Application of Machine Learning to 21 cm Cosmology

Technical Summary: Application of Machine Learning to 21 cm Cosmology

Problem Statement
The extraction of astrophysical and cosmological information from redshifted 21 cm observations of the Cosmic Dawn and the Epoch of Reionization (EoR) faces three primary bottlenecks:

1. Intrinsic Non-Gaussianity: The 21 cm signal arises from the nonlinear coupling of density, ionization, thermal state, and radiation backgrounds. As reionization progresses, the signal becomes strongly non-Gaussian, rendering standard two-point statistics (like the power spectrum) insufficient to capture all morphological and topological information.
2. Observational Systematics: The cosmological signal is orders of magnitude fainter than foreground emission (Galactic and extragalactic). Instrumental effects, including chromatic beams, radio-frequency interference (RFI), ionospheric distortions, and calibration errors, cause foreground power to leak into cosmological modes (the "foreground wedge"). Additionally, irregular time-frequency sampling due to RFI flagging complicates imaging.
3. Computational Burden: Robust Bayesian inference requires repeated forward modeling in high-dimensional parameter spaces. This includes varying astrophysical source parameters (e.g., X-ray efficiency, star-formation efficiency) alongside nuisance parameters for calibration and foregrounds. Brute-force sampling is prohibitively slow, and constructing explicit likelihoods with realistic covariance is difficult.

Methodology
The paper organizes the application of Machine Learning (ML) not by algorithm architecture, but by its functional role within the analysis pipeline. The review categorizes approaches into three domains:

1. Observation-Domain Methods:

   • RFI Screening: Supervised classifiers and anomaly detection tools automate the identification of contaminated time-frequency samples, scaling to SKA-Low data volumes.
   • Data Reconstruction: Learned inpainting models address irregular sampling gaps left by flagging, aiming to refill missing regions with statistically consistent surrogates without inventing structure.
   • Foreground Mitigation: Neural reconstruction and Gaussian processes with ML-enhanced kernels are used to learn the joint spatial-spectral structure of contamination, attempting to recover wedge-filtered light-cones and separate foregrounds beyond fixed analytic models.
   • Calibration and Morphology: Data-driven corrections address gain errors and beam mismatches. Segmentation networks identify ionized/neutral regions to enable morphology-based analyses (e.g., bubble-size statistics), while cross-domain translation (e.g., GANs) uses external tracers like Lyman-α emitters to condition 21 cm map reconstruction.

2. Theory-Domain Methods:

   • Emulation: Fast surrogate models (neural networks, Gaussian processes) replace repeated simulation calls to interpolate observables (like the power spectrum) across astrophysical parameter spaces, accelerating Bayesian exploration.
   • Compression and Field-Level Surrogates: Beyond simple summary-statistic emulation, recent work explores learning compact representations directly from simulated maps or light-cones. This includes fully field-level surrogates and operator-level acceleration (emulating only costly substeps or learning residuals around physical approximations) to preserve non-Gaussian information.

3. Inference-Domain Methods:

   • Direct Regression/Classification: Early applications used Convolutional Neural Networks (CNNs) to directly classify source populations or regress parameters from maps and light-cones, demonstrating information content beyond two-point summaries.
   • Simulation-Based Inference (SBI): Likelihood-free approaches (e.g., neural posterior estimation, likelihood ratio estimation using normalizing flows) learn the posterior distribution directly from simulated data. SBI allows for the conditioning on multiple summaries (power spectra, learned embeddings, multi-redshift data) without pre-selecting a single "best" statistic.

Key Contributions and Findings

• Pipeline Integration: The paper argues that ML is most effective when integrated as a specific component within a physics-based pipeline rather than as a wholesale replacement.
• Uncertainty Propagation: A critical finding is that ML models must explicitly propagate uncertainty. Fast mean predictions are insufficient; emulator errors and simulator mismatches must be quantified and marginalized to avoid biased posteriors.
• Validation Hierarchy: The review highlights a maturity gradient. RFI screening and covariance-based foreground models are relatively mature and operational. In contrast, image recovery and segmentation networks remain largely simulation-validated; their deployment requires rigorous testing against "hidden signal loss" and robustness to domain shifts (e.g., beam mismatches, calibration drift).
• Information Content: The 21 cm signal distributes information across scale, morphology, and redshift. The paper notes that while SBI can combine diverse summaries to tighten constraints, a sophisticated density model cannot recover information discarded by an impoverished summary representation.

Significance and Claims
The paper positions ML as a necessary tool for the SKA-Low era to handle data rates, calibration complexity, and high-dimensional inference. However, it maintains a modest and cautious stance regarding the current state of the field:

• Hybrid Strategy: The most defensible strategy is a hybrid workflow where ML accelerates or regularizes difficult subproblems (e.g., emulation, RFI flagging) while physics-based modeling retains responsibility for interpretation and error budgeting.
• Risk of Opaque Replacement: ML should not act as an opaque replacement for the underlying forward model. The central message is that ML is useful only when it preserves physically relevant structure and propagates uncertainty explicitly.
• Calibration as a Priority: The decisive challenge for inference-domain ML is not merely parameter recovery on held-out simulations, but ensuring that recovered posteriors remain trustworthy under perturbations (simulator mismatch, domain shift). The paper emphasizes that coverage tests and simulation-based calibration are now first-class requirements for any ML application in this domain.

In summary, the paper asserts that while ML offers powerful capabilities for accelerating forward modeling and extracting non-Gaussian information, its scientific value in 21 cm cosmology depends entirely on rigorous validation, uncertainty quantification, and the preservation of physical interpretability.