Machine Learning insights on the Z3 3HDM with Dark Matter

This paper employs Evolutionary Strategies augmented with Novelty Reward to identify viable Z3-symmetric 3-Higgs Doublet Model dark matter candidates with mass-degenerate WIMPs across two distinct mass regimes, successfully satisfying theoretical and experimental constraints while revealing the challenges of navigating the model's non-convex parameter space.

The Gist

Imagine the universe as a giant, complex puzzle. For decades, physicists have been trying to solve the "Standard Model" puzzle, which explains how particles like electrons and quarks behave. We found a crucial piece in 2012: the Higgs boson (the "God particle"). But the puzzle is still missing a huge chunk: Dark Matter.

We know Dark Matter exists because it holds galaxies together, but we can't see it, touch it, or detect it directly. It's like a ghost that makes up 85% of all the matter in the universe.

This paper is about a new theory trying to solve that missing piece, using a very smart, computer-based detective method. Here is the breakdown in simple terms:

1. The Theory: A "Three-House" Neighborhood

The Standard Model has one "Higgs field" (think of it as one house in a neighborhood). This paper proposes a 3-Higgs Doublet Model (3HDM). Imagine a neighborhood with three houses instead of one.

• House 1 (The Active House): This is the one we know. It's where the 125 GeV Higgs boson lives. It interacts with everything.
• Houses 2 & 3 (The Inert Houses): These are the "quiet neighbors." They don't talk to regular matter much. They are hidden behind a special "Z3 symmetry" fence.
  • Because of this fence, the particles inside these houses (called Dark Matter) can't just disappear or decay. They are stable.
  • The theory predicts two specific "ghosts" living in these houses: H1 and A1. They are twins with the same mass but opposite "spins" (CP quantum numbers).

2. The Problem: Finding the Right Goldilocks Zone

Physicists need to find the exact settings for these three houses so that:

1. The universe is stable (the houses don't collapse).
2. The amount of Dark Matter matches what we see in the sky (the "relic density").
3. The ghosts don't get caught by our detectors (which have been looking for them and finding nothing so far).

It's like trying to bake a cake where the recipe has thousands of variables (sugar, flour, temperature, time). If you get the temperature wrong by a tiny bit, the cake burns. If you get the sugar wrong, it's inedible. We need to find the exact combination that makes a perfect, edible cake that also happens to be invisible to the eye.

3. The Solution: The "Evolutionary Detective" (Machine Learning)

Trying to guess the right numbers by hand is impossible. There are too many combinations. So, the authors used Machine Learning, specifically a method called an Evolutionary Strategy.

Think of this like natural selection for computer code:

• Generation 1: The computer generates 1,000 random "recipes" (sets of numbers for the model).
• The Test: It checks each recipe against the rules (Does it match the universe's mass? Does it break physics? Is it caught by the LZ detector?). Most fail immediately.
• The "Novelty Reward": This is the clever part. Usually, computers get stuck in a rut, finding the same "good" solution over and over. This system gives a "bonus point" (a novelty reward) to recipes that are different from the ones already found. It forces the computer to explore the weird, empty corners of the map where no one has looked before.
• Seeding: Once the computer finds a few good spots, it uses them as "seeds" to grow new, better searches in those specific areas, ensuring it doesn't miss any hidden valleys.

4. The Results: Two Safe Havens

After running this high-tech detective work, they found two "safe zones" where the Dark Matter ghosts could exist without getting caught:

• The Light Zone (50 to 80 GeV): These ghosts are relatively light, about half the weight of a W-boson (a heavy particle).
• The Heavy Zone (380 to 1,000 GeV): These ghosts are very heavy, almost 10 times the weight of a proton.

The Twist:
In previous studies, physicists assumed a very specific, "safe" angle (called $\theta = \pi/4$) to make the math easy. This paper dared to look at all possible angles.

• When they relaxed this rule, they found that the "Dark Matter" could interact much more strongly with the Higgs boson (up to 100 times stronger than before thought!) and still hide from our detectors.
• It turns out the "safe zone" is much bigger and more flexible than we thought.

5. Why This Matters

• The "Neutrino Fog": There is a background noise in our detectors caused by neutrinos (tiny particles from the sun). The authors found that many of their valid solutions hide inside this fog. This means even our next-generation super-sensitive detectors might not be able to rule this model out. The ghosts are hiding in plain sight, disguised as background noise.
• The Method: The paper proves that using AI to "evolve" solutions is a powerful way to explore complex physics models that are too hard for humans to calculate manually.

