Application of Machine Learning to 21 cm Cosmology

Abstract: In this chapter, the use of ML in redshifted 21 cm cosmology is discussed, especially for the cosmic dawn, the Epoch of Reionization, and the scientific program of SKA-Low. The 21 cm signal is useful because it can directly probe diffuse neutral hydrogen. At the same time, it is a difficult signal, since the observable depends on density, ionization, heating, radiation backgrounds, and instrumental response in a nonlinear way. The first part of this chapter reviews the basic physical ingredients needed for the later discussion, including the global signal, spatial fluctuations, morphology-aware summaries, and the 21 cm forest. The next part describes the main difficulties for realistic analysis, such as bright foregrounds, radio-frequency interference, ionospheric and calibration effects, incomplete sampling, and the cost of forward modeling in large parameter spaces. Based on this background, ML applications are grouped by their role in the analysis pipeline. Observation-domain methods work on contaminated data products; theory-domain methods accelerate or compress forward modeling; and inference-domain methods connect observables to astrophysical and cosmological constraints. The 21 cm forest is also discussed as a case where one-dimensional spectra, small-scale information, and uncertain source populations make ML useful.

• The paper shows that machine learning methodologies efficiently overcome computational and systematic challenges in analyzing the faint 21 cm cosmological signal.
• It details advanced techniques such as emulators and simulation-based inference to accelerate parameter estimation and improve uncertainty quantification.
• The study emphasizes a hybrid approach that integrates ML tools with physical models to address domain shift and ensure robust, interpretable cosmological inferences.

Application of Machine Learning to 21 cm Cosmology: An Expert Review

Context and Motivation

The analysis of the redshifted 21 cm signal provides a direct means of probing the thermal and ionization history of the intergalactic medium during the Dark Ages, cosmic dawn, and the Epoch of Reionization (EoR). However, extracting and interpreting this faint cosmological signal are hampered by the intrinsic nonlinearity of the underlying astrophysics, severe foreground contamination, instrumental systematics, and the computational demands of high-dimensional inference pipelines. This work systematically delineates the ways in which ML addresses these challenges, focusing on the SKA-Low era, where analysis pipelines must process unprecedented data volumes and handle complex calibration, uncertainty quantification, and model validation requirements (2605.10105).

The Physics of the 21 cm Signal

The 21 cm emission/absorption line arises from the hyperfine transition of neutral hydrogen. The observable brightness temperature $\delta T_b(x, z)$ is a nonlinear function of local density, ionization state, kinetic and spin temperatures, and background radiation fields. Its detectability depends critically on the difference between the spin temperature $T_s$ and the cosmic microwave background (CMB) temperature $T_\gamma$. Observational summaries bifurcate into the global signal—representing the sky-averaged thermal and ionization history—and spatial fluctuations characterized via the power spectrum and higher-order statistics.

The inherent non-Gaussianity of the 21 cm field during EoR renders two-point statistics incomplete, motivating the use of additional descriptions such as the one-point PDF, bispectrum, bubble-size distributions, Minkowski functionals, Betti numbers, and wavelet-based summaries. These provide access to information on field connectivity, morphological structure, and non-Gaussian signatures, each with complementary sensitivities and limitations.

Of particular note is the 21 cm forest—narrow absorption features observed against high-redshift radio sources—offering sensitivity to cold, small-scale structure and dark-matter microphysics that evade traditional imaging analyses.

Practical and Computational Challenges

Key bottlenecks in extracting physical information from 21 cm data include:

• Degeneracy: Astrophysical parameters often yield similar summary statistics.
• Nonlinearity and High Dimensionality: Morphology, redshift evolution, and scale-dependent structure demand analyses that do not overcompress the data.
• Foregrounds and Systematics: Foreground emission and calibration errors exceed the cosmological signal amplitude by several orders of magnitude, while instrumental transfer functions induce mode mixing ("foreground wedge").
• Computational Expense: Bayesian inference over high-dimensional parameter spaces with full uncertainty propagation requires thousands of forward simulations, further complicated by emulator error and instrumental model uncertainty.

Machine Learning Applications Across the Analysis Pipeline

ML methods are systematically categorized according to their operational locus:

Observation-Domain Methods

These methods process data prior to cosmological inference. Major applications include:

• RFI Mitigation: Supervised classifiers and anomaly detection flag contaminated samples, improving robustness but requiring stringent validation to avoid cosmological signal loss.
• Missing Data and Inpainting: Learned interpolation algorithms reconstruct flagged regions; the principal risk is artificial structure insertion that biases statistical summaries.
• Foreground Mitigation and Calibration: Neural models and ML-enhanced Gaussian processes model spatial–spectral foreground covariance, and learned corrections stabilize calibration residuals under complex instrumental transfer functions.
• Morphology Reconstruction: Segmentation and image-recovery networks infer bubble morphologies or topological features from noisy, foreground-degraded maps, but require benchmarks based on information-preserving statistical summaries rather than visual fidelity.
• Cross-Modality Translation: Conditional GANs and related models translate between Lyman-α emission maps and 21 cm fields for cross-survey inference or conditioning.

