Two Flavon Froggatt-Nielsen Models with Genetic Algorithms

This paper presents the first systematic scan of two-flavon Froggatt-Nielsen models using the Non-dominated Sorting Genetic Algorithm III to efficiently identify over 100,000 phenomenologically viable configurations that naturally explain CP violation, fermion masses, and mixing patterns across both quark and lepton sectors.

The Gist

Imagine the universe as a giant, complex recipe book. For decades, physicists have been trying to figure out why the ingredients in this book—specifically the fundamental particles like electrons and quarks—have such wildly different "weights" (masses) and mix together in such specific ways. Some are heavy as lead, others light as feathers, and they dance together in patterns that seem random but are actually governed by hidden rules.

This paper is about cracking the code of those hidden rules using a new kind of digital detective work.

The Mystery: The "Flavor Puzzle"

In the Standard Model of physics, we know that particles have different masses, but we don't know why. It's like knowing a cake has sugar, flour, and eggs, but having no idea why the baker used a cup of sugar instead of a cup of salt.

Physicists use a theory called the Froggatt-Nielsen (FN) mechanism to explain this. Think of it as a "flavor symmetry." Imagine there are invisible "flavon" fields (like invisible spices) that get sprinkled over the particles. The more "spice" a particle gets, the heavier it becomes. The amount of spice is determined by a secret number (a "charge") assigned to each particle.

The Old vs. The New: One Spice Jar vs. Two

For a long time, scientists tried to explain all the particle masses using just one jar of spice (one flavon field).

• The Problem: With only one jar, the recipe is too rigid. It can't explain why some particles mix in a way that creates "CP violation" (a subtle difference between matter and antimatter that is crucial for our existence). It's like trying to bake a perfect cake with only one flavor of extract; you can't get the complex taste you need.

This paper introduces a new approach: What if we use two different jars of spice?

• Jar A (Up-Flavon): Only sprinkles on "up-type" particles (like up quarks).
• Jar B (Down-Flavon): Only sprinkles on "down-type" particles (like down quarks).

The Magic: By having two jars, the scientists discovered a new source of complexity. The "timing" or "phase" between when Jar A and Jar B are sprinkled creates a natural source of CP violation. It's like having two chefs adding spices at slightly different times; the interaction between their rhythms creates a flavor profile that a single chef could never achieve.

The Search: Finding the Perfect Recipe

The problem is that there are billions of possible combinations of spice amounts and particle charges. Trying to check them all by hand is impossible.

The Old Way (Reinforcement Learning):
Previous attempts used AI that had to "train" for a long time, like a student studying for a test before taking it. It would learn a strategy first, then try to find the recipe. This was slow and often got stuck in local loops.

The New Way (Genetic Algorithms):
The authors used a different AI technique called NSGA-III, which works like evolution.

1. The Population: They start with a huge group of random "recipes" (models).
2. The Test: They check each recipe against 17 different rules (like "the electron mass must be right," "the mixing angles must match," etc.).
3. Survival of the Fittest: Instead of just picking the single "best" recipe, they keep a diverse group of recipes that are good at different things. Some are great at getting the masses right, others at getting the mixing angles right.
4. Breeding: They mix and match these good recipes to create new, better ones.
5. No Training: Unlike the old method, this AI doesn't need to "study" first. It starts finding good recipes immediately.

The Results: A Treasure Trove of Recipes

The results were staggering:

• Speed: They found viable models orders of magnitude faster than previous methods.
• Quantity: They discovered over 100,000 unique, valid recipes.
• Diversity: Because they used two spice jars, they found two distinct types of universes:
  • Normal Ordering: Where the heaviest neutrino is the heaviest.
  • Inverted Ordering: Where the lightest neutrino is actually the heaviest of the light ones.
• Simplicity: They found that you don't need complex, messy recipes. Some of the best models only required the "spice" to be sprinkled a tiny number of times (exponents as small as 3), suggesting the underlying physics could be very simple.
• Precision: Even without fine-tuning the numbers, their AI found recipes that predicted particle masses within 6% of the real values.

The Bottom Line

This paper is a proof-of-concept that using two flavor fields instead of one, combined with a smart, evolutionary AI search, can solve a decades-old physics puzzle much faster and more thoroughly than before.

They didn't just find one answer; they found a massive library of over 100,000 possible ways the universe could be built to match what we see today. This suggests the "space" of possible theories is vast and far from exhausted, and that the universe might be built on a surprisingly simple, two-spice framework.