Summary Analogy

Imagine you are looking for a specific type of rare bird in a massive, foggy forest.

• Old way: You walk in a straight line, checking every bush. You get tired and miss the birds hiding in the dense thickets.
• This paper's way: You release a swarm of drones (the AI). They fly randomly at first. When they find a bird, they don't just stop; they send out new drones to explore the edges of that area and the empty spaces nearby, looking for different types of birds.
• The Discovery: They found that the birds aren't just in one small clearing; they are hiding in two distinct regions of the forest, and they are much better at camouflage (hiding in the "neutrino fog") than we previously believed.

This paper gives us a new map of where to look for Dark Matter and a new, smarter tool to find it.

Technical Summary

Here is a detailed technical summary of the paper "Machine Learning insights on the Z3 3HDM with Dark Matter."

1. Problem Statement

The Standard Model (SM) fails to explain Dark Matter (DM), which constitutes roughly 85% of the universe's matter. While minimal extensions like the Inert Doublet Model (IDM) with a $Z_2$ symmetry have been extensively studied, they face increasingly stringent constraints from direct detection experiments (e.g., LZ) and collider data (LHC).

This paper investigates a Three-Higgs Doublet Model (3HDM) with an imposed $Z_3$ symmetry. In this model:

• One doublet is "active" (acquires a vacuum expectation value, $v$), while two are "inert" (do not acquire a vev).
• The $Z_3$ symmetry ensures the stability of the inert sector, leading to two mass-degenerate DM candidates ($H_1$ and $A_1$) with opposite CP quantum numbers.
• The Challenge: The parameter space of this model is high-dimensional and constrained by a complex interplay of theoretical bounds (boundedness from below, global minimum, unitarity) and experimental data (relic density, direct/indirect detection, Higgs signal strengths).
• Specific Difficulty: Previous studies often fixed the mixing angle $\theta$ between the inert doublets to $\pi/4$ to eliminate a specific gauge annihilation channel ($Z H_i A_i$) that depletes the relic density. Exploring the general case where $\theta$ varies is computationally difficult because the valid parameter space is non-convex, multi-dimensional, and contains disconnected regions that are difficult to find using standard sampling methods.

2. Methodology

The authors employ a Machine Learning (ML)-driven scanning strategy to efficiently navigate the complex parameter space without requiring pre-existing training data.

A. Model Formulation

• Scalar Potential: The authors derive the scalar potential for the $Z_3$ 3HDM, establishing the conditions for the potential to be Bounded From Below (BFB) and ensuring the desired vacuum alignment $(0, 0, v)$ is the global minimum. They identify competing minima $(v, 0, 0)$ and $(0, v, 0)$ and impose constraints to discard points where these are deeper than the inert vacuum.
• Parameters: The scan varies 13 independent parameters, including masses ($m_{H_1}, m_{H_2}, m_{H^\pm}$), mixing angle $\theta$, and couplings ($g_1, g_2$, and quartic couplings $\lambda_i$).

B. Machine Learning Framework

The core of the methodology is an Evolutionary Strategy augmented with Novelty Reward:

1. Optimizer: The Covariant Matrix Adaptation Evolutionary Strategy (CMA-ES) is used to minimize a loss function $L$.
2. Loss Function: $L$ is the sum of constraint violations ($C(O)$) for theoretical bounds, experimental limits, and astrophysical data. $C(O) = 0$ if a point satisfies all constraints; otherwise, it quantifies the deviation.
3. Novelty Reward: To prevent the algorithm from getting stuck in local optima or dense regions, a penalty based on density estimation (Histogram Based Outlier Score - HBOS) is added. Points in sparse regions receive lower penalties, encouraging the algorithm to explore uncharted areas of the parameter space.
4. Prototype Selection (Seed Acquisition): To overcome the "local search" nature of CMA-ES, the authors developed three automated methods to select seeds (starting points) for new runs from previously found valid points:
   • Histogram Gradient Boosting Classifier (HGBC): Identifies globally sparse regions by distinguishing valid data from uniform noise.
   • Mini-Batch k-Means (MBKM): Clusters data and selects points furthest from cluster centroids (edges of clusters) to expand boundaries.
   • Local Outlier Factor (LOF): Detects local density deviations to find outliers on the fringes of local neighborhoods.