Validation for these methods must focus on scientific metrics (information preservation, error calibration, robustness to distribution shift), not merely empirical performance or visual plausibility.

Theory-Domain Methods

ML accelerates forward modeling and summary statistical computation:

• Emulators: Neural nets, Gaussian processes, and VAEs interpolate expensive summary statistics (e.g., the 21 cm power spectrum) across parameter space, as in 21cmEMU or globalemu. Uncertainty quantification and explicit error propagation are essential to avoid biased constraints.
• Field-Level Surrogates: Generative models for 3D fields remain nascent; their reliability under variation in underlying astrophysical conditions and their preservation of physically meaningful structure remain open questions.
• Operator-Level Acceleration: Learning surrogates for expensive substeps (e.g., radiative transfer, ionization bubble growth) is theoretically appealing but less developed.

Appropriate usage of ML in the theory domain maintains visibility of the underlying physical model and treats surrogate discrepancy as part of the error budget (2605.10105).

Inference-Domain Methods

ML-based inference ranges from deterministic regression on simulation outputs to fully probabilistic SBI:

• Direct Regression and Classification: Early work demonstrated that CNNs and similar models extract astrophysical parameters or classify EoR scenarios from maps or power spectra; these are computationally efficient but limited in uncertainty quantification.
• Simulation-Based Inference (SBI): Normalizing flows and other neural density estimators implement likelihood-free inference, enabling flexible analysis of complex observables and direct posterior estimation. Recent studies highlight the importance of joint summary combinations to leverage non-Gaussian information and the criticality of rigorous calibration (simulation-based diagnostics, posterior predictive checks) [87].
• 21 cm Forest Analysis: ML enables absorption feature identification and latent compression in low-SNR, high-degeneracy regimes. Persistent homology and other topological summaries provide information on absorption structure not captured by amplitude-based statistics.

In all cases, robustness to domain shift and careful propagation of uncertainties associated with simulation assumptions, foregrounds, and source populations are mandatory.

Strong Numerical Results, Claims, and Limitations

The paper makes specific, authoritative points:

• No single summary is sufficient: The power spectrum alone fails to capture all the information content of the 21 cm field during EoR, as structure and morphology encode independent and complementary constraints.
• Emulation has proven quantitative utility: Emulators yield high-precision, order-of-magnitude speedups in MCMC or Bayesian parameter estimation workflows while maintaining accuracy when error propagation is implemented.
• SBI tightens parameter posteriors: Combining multiple summary statistics or learned embeddings in an SBI framework produces statistically tighter constraints than traditional approaches utilizing the power spectrum alone.
• Domain shift is a principal failure mode: Neural methods can produce overconfident or misleading predictions when trained on simulations that do not span the full range of instrumental and astrophysical variability encountered in real data.

Implications and Future Development

In the SKA-Low context, the main implication is that end-to-end ML solutions are, at present, insufficiently mature to replace physically-motivated, modular pipelines. The optimal strategy is a hybrid approach: ML should target specific subproblems—RFI flagging, summary compression, forward emulation—while maintaining physical transparency, interpretability, and visible uncertainty budgets.

Theoretically, the integration of morphological and topological summaries with advanced density estimation promises improved discrimination between astrophysical scenarios that are degenerate in standard statistics. Practically, robust ML adoption will require scalable, auditable, and uncertainty-calibrated workflows. As data rates and dimensionality increase, the demand for high-fidelity generative models and transfer learning techniques capable of mitigating domain shift will intensify. The evolving interface between data-driven compression and physically interpretable summaries merits ongoing scrutiny.

The 21 cm forest, with its sensitivity to small-scale IGM physics and dark-matter phenomenology, highlights the need for ML methodologies that are rigorous not only in data analysis but also in their accommodation of astrophysical and instrumental uncertainties inherent in next-generation surveys.

Conclusion

The reviewed chapter provides a comprehensive, critical account of ML applications in 21 cm cosmology, organized by their analytical role. The methodological consensus is that while ML enables essential advances in computational tractability and information extraction, its deployment must be selective, interpretable, and validated under physically realistic conditions. The field’s trajectory suggests increasing integration of ML-accelerated summaries, uncertainty-aware inference, and hybrid pipelines, but with persistent emphasis on transparency and robust validation. This balanced approach will be central for exploiting the high precision and information richness anticipated from current and forthcoming 21 cm cosmological surveys (2605.10105).