Probing Freeze-In Dark Matter via a Spin-2 Portal at the LHC with Vector Boson Fusion and Machine Learning

This paper investigates the collider phenomenology of feebly interacting dark matter produced via a spin-2 portal, demonstrating that machine learning-enhanced searches at the High-Luminosity LHC can effectively probe cosmologically viable freeze-in parameter space through Vector Boson Fusion channels.

The Gist

The Big Mystery: The Invisible Ghost

Imagine the universe is a giant, bustling city. We can see the buildings, the cars, and the people (this is Normal Matter). But astronomers know that 85% of the city is actually made of invisible "ghosts" (this is Dark Matter). These ghosts hold the city together with gravity, but they don't shine, they don't reflect light, and they barely bump into anything.

For decades, scientists have been trying to catch these ghosts by building giant traps (colliders) to see if they bump into normal people. They've been looking for "Weakly Interacting Massive Particles" (WIMPs)—ghosts that are heavy and bump into things occasionally. But so far, the traps are empty. The ghosts are too shy to be caught.

The New Theory: The "Feeble" Ghost and the Gravity Bridge

This paper suggests a new idea: Maybe the ghosts aren't just shy; maybe they are feeble. They interact so weakly with us that they are practically invisible. This is called Feebly Interacting Dark Matter (FIDM).

But how do we catch something that barely interacts? The authors propose a new "bridge" to catch them: a Spin-2 Portal.

• The Analogy: Imagine the Standard Model (us) and the Dark Sector (the ghosts) are two separate islands. Usually, we try to build a bridge using heavy trucks (gluons/quarks). But in this new theory, the bridge is a massive, invisible trampoline (a Spin-2 particle, like a heavy version of a graviton).
• The Catch: This trampoline doesn't care about the heavy trucks; it only cares about light (photons). It connects the two islands by bouncing light back and forth. Because the connection is so weak, the ghosts are produced very slowly in the early universe, never reaching a "thermal equilibrium" (they never get hot enough to mix with us). This is the "Freeze-In" mechanism: they slowly "freeze in" to existence over time, rather than freezing out like ice cubes.

The Challenge: Finding a Needle in a Haystack

The problem is that because these ghosts interact so weakly, creating them in a particle collider (like the Large Hadron Collider, or LHC) is incredibly rare. It's like trying to find a specific, invisible grain of sand in a hurricane.

Furthermore, the usual way scientists look for new particles is by smashing two heavy trucks together and seeing what flies out. But in this model, the "trucks" (quarks) don't talk to the bridge. Instead, the bridge is built by two beams of light (photons) crashing together. This is a much rarer event, making the signal even harder to find.

The Solution: Vector Boson Fusion and the "AI Detective"

The authors realized that to find these rare events, they needed a specific strategy and a super-smart assistant.

1. The Strategy (Vector Boson Fusion): Instead of looking for a direct crash, they look for a specific "fingerprint" left behind. When the light beams crash to create the dark matter, they kick two jets of normal particles (quarks) forward, like two skiers jumping off a ramp.

   • The Metaphor: Imagine you are trying to find a ghost in a room. You can't see the ghost, but you know that when the ghost appears, it kicks two chairs across the room. You don't look for the ghost; you look for the two chairs flying apart and the empty space (missing energy) where the ghost went.

2. The AI Detective (Machine Learning): The "chairs flying apart" (the jets) create a messy signal mixed with millions of other background events (noise). Traditional methods use simple "cut-and-paste" rules (e.g., "if the chair moves faster than X, keep it"). But this is too blunt.

   • The authors trained a Machine Learning algorithm (specifically a Gradient Boosted Decision Tree, or BDT). Think of this AI as a super-detective. Instead of just looking at one clue, it looks at everything at once: the angle of the chairs, how fast they are moving, the shape of the room, and the timing. It learns the subtle, complex patterns that distinguish a "ghost event" from a "background noise event."

The Results: A New Hope for the Future

The team simulated what would happen if they ran this experiment at the High-Luminosity LHC (the upgraded, super-powerful version of the collider coming in the future).

