Rico: An Accurate Cosmological Recombination Code

TL;DR

Rico is a machine learning-based code that rapidly and accurately computes the Universe's ionization history during recombination, facilitating precise cosmological parameter estimation from CMB data.

Contribution

It introduces Rico, a novel machine learning approach trained on detailed recombination calculations, enabling fast and accurate ionization history predictions without complex fudge factors.

Findings

Rico computes ionization history in ~10 milliseconds with negligible error.

It reproduces full recombination code results with below 0.1% accuracy.

Rico accelerates cosmological analyses by over a million times.

Abstract

We present Rico, a code designed to compute the ionization fraction of the Universe during the epoch of hydrogen and helium recombination with an unprecedented combination of speed and accuracy. This is accomplished by training the machine learning code Pico on the calculations of a multi-level cosmological recombination code which self-consistently includes several physical processes that were neglected previously. After training, Rico is used to fit the free electron fraction as a function of the cosmological parameters. While, for example at low redshifts (z<~900), much of the net change in the ionization fraction can be captured by lowering the hydrogen fudge factor in Recfast by about 3%, Rico provides a means of effectively using the accurate ionization history of the full recombination code in the standard cosmological parameter estimation framework without the need to add new or refined fudge factors or functions to a simple recombination model. Within the new approach presented here it is easy to update Rico whenever a more accurate full recombination code becomes available. Once trained, Rico computes the cosmological ionization history with negligible fitting error in ~10 milliseconds, a speed-up of at least 10^6 over the full recombination code that was used here. Also Rico is able to reproduce the ionization history of the full code to a level well below 0.1%, thereby ensuring that the theoretical power spectra of CMB fluctuations can be computed to sufficient accuracy and speed for analysis from upcoming CMB experiments like Planck. Furthermore it will enable cross-checking different recombination codes across cosmological parameter space, a comparison that will be very important in order to assure the accurate interpretation of future cosmic microwave background data.