ML-GPR user guide

ML-GPR separates frequency-dependent 21-cm observations by assigning a Gaussian-process covariance to each physical component. A typical model contains smooth intrinsic foregrounds, less-smooth chromatic contamination, a 21-cm component, and heteroscedastic thermal noise. The fit produces a posterior distribution of component reconstructions and power spectra rather than one deterministic foreground subtraction.

This guide explains model construction, fitting, inference, and posterior products. Package installation is covered in Getting started. The API reference documents individual Python objects.

• Migrating ML-GPR from 0.x to 1.0
  • The high-level API is unchanged
  • Advanced API changes
  • VAE kernels
  • Saved results and trained VAEs
• Running an ML-GPR fit
  • Configuration-driven interface
  • Advanced interface
• Building the covariance model
  • Configuration example
  • Component names and composition
  • Available frequency kernels
  • Available priors
  • UV-dependent parameterizations
  • Units and normalization
  • Noise model
• Choosing an inference method
  • Ensemble MCMC
  • MAP and Laplace approximation
  • Dynesty nested sampling
  • NUTS
  • UltraNest
  • Running inference
  • Running sampler classes directly
  • Assessing inference quality
• Posterior components and power spectra
  • Common products
  • A simulated posterior
  • Selecting components
  • Saving and loading
• Multi-epoch and Cross-GPR models
  • Data requirements
  • Configuring a time-dependent component
  • Power spectra of the coherent and independent parts
  • Choosing a time covariance
• Tutorial notebooks