• The Good News: Even though the signal is tiny, the combination of the "two-chair" strategy and the "AI detective" allows them to see through the noise. They found that the LHC could potentially probe a huge chunk of the "Ghost Universe" that was previously thought to be invisible.
• The Limit: They can't find the ghosts if the "bridge" (the mediator) is too heavy or the connection is too weak (which happens if the universe reheated to very high temperatures early on). But for a wide range of realistic scenarios, the search is viable.

The Takeaway

This paper is a roadmap for how to catch the "feeblest" ghosts in the universe. It tells us:

1. Stop looking for heavy trucks; look for light beams.
2. Don't just use simple rules; use AI to find the subtle patterns in the chaos.
3. The future of Dark Matter hunting isn't just about building bigger machines; it's about building smarter analysis tools.

By connecting the history of the early universe (Cosmology) with the experiments of today (Colliders), this study bridges the gap between how the universe was born and how we can test it today, using a massive trampoline and a super-smart AI to finally catch the invisible.

Technical Summary

1. Problem Statement

The search for Dark Matter (DM) has historically focused on Weakly Interacting Massive Particles (WIMPs), but the lack of signals in direct detection and collider experiments has shifted interest toward Feebly Interacting Dark Matter (FIDM). In FIDM scenarios, particles interact so weakly with the Standard Model (SM) that they never reach thermal equilibrium in the early Universe, instead being produced via the freeze-in mechanism.

The specific challenges addressed in this work are:

• Theoretical Gap: Connecting early-Universe cosmology (freeze-in production) with collider phenomenology for models mediated by a spin-2 particle (massive graviton-like mediator).
• Experimental Limitations: Traditional LHC searches for spin-2 resonances (e.g., Randall-Sundrum gravitons) rely on gluon-fusion production and diphoton decay channels. However, in the specific "photon-only portal" model studied here, production is dominated by photon-photon fusion (which has a much lower cross-section), and the mediator may decay invisibly into DM. Consequently, standard inclusive diphoton resonance searches have negligible sensitivity to these feebly coupled scenarios.
• Signal Discrimination: The signal involves Vector Boson Fusion (VBF) topology with missing transverse momentum ($E_T^{miss}$), which is buried under large SM backgrounds (e.g., $Z(\to \nu\bar{\nu}) + \text{jets}$). Distinguishing this subtle signal requires advanced analysis techniques beyond simple cut-based strategies.

2. Methodology

A. Theoretical Framework

• Model: The authors propose a minimal extension of the SM where a massive spin-2 mediator ($G_{\mu\nu}$) couples universally to the energy-momentum tensors of both the SM (specifically photons in the benchmark) and a scalar DM field ($\chi$).
• Lagrangian: The interaction is governed by effective operators suppressed by high mass scales $\Lambda_\gamma$ (coupling to photons) and $\Lambda_\chi$ (coupling to DM).
  $$\mathcal{L}_{int} = \frac{1}{\Lambda_\gamma} G_{\mu\nu} T^{(\gamma)\mu\nu} + \frac{1}{\Lambda_\chi} G_{\mu\nu} T^{(\chi)\mu\nu}$$
• Cosmology: The DM relic abundance is generated via freeze-in through the process $\gamma\gamma \to G^* \to \chi\chi$. The authors calculate the relic density for both off-resonant (production far from the mediator mass pole) and resonant regimes, imposing constraints on the reheating temperature ($T_R$) and coupling scales to match the observed $\Omega_{DM}h^2 \approx 0.12$.

B. Collider Simulation and Event Generation

• Process: The signal process is $pp \to G + jj \to \chi\chi + jj$ (VBF topology with two forward jets and missing energy).
• Tools:
  • FeynRules/MadGraph5_aMC@NLO: Used for matrix element generation at Leading Order (LO) with NNPDF2.3 PDFs at $\sqrt{s} = 13.6$ TeV.
  • Pythia 8: For parton showering and hadronization.
  • Delphes 3: For fast detector simulation (CMS configuration), including pileup effects (average 140 interactions) and object reconstruction efficiencies.
• Backgrounds: Dominant backgrounds include $Z(\to \nu\bar{\nu}) + \text{jets}$, $W(\to \ell\nu) + \text{jets}$, diboson production ($VV$), and top quark production.

C. Machine Learning Strategy

To overcome the low signal-to-background ratio, the authors employ a Gradient-Boosted Decision Tree (BDT) classifier (using XGBoost).

