or MCMC.

4. Results

• Statistical Fit:
  • Scenario III provides the best fit to the global data with a **pull of $6.25\sigma$** ($\Delta \chi^2_{SM} = 46.66$).
  • Scenarios I and II (assuming $C_1=C_3$) yield significantly worse fits (pulls of $5.71\sigma$and$5.51\sigma$ respectively).
• Parameter Values (Scenario III):
  • $C_1 = -0.205 \pm 0.015$
  • $C_3 = -0.12^{+0.02}_{-0.01}$
  • $\beta_q = 0.64^{+1.36}_{-0.24}$
  • $\alpha_\ell = \beta_\ell = 0$ (No lepton mixing).
• Observable Predictions:
  • $R_{D^{(*)}}$ and $R_{J/\psi}$: The fit successfully reproduces the experimental excesses (e.g., $R_{D^*} \approx 0.288$, consistent with data).
  • $B^+ \to K^+ \nu \bar{\nu}$: The branching ratio prediction improves significantly, moving closer to the Belle II measurement ($\approx 2.3 \times 10^{-5}$).
  • $B^0 \to K^{*0} \nu \bar{\nu}$: The fit worsens the tension for the neutral mode compared to the SM, suggesting that the minimal left-handed operator structure of Scenario III cannot fully reconcile the charged and neutral $B \to K^{(*)} \nu \bar{\nu}$ channels simultaneously.
  • $R_K$ Ratios: The fit maintains consistency with the SM ($R_K \approx 1$), as required by the new data.
• Correlations: A critical finding is the loss of correlation between $R_{D^*}$ and $BR(B^+ \to K^+ \nu \bar{\nu})$ in Scenario III. In previous models (Scenario II), these were tightly correlated; in Scenario III, they are decoupled, allowing independent adjustment to fit the new data.

5. Significance

• Phenomenological Shift: The paper highlights a paradigm shift in B-physics. The resolution of $R_K$ anomalies and the emergence of $B \to K \nu \bar{\nu}$ excesses necessitate a move away from unified $C_1=C_3$ models toward models with independent singlet and triplet couplings and flavor-specific mixing (U(2) symmetry-like structures).
• Methodological Advancement: The successful application of XGBoost-based ML emulators demonstrates a powerful new tool for precision phenomenology. It overcomes the computational bottlenecks of global fits with complex likelihood landscapes, enabling high-resolution confidence regions and robust sensitivity analyses (via SHAP) that were previously infeasible.
• Future Directions: The results point toward New Physics interacting primarily with the third generation of leptons and mixing the second and third quark generations. The paper suggests that future measurements of $R_{J/\psi}$ precision and polarisation observables are crucial to further constrain these models and potentially resolve the remaining tension in the neutral $B \to K^* \nu \bar{\nu}$ channel, which may require right-handed currents or more complex flavor structures.