Uncertainty-Aware Neural Networks for Fuzzy Dark Matter Model Selection from \texorpdfstring{$x_{\rm HI}$}{x_HI} Measurements

This study employs an uncertainty-aware hybrid machine learning framework trained on Bayesian-inferred observational probability density functions to constrain fuzzy dark matter models using 21-cm neutral hydrogen fraction data, identifying a best-fit scenario with a particle mass of approximately $10^{-22}\,\mathrm{eV}$ and a fraction of 0.04 while ruling out lighter masses.

The Gist

Imagine the universe as a giant, dark ocean. For a long time, scientists have believed this ocean is made of "Cold Dark Matter" (CDM)—think of it as invisible, heavy sand grains that clump together easily to form islands (galaxies).

But there's a problem. When we look at the smallest islands, the "sand grain" theory predicts they should be messy and spiky. However, observations show they are actually smooth and round. It's like expecting a pile of sand to look like a jagged rock, but finding it looks like a smooth pebble instead.

This paper proposes a different kind of ocean: one made of Fuzzy Dark Matter (FDM). Instead of heavy sand, imagine the dark matter is made of ghostly, ultra-light waves (like ripples on a pond). Because these waves are so fuzzy, they can't clump together in tiny spaces. They smooth out the small islands, creating the smooth pebbles we actually see.

Here is how the authors solved the mystery of which type of dark matter is real, using a mix of cosmic simulations and AI detectives.

1. The Cosmic "What-If" Machine

The researchers used a supercomputer to run thousands of simulations of the early universe. They asked: "What if the dark matter waves have different weights?"

• Some waves were super light (very fuzzy).
• Some were slightly heavier.
• They mixed in different amounts of this "fuzzy stuff" (from 2% to 10% of the total dark matter).

For every scenario, they calculated how much neutral hydrogen (the gas that hasn't been turned into stars yet) existed at different times in the universe's history. Think of this as predicting how much "fog" was in the universe at different ages.

2. The New "Cosmic Camera" (JWST)

Enter the James Webb Space Telescope (JWST). It's like a brand-new, super-powerful camera that can see deeper into the past than any telescope before. It took pictures of the early universe and measured exactly how much "fog" (neutral hydrogen) was there.

But here's the catch: The camera isn't perfect. The measurements have "fuzziness" (uncertainty) too. Sometimes the fog looks like it's 50% gone, sometimes 60%. It's not a single number; it's a range of possibilities.

3. The AI Detective (The Hybrid Neural Network)

This is where the paper gets clever. Instead of just comparing the simulation numbers to the telescope numbers, the authors built a hybrid AI detective.

• The CNN (The Pattern Spotter): Imagine a detective looking at a map. This part of the AI looks at the shape of the data to find local patterns.
• The RNN (The Time Traveler): This part looks at the story over time. It understands that the fog doesn't disappear all at once; it fades gradually as the universe ages.

The Secret Sauce: Most AI models just look for the "average" answer. This AI was trained to understand uncertainty. It didn't just say, "The fog is 50%." It said, "The fog is likely between 45% and 55%, and here is the exact probability curve of that." It learned to respect the "fuzziness" of the telescope data, just like the fuzziness of the FDM waves.

4. The Verdict: Who Won the Race?

The AI compared the "ghost wave" simulations against the "real photo" from the JWST.

• The Losers: The simulations where the waves were too light (super fuzzy) were ruled out. They kept the fog around for too long, which didn't match the photos.
• The Winner: The simulation that matched the JWST photos best was one where the dark matter waves had a specific "weight" (mass) of about $10^{-22}$ electron-volts and made up about 4% of the total dark matter.

In this winning scenario, the "ghost waves" slowed down the formation of the very first stars just enough to match what we see today. It's like a traffic light that turned red a split second later than expected, causing a specific pattern of cars (stars) that matches our observations.

Why Does This Matter?

This paper is a breakthrough because it didn't just guess; it used a smart, uncertainty-aware AI to bridge the gap between complex physics simulations and real telescope data.

It tells us that the "fuzzy wave" theory is a strong contender for explaining the universe's dark matter. It suggests that the early universe was a bit more "patient" in forming stars than we thought, and that the invisible stuff holding galaxies together might be made of quantum waves rather than invisible particles.

In short: The authors built a time-traveling AI detective that looked at the universe's childhood photos, figured out the "fuzziness" of the data, and concluded that the universe is likely made of "fuzzy waves" rather than "cold sand."

Technical Summary

1. Problem Statement

The nature of dark matter remains a central unsolved problem in cosmology. While the Cold Dark Matter (CDM) paradigm is the standard model, it faces small-scale challenges (e.g., the core-cusp problem, missing satellites). Fuzzy Dark Matter (FDM), composed of ultra-light axions ($m \sim 10^{-22}$ eV), offers an alternative where quantum wave pressure suppresses small-scale structure formation.

A critical test for FDM is the Epoch of Reionization (EoR). FDM delays the collapse of low-mass halos, postponing star formation and altering the evolution of the neutral hydrogen fraction ($x_{\text{HI}}$) compared to CDM. However, constraining FDM parameters (particle mass $m_{\text{FDM}}$ and fraction $f_{\text{FDM}}$) is difficult because:

• Observational data (from JWST, quasars, GRBs) contains significant, often non-Gaussian and correlated uncertainties.
• Standard machine learning approaches often treat data points as independent with Gaussian errors, failing to capture the full probability density of the observations.
• There is a need to distinguish between FDM scenarios and standard CDM using high-redshift $x_{\text{HI}}$ measurements.

