Testing a 95 GeV Scalar at the CEPC with Machine Learning

This paper demonstrates that the Circular Electron Positron Collider (CEPC) can efficiently test the 95 GeV scalar hypothesis in the Higgsstrahlung channel by identifying 210 GeV as the optimal center-of-mass energy and utilizing deep neural networks to significantly reduce the luminosity required for discovery and precision measurement.

The Gist

Imagine the Standard Model of physics as a completed puzzle. For years, scientists have been trying to find the missing pieces that might explain things like dark matter or why the universe has more matter than antimatter. Recently, several different experiments have spotted a faint, blurry "ghost" in the data—a potential new particle weighing in at about 95 GeV (roughly the mass of a silver atom). It's not a clear picture yet; it's just a hint that something is there.

This paper is like a detective's manual for the CEPC (Circular Electron-Positron Collider), a massive particle accelerator planned for China. The authors are asking: "If we build this machine, how can we best catch this 95 GeV ghost?"

Here is the breakdown of their investigation, explained through everyday analogies:

1. The Right Speed for the Hunt

The CEPC is designed to run at a specific speed (energy) to study the famous 125 GeV Higgs boson. But the new "95 GeV ghost" is lighter.

• The Analogy: Imagine you are trying to catch a specific type of bird. If you fly your plane too fast, you might overshoot it. If you fly too slow, you might not reach its altitude.
• The Discovery: The authors ran simulations and found that the CEPC shouldn't fly at its "design speed" (240 GeV) to catch this specific bird. Instead, it needs to slow down to 210 GeV. At this specific speed, the conditions are perfect to spot the 95 GeV particle.

2. The Needle in a Haystack Problem

Finding this particle is incredibly hard because it looks almost exactly like background noise.

• The Analogy: Imagine you are trying to hear a specific whisper (the signal) in a crowded stadium where everyone is shouting (the background noise). The "whisper" is the collision creating the new particle, and the "shouting" is the billions of other collisions happening at the same time.
• The Solution: The team used a clever trick called "Recoil Mass." Instead of trying to listen to the whisper directly (which is messy), they measured the reaction of the crowd. By looking at the momentum of the muons (particles that fly out cleanly), they could calculate the weight of the invisible particle without needing to see it directly. It's like deducing the weight of a hidden box by how much the scale tips when you lift it.

3. The AI Super-Helper

Even with the right speed and the recoil trick, the "shouting" background is still too loud. The signal is still buried.

• The Analogy: You have a security camera feed with thousands of people walking by. You need to find one specific person wearing a red hat. A human looking at the screen would get tired and miss them.
• The Solution: The authors trained Machine Learning (AI) algorithms (specifically Deep Neural Networks) to be the super-observer. They fed the AI thousands of examples of "noise" and "signal."
• The Result: The AI became incredibly good at spotting the pattern. It acted like a filter that instantly muted the shouting crowd and amplified the whisper. Using this AI, they found they needed half the amount of data (luminosity) to confirm the particle's existence compared to using old-school methods.

4. The Game Plan

The paper concludes with a clear strategy for the future:

• Run at 210 GeV: Don't stick to the original 240 GeV plan; adjust the machine to hunt this specific particle.
• Use AI: Turn on the machine learning filters to cut through the noise.
• The Payoff: With this strategy, the CEPC could confirm or rule out the existence of this 95 GeV particle with very little data (less than a year of running time at full power). If the particle exists, the CEPC could measure its properties with incredible precision (within 5% accuracy).

The Bottom Line

Think of this paper as a GPS route update for the CEPC. The original map said, "Go straight at top speed." The authors say, "Actually, take this specific exit ramp at a slightly slower speed, and use this new AI navigation system." If they do this, they have the best possible chance of finally solving the mystery of the 95 GeV particle and discovering what lies beyond our current understanding of the universe.

Technical Summary

Here is a detailed technical summary of the paper "Testing a 95 GeV Scalar at the CEPC with Machine Learning."

1. Problem Statement

Recent experimental data from LEP, CMS, and ATLAS have reported local excesses around 95 GeV in various channels ($b\bar{b}$, $\gamma\gamma$, and $\tau^+\tau^-$). These anomalies suggest the existence of a light scalar particle ($S$) beyond the Standard Model (BSM).

• The Challenge: The Circular Electron-Positron Collider (CEPC) is primarily designed to operate at $\sqrt{s} = 240$ GeV for precision measurements of the 125 GeV Higgs boson. However, a 95 GeV scalar is significantly lighter than this design target.
• The Gap: It is unclear whether the CEPC's standard design energy is optimal for discovering a 95 GeV scalar, nor is it known how to maximize sensitivity against dominant backgrounds (specifically the irreducible $Z\tau\tau$ background) without excessive luminosity requirements.
• Goal: To determine the optimal center-of-mass energy ($\sqrt{s}$) for the CEPC to probe a 95 GeV scalar in the Higgsstrahlung channel ($e^+e^- \to ZS$) with $Z \to \mu^+\mu^-$ and $S \to \tau^+\tau^-$, and to evaluate the impact of Machine Learning (ML) on discovery potential.

