PINION: Physics-informed neural network for accelerating radiative transfer simulations for cosmic reionization

TL;DR

PINION is a physics-informed neural network that accelerates cosmic reionization simulations by accurately predicting hydrogen ionization evolution from density fields, reducing computational costs and enabling large-scale studies.

Contribution

The paper introduces PINION, a novel neural network model that incorporates physical constraints to efficiently simulate cosmic reionization history from N-body simulation data.

Findings

PINION accurately predicts reionization history between redshifts 6 and 12.

The model's predictions agree well with detailed radiative transfer simulations for z>7.

Performance decreases for z<7 due to oversimplified propagation modeling.

Abstract

With the advent of the Square Kilometre Array Observatory (SKAO), scientists will be able to directly observe the Epoch of Reionization by mapping the distribution of neutral hydrogen at different redshifts. While physically motivated results can be simulated with radiative transfer codes, these simulations are computationally expensive and can not readily produce the required scale and resolution simultaneously. Here we introduce the Physics-Informed neural Network for reIONization (PINION), which can accurately and swiftly predict the complete 4-D hydrogen fraction evolution from the smoothed gas and mass density fields from pre-computed N-body simulation. We trained PINION on the C$^2$-Ray simulation outputs and a physics constraint on the reionization chemistry equation is enforced. With only five redshift snapshots and a propagation mask as a simplistic approximation of the ionizing photon mean free path, PINION can accurately predict the entire reionization history between $z=6$ and $12$. We evaluate the accuracy of our predictions by analysing the dimensionless power spectra and morphology statistics estimations against C$^2$-Ray results. We show that while the network's predictions are in good agreement with simulation to redshift $z>7$, the network's accuracy suffers for $z<7$ primarily due to the oversimplified propagation mask. We motivate how PINION performance can be drastically improved and potentially generalized to large-scale simulations.