2. Methodology

The authors propose a hybrid uncertainty-aware machine learning framework that integrates cosmological simulations with observational data constraints.

A. Data Generation and Simulation

• Simulation Tool: They use 21cmFirstCLASS, a semi-numerical code combining 21cmFAST (for 21-cm signal simulation) and CLASS (for cosmological Boltzmann equations), specifically modified to include FDM physics.
• Parameter Space: A discrete grid was explored with:
  • Particle masses: $m_{\text{FDM}} \in \{10^{-21}, \dots, 10^{-24}\}$ eV.
  • FDM fractions: $f_{\text{FDM}} \in \{0.02, 0.04, 0.06, 0.08, 0.10\}$.
• Output: Global neutral hydrogen fractions ($x_{\text{HI}}$) as a function of redshift ($z$).

B. Observational Data Processing

• Source: The study primarily utilizes JWST observations (Ly$\alpha$ damping wings, luminosity functions, equivalent widths) for $z \gtrsim 7$, supplemented by lower-redshift constraints (quasars, GRBs).
• Bayesian Inference: Instead of using point estimates with error bars, the authors use the No-U-Turn Sampler (NUTS) to estimate full probability density functions (PDFs) for $x_{\text{HI}}$ and $z$.
• Result: This yields non-Gaussian, multivariate uncertainty distributions that capture correlations between redshift and neutral fraction, avoiding assumptions of Gaussian errors.

C. Neural Network Architecture

A Hybrid CNN-RNN architecture is designed to learn the mapping between simulated FDM scenarios and the observational PDFs:

1. Convolutional Neural Networks (CNN): Extract spatial features and correlated patterns within the uncertainty maps of the $x_{\text{HI}}$ distributions.
2. Recurrent Neural Networks (RNN): Specifically GRU (Gated Recurrent Units) and LSTM (Long Short-Term Memory) layers are used to model temporal dependencies across redshift steps (cosmic time evolution).
3. Training Strategy:
   • Uncertainty-Weighted Fine-Tuning: The loss function incorporates observational uncertainties as multiplicative scaling factors. Samples with higher posterior probability (from NUTS) contribute more to the loss, while outliers are downweighted.
   • Regularization: Dropout (15%) and Early Stopping are used to prevent overfitting.
   • Optimization: Adam optimizer with Bayesian optimization for hyperparameter tuning.

3. Key Contributions

1. Uncertainty-Aware Framework: The paper introduces a novel method to train neural networks directly on full non-Gaussian posterior distributions derived from Bayesian inference, rather than simplified point estimates. This preserves the true statistical structure of JWST data.
2. Hybrid Deep Learning Model: The integration of CNNs (for spatial uncertainty structure) and RNNs (for redshift evolution) allows the model to capture complex, non-linear relationships between FDM parameters and the reionization history.
3. Model Selection via Simulation Comparison: The trained network acts as an emulator to compare theoretical FDM simulations against the "machine-learned" observational trends, objectively identifying the best-fit parameters.

4. Key Results

• Best-Fit Parameters: The analysis identifies the FDM model with $m_{\text{FDM}} \simeq 10^{-22}$ eV and $f_{\text{FDM}} \simeq 0.04$ as the configuration that best matches current JWST and other observational constraints.
• Performance Metrics: The hybrid model achieved a coefficient of determination $R^2 = 0.975$ and a Mean Squared Error (MSE) of $1.93 \times 10^{-3}$ when reproducing simulation outputs trained on noisy observational data.
• Constraints on Lighter Masses: Models with significantly lighter masses ($m_{\text{FDM}} < 10^{-22}$ eV) are strongly constrained. They suppress structure formation too aggressively, leading to $x_{\text{HI}}$ curves that overshoot observational upper bounds or diverge from the JWST-inferred range at $z \geq 8$.
• Physical Implications:
  • The preferred FDM model predicts a delayed and more gradual reionization compared to standard $\Lambda$CDM.
  • The global 21-cm signal ($T_{21}$) shows an earlier onset of the absorption trough and a longer duration of the absorption phase due to suppressed small-scale halo collapse.
  • The model remains consistent with independent constraints from the CMB, Ly$\alpha$ forest, galaxy rotation curves, and black hole spin limits.

5. Significance

• Bridging Theory and Observation: This work demonstrates how advanced machine learning can effectively bridge the gap between complex cosmological simulations and high-dimensional, uncertain observational data.
• JWST Synergy: It highlights the critical role of JWST in constraining dark matter physics, specifically showing that high-redshift ($z > 7$) neutral hydrogen measurements are sensitive to the quantum nature of dark matter.
• Future Directions: The framework provides a robust template for future studies involving next-generation facilities like the Square Kilometre Array (SKA) and HERA, suggesting that likelihood-free inference and physics-informed neural networks can further refine constraints on the dark matter sector.

In conclusion, the paper successfully utilizes an uncertainty-aware neural network to select a specific FDM scenario ($m \approx 10^{-22}$ eV, $f \approx 4\%$) that reconciles theoretical predictions with the latest JWST observations, offering a compelling alternative to the standard CDM paradigm on small scales.