Technical Summary

Technical Summary: Two Flavon Froggatt-Nielsen Models with Genetic Algorithms

Problem Statement
The origin of fermion mass hierarchies and mixing angles remains a central open problem in particle physics. The Froggatt-Nielsen (FN) mechanism addresses this by introducing a $U(1)_{FN}$ flavor symmetry and scalar flavon fields, where hierarchies arise from the suppression of higher-dimensional operators by powers of the flavon vacuum expectation value (vev) to the UV cutoff scale ratio. While extensively studied, systematically exploring the FN model space is computationally challenging due to its high-dimensional, mixed discrete-continuous nature. Specifically, finding integer charge assignments for Standard Model fields that simultaneously reproduce all measured quark and lepton masses, CKM and PMNS mixing angles, and CP-violating phases is a non-linear combinatorial problem. Previous attempts using Reinforcement Learning (RL) were restricted to single-flavon scenarios, employed loose constraints, and required computationally expensive training phases. Furthermore, single-flavon models lack a natural source of CP violation, as flavon-induced phases can typically be removed via field redefinitions.

Methodology
This work presents a systematic scan of a two-flavon FN framework where distinct flavon fields, $\Phi_u$ and $\Phi_d$, couple exclusively to the up-type and down-type sectors, respectively. The authors employ the Non-dominated Sorting Genetic Algorithm III (NSGA-III), a multi-objective optimization technique, to navigate the parameter space.

• Model Setup: The Lagrangian includes separate flavon couplings for up-type quarks, down-type quarks, charged leptons, and neutrinos (via a Type-I see-saw mechanism). The relative phase $\alpha$ between the vevs $\langle \Phi_u \rangle$ and $\langle \Phi_d \rangle$ is retained as a physical parameter, providing a generic source of CP violation absent in single-flavon models.
• Optimization Strategy: Instead of aggregating constraints into a single scalar loss (which risks local minima and scale issues), the authors formulate the search as a multi-objective problem with 17 distinct objectives. These include:
  • CKM and PMNS mixing angles and CP phases (within $3\sigma$ of NuFit and PDG data).
  • Neutrino squared-mass differences (within $3\sigma$).
  • Charged fermion masses (within 20% of central values).
  • A constraint enforcing non-negative flavon exponents.
• Algorithm Implementation: NSGA-III maintains population diversity using reference directions in the objective space, allowing the algorithm to explore the Pareto front of solutions where constraints are satisfied simultaneously. The search space includes 18 discrete FN charges, 45 continuous Wilson coefficients (constrained to order unity), and flavon vev parameters. The algorithm requires no separate training phase, beginning the search for viable models immediately.
• Scan Parameters: The scan explores two regimes: "Up Leading" ($|\epsilon_u| > |\epsilon_d|$) and "Down Leading" ($|\epsilon_d| > |\epsilon_u|$), with vev magnitudes restricted to ensure genuine FN suppression rather than democratic textures.

Key Contributions and Results

1. Discovery of Viable Models: The scan identified over 100,000 unique phenomenologically viable models. The extremely low duplication rate (26 duplicates across independent runs) indicates that the space of valid two-flavon FN realizations has not been exhausted.
2. CP Violation: The study demonstrates that the relative phase between the two flavon vevs provides a natural and generic source of CP violation, a feature impossible in single-flavon models without fine-tuning.
3. Neutrino Mass Ordering: Both Normal Ordering (NO) and Inverted Ordering (IO) solutions were realized.
   • NO Models: Populated the $(m_{\text{lightest}}, m_{ee})$ plane consistently with current NuFit bounds.
   • IO Models: Found primarily in the "Down Leading" scenario, yielding larger effective neutrinoless double beta decay masses ($m_{ee}$) compared to the "Up Leading" scenario.
4. Minimality: The authors identified "minimal" FN realizations where the maximal flavon exponent does not exceed three, suggesting simple UV completions are possible.
5. Mass Reproduction: Without dedicated continuous parameter optimization (e.g., $\chi^2$ minimization), the framework successfully reproduced charged fermion masses with relative deviations as small as 6% (and generally below 10% in dedicated scans).
6. Efficiency: The GA-based approach found thousands of distinct models satisfying all constraints within minutes, significantly outperforming prior RL methods which required hours of training and found far fewer models.

Significance
The paper claims to provide the first systematic and comprehensive scan of two-flavon FN models. Its significance lies in:

• Methodological Advancement: Demonstrating that multi-objective genetic algorithms are superior to reinforcement learning for this specific class of BSM model building, offering immediate viability without training overhead and better handling of high-dimensional, mixed parameter spaces.
• Physical Insight: Establishing that extending the FN framework to two flavons naturally resolves the CP violation origin problem and allows for a rich variety of phenomenological outcomes, including distinct predictions for $m_{ee}$ based on the hierarchy of flavon vevs.
• Exploration of Model Space: By finding over 100,000 viable models with low duplication, the work suggests that the viable parameter space for two-flavon FN models is vast and largely unexplored, providing a robust foundation for future phenomenological studies and precision fits.

The authors note that while the current scan uses broad constraints (e.g., 20% mass tolerance), the framework is capable of tighter constraints, and future work could involve post-processing with dedicated $\chi^2$ fits and renormalization group running for precision comparisons with upcoming experimental data.