2. Methodology

A. Simulation and Event Selection

• Process: The study focuses on $e^+e^- \to Z(\to \mu^+\mu^-)S(\to \tau^+\tau^-)$.
• Backgrounds:
  • Irreducible: $e^+e^- \to Z(\to \mu^+\mu^-)\tau^+\tau^-$ ($Z\tau\tau$).
  • Reducible: $e^+e^- \to Z(\to \mu^+\mu^-)XX$ ($Zjj$), where jets are misidentified as $\tau$-jets.
• Tools: Events were generated using MadGraph5_aMC@NLO, showered/hadronized with PYTHIA 8, and simulated through a CEPC detector model using Delphes 3.5.0 (with 80% $\tau$-tagging efficiency).
• Key Observable: Due to the difficulty of reconstructing $\tau$ decays, the authors utilize the Recoil Mass ($M_{recoil}$) calculated from the well-measured muon pair:
  $$M_{recoil} \equiv \sqrt{s + M_{\mu^+\mu^-}^2 - 2\sqrt{s}(E_{\mu^+} + E_{\mu^-})}$$
  This allows inference of the scalar mass without relying on $\tau$ reconstruction.
• Baseline Cuts: Requirements include $p_T(\mu) > 10$ GeV, $p_T(\tau) > 20$ GeV, and a recoil mass window $|M_{recoil} - M_S| < 1.5$ GeV.

B. Machine Learning Optimization

To further suppress backgrounds and enhance signal significance, the authors compared three classifiers:

1. Gradient Boosting Decision Tree (GBDT)
2. eXtreme Gradient Boosting (XGBoost)
3. Deep Neural Network (DNN)

• Input Features: 22 kinematic variables including four-momenta, transverse momenta, pseudorapidities, angular separations ($\Delta R$), and the recoil mass.
• DNN Architecture: Six hidden layers (64, 48, 32, 24, 16, 8 neurons) with ReLU activation, 20% dropout, and Adam optimizer.
• Performance: The DNN achieved the highest classification accuracy (~78%) and Area Under the Curve (AUC ~0.87) at $\sqrt{s}=210$ GeV.

C. Theoretical Framework

The study applied these methods to the N2HDM-Flipped (Next-to-Two Higgs Doublet Model), a specific BSM scenario that can accommodate the 95 GeV excesses. The analysis scanned 11 parameters (masses, $\tan\beta$, couplings, mixing angles) subject to theoretical constraints (unitarity, stability) and experimental limits (HiggsBounds, HiggsSignals).

3. Key Contributions & Results

A. Optimal Center-of-Mass Energy

• Scanning $\sqrt{s}$ from 190 to 240 GeV revealed that $\sqrt{s} = 210$ GeV is the optimal energy for discovering a 95 GeV scalar.
• At the benchmark point ($M_S = 95.5$ GeV, $\sqrt{s} = 210$ GeV), the signal significance reaches $Z \approx 5.1\sigma$ with an integrated luminosity of $500 \text{ fb}^{-1}$.
• Achieving the same significance at the design energy of 240 GeV would require $\approx 720 \text{ fb}^{-1}$ (a 1.4x increase in luminosity).

B. Impact of Machine Learning

• ML classifiers significantly reduce the luminosity required for discovery.
• Luminosity Reduction: At $\sqrt{s} = 210$ GeV, the DNN reduces the required luminosity for $5\sigma$discovery from **480 fb$^{-1}$** (baseline cuts only) to **210 fb$^{-1}$**.
• At $\sqrt{s} = 240$ GeV, the reduction is from 690 fb$^{-1}$ to 310 fb$^{-1}$.
• The DNN consistently outperformed GBDT and XGBoost across all signal efficiencies.

C. Sensitivity and Precision

With the full CEPC design luminosity ($L = 20 \text{ ab}^{-1}$):

• Discovery Sensitivity: The CEPC can probe signal strengths ($\mu_{ZS}^{\tau\tau}$) down to 0.016 at 210 GeV and 0.020 at 240 GeV (5$\sigma$ level).
• Precision Measurement: A 5% precision measurement on the signal strength is achievable for $\mu_{ZS}^{\tau\tau} > 0.10$ (at 210 GeV) and $> 0.12$ (at 240 GeV).

D. N2HDM-Flipped Coverage

• The study tested the model's viability against current experimental excesses ($\chi^2_{h95} < 7.82$).
• Coverage: A 210 GeV run with $L = 800 \text{ fb}^{-1}$ is sufficient to cover 100% of the surviving N2HDM-Flipped parameter space at the 5$\sigma$ level.
• At 240 GeV, $L = 1.22 \text{ ab}^{-1}$ is required to achieve the same coverage.
• The best-fit point ($\chi^2_{h95} = 0.24$) could be confirmed with as little as $58 \text{ fb}^{-1}$ at 210 GeV.

4. Significance

• Strategic Recommendation: The paper establishes that an early run of the CEPC at 210 GeV, augmented by Deep Neural Network selection, is the most efficient strategy to confirm or refute the 95 GeV excess. This is more efficient than waiting for the standard 240 GeV run.
• Methodological Advance: It demonstrates that ML techniques can halve the luminosity requirements for light scalar searches, making the CEPC a powerful tool for BSM physics even outside its primary design parameters.
• Model Validation: The results provide a concrete roadmap for testing the N2HDM-Flipped and similar extended Higgs sector models, showing that the CEPC has the potential to fully cover the parameter space compatible with current experimental hints with relatively modest integrated luminosity.