Neural network emulation of reionization to constrain new physics with early- and late-time probes

Abstract: The optical depth to reionization, a key parameter of the $\Lambda$CDM model, can be computed within astrophysical frameworks for star formation by modeling the evolution of the intergalactic medium. Accurate evaluation of this parameter is thus crucial for joint statistical analyses of CMB data and late-time probes such as the 21 cm power spectrum, requiring consistent integration into cosmological solvers. However, modeling the optical depth with sufficient precision in a computationally feasible manner for MCMC analyses is challenging due to the complexities of the nonlinear astrophysics. We introduce NNERO (Neural Network Emulator for Reionization and Optical depth), a framework that leverages neural networks to emulate the evolution of the free-electron fraction during cosmic dawn and reionization. We demonstrate its effectiveness by simultaneously constraining cosmological and astrophysical parameters in both standard cold dark matter and non-cold dark matter scenarios, including models with massive neutrinos and warm dark matter, showcasing its potential for efficient and accurate parameter inference.