• Input Features: 8 kinematic variables were selected to maximize discrimination:
  1. Missing transverse momentum ($E_T^{miss}$).
  2. Transverse momenta of leading ($p_T(j_1)$) and subleading ($p_T(j_2)$) jets.
  3. Dijet invariant mass ($m_{jj}$).
  4. Pseudorapidities of jets ($\eta(j_1), \eta(j_2)$).
  5. Angular separations ($\Delta\eta_{jj}, \Delta\phi_{jj}$).
• Training: The dataset (approx. $10^6$ signal and $10^7$ background events) was split 90/10 for training/testing. The BDT learns complex, non-linear correlations between these variables that simple rectangular cuts cannot capture.

D. Statistical Analysis

• Sensitivity Projection: A shape-based analysis using the full BDT output distribution was performed with the Combine framework.
• Uncertainties: Systematic uncertainties (PDFs, scale variations, luminosity, jet energy scale, and a conservative 10% background normalization uncertainty) were included as nuisance parameters.
• Scenario: Projections assume the High-Luminosity LHC (HL-LHC) with an integrated luminosity of $L = 3000 \text{ fb}^{-1}$.

3. Key Contributions

1. Novel Search Strategy: The paper establishes that Vector Boson Fusion (VBF) combined with Machine Learning is the optimal strategy to probe spin-2 mediated freeze-in DM, a regime where traditional resonance searches fail.
2. Cosmology-Collider Bridge: It systematically maps the parameter space ($m_G, m_\chi, \Lambda_\gamma, \Lambda_\chi$) consistent with freeze-in cosmology to collider exclusion limits, identifying which cosmological scenarios are testable at the HL-LHC.
3. Demonstration of ML Efficacy: The study quantifies how BDTs significantly enhance sensitivity compared to cut-based methods, allowing the exploration of parameter spaces with production cross-sections as low as $\mathcal{O}(10^{-3})$ pb.
4. Reinterpretation of Limits: The authors demonstrate that existing diphoton resonance limits from ATLAS and CMS are not applicable to this specific "photon-only" portal due to the different production mechanism (photon fusion vs. gluon fusion) and invisible decay modes.

4. Results

• Kinematic Discrimination: The signal exhibits a distinct VBF topology characterized by:
  • Large dijet invariant mass ($m_{jj}$) and large rapidity separation ($\Delta\eta_{jj}$).
  • Harder $p_T$ spectra for the leading jet compared to SM backgrounds.
  • A harder $E_T^{miss}$ tail due to the heavy mediator mass.
• Exclusion Reach:
  • For a benchmark mediator mass $m_G = 1000$ GeV and couplings $\Lambda_\gamma = \Lambda_\chi = 1$ TeV, the HL-LHC can probe cross-sections down to $\sim 2 \times 10^{-3}$ pb.
  • The projected 95% CL exclusion extends to mediator masses of $\sim 800$ GeV for $\Lambda \sim 1$ TeV, and up to the TeV scale for smaller coupling scales (larger production rates).
• Cosmological Viability:
  • Low Reheating Temperature ($T_R \sim 10-100$ MeV): The parameter space consistent with freeze-in production and relic density lies entirely within the region accessible to the HL-LHC.
  • High Reheating Temperature ($T_R \sim 1$ GeV) / Resonant Regime: These scenarios require much larger coupling scales ($\Lambda \sim 10^8$ GeV) and remain largely beyond the reach of the HL-LHC.
• Feature Importance: The BDT analysis identified $E_T^{miss}$, $m_{jj}$, and $\Delta\eta_{jj}$ as the most critical variables for separating signal from background.

5. Significance

This work provides a concrete and complementary pathway to test freeze-in dark matter scenarios, which are theoretically well-motivated by extra-dimensional models and effective gravity theories. By bridging the gap between early-Universe cosmology and high-energy collider physics, the study highlights the High-Luminosity LHC as a powerful laboratory for discovering feebly interacting sectors.

Crucially, it demonstrates that machine learning is not just an optimization tool but a necessity for probing these "dark" sectors where signal rates are suppressed and backgrounds are overwhelming. The findings suggest that a significant fraction of the cosmologically viable parameter space for spin-2 mediated freeze-in DM can be tested in the upcoming HL-LHC run, motivating dedicated search strategies tailored to VBF topologies and invisible decays.