C. Constraints

The scan enforces a comprehensive set of constraints:

• Theoretical: BFB conditions, global minimum stability, perturbative unitarity, and oblique parameters ($S, T, U$).
• Collider: Higgs signal strengths (ATLAS/CMS), invisible Higgs branching ratio ($BR < 0.107$), and searches for additional scalars (HiggsBounds).
• Dark Matter:
  • Relic Density: Must match Planck data ($\Omega_{DM}h^2 = 0.1200 \pm 0.0012$) within $3\sigma$. The calculation uses micrOMEGAs with the DarkOmegaN routine to account for co-scattering processes essential for chemical equilibrium.
  • Direct Detection: Spin-independent cross-sections must satisfy LZ 2025 limits.
  • Indirect Detection: Annihilation cross-sections ($\langle \sigma v \rangle$) are checked against Fermi-LAT, AMS-02, and H.E.S.S. bounds (used as a validation step).

3. Key Contributions

1. Rigorous Theoretical Consistency: The authors revisit the scalar potential formulation, explicitly deriving the BFB conditions for the $Z_3$ potential (including neutral, charge-breaking, and $Z_3$-specific terms) and enforcing the global minimum condition, which previous benchmark studies often overlooked.
2. Advanced ML Sampling: The paper introduces and validates a novel prototype selection framework (HGBC, MBKM, LOF) to automate the seeding of evolutionary algorithms. This allows for efficient global exploration of non-convex, disconnected parameter spaces that standard scans miss.
3. Generalization of the Mixing Angle: Unlike previous works that fixed $\theta = \pi/4$, this study explores the full domain of $\theta$ ($-\pi/2 \leq \theta \leq \pi/2$), revealing how deviations from this limit affect the phenomenology and allowed coupling strengths.

4. Results

The study identifies viable Dark Matter candidates in two distinct mass regimes under all constraints:

A. The Conservative Limit ($\theta = \pi/4$)

• Low Mass Region: $50 \text{ GeV} \lesssim m_{DM} \lesssim m_W$.
  • The coupling $g_1$ (DM-Higgs) is tightly constrained to $\mathcal{O}(10^{-5})$ due to direct detection limits.
  • Points in this region often lie within the "neutrino fog," making them difficult to probe even with next-generation detectors.
• High Mass Region: $380 \text{ GeV} \lesssim m_{DM} \lesssim 1000 \text{ GeV}$.
  • The coupling $g_1$ scales up to $\mathcal{O}(10^{-4})$.
  • The intermediate mass region ($m_W < m_{DM} < 380 \text{ GeV}$) is excluded because DM is under-produced due to efficient annihilation into gauge bosons ($W^+W^-$).
• Benchmark Reproduction: The authors successfully reproduced previous benchmark scenarios but found that many were excluded once the full BFB and global minimum conditions were applied.

B. The General Case (Variable $\theta$)

• Significant Expansion of Parameter Space: Allowing $\theta$ to vary opens up new regions.
• Enhanced Couplings: The DM-Higgs coupling $g_1$ can reach values as high as $|g_1| \sim 0.5$ (orders of magnitude larger than in the $\theta = \pi/4$ limit).
  • Mechanism: As $\theta$ deviates from $\pi/4$, the $Z H_i A_i$ vertex (proportional to $\cos 2\theta$) becomes non-zero. This introduces a new depletion channel, requiring larger couplings to other sectors (specifically the Higgs portal) to maintain the correct relic density.
• Mass Regimes: The viable mass ranges remain similar ($50 \text{ GeV} - m_W$and$380 \text{ GeV} - 1 \text{ TeV}$), but the allowed parameter space is significantly broader and more complex.
• Challenges: The general case is computationally intensive due to the additional depletion channel, necessitating the use of the prototype selection methods (LOF and MBKM) to find viable points efficiently.

5. Significance

• Viability of Multi-Component DM: The paper demonstrates that a $Z_3$ 3HDM with two mass-degenerate DM states is a robust candidate for Dark Matter, surviving all current theoretical and experimental constraints.
• Methodological Advancement: The integration of Evolutionary Strategies with Novelty Reward and automated prototype selection provides a powerful template for exploring other complex Beyond Standard Model (BSM) scenarios where the valid parameter space is fragmented or non-convex.
• Phenomenological Implications:
  • The model predicts viable DM candidates that may lie below the "neutrino fog," implying that even next-generation direct detection experiments (like XLZD) might not definitively rule out this model.
  • The finding that large couplings ($g_1 \sim 0.5$) are possible in the general case suggests that if $\theta \neq \pi/4$, the model could be more accessible to collider searches or indirect detection than previously thought.
• Theoretical Rigor: By strictly enforcing global minimum conditions and full BFB constraints, the study corrects potential inconsistencies in previous literature, ensuring that the identified benchmarks are physically realizable.