Presentation #202.04 in the session Cosmology — iPoster Session.
The reionization history uniquely determines the optical depth to reionization, “tau”, for a fixed cosmology. One method to determine tau is to have neural networks trained on simulated data to predict the ionization history of hydrogen from HERA data. Two leading simulators for modeling a reionization history quickly are 21cmFAST and Zreion. Zreion uses a biased density field prescription to create the ionization field (in reference to Battaglia et al. 2013; ApJ 776 31), whereas 21cmFAST (Murray et al. 2020; JOSS 5(54) 2582; Mesinger et al. 2011; MNRAS 411 955) uses a more physical description of the star formation and ionized photon creation process. To create a more diverse set of training data and thus in principle a more robust prediction of tau, we need to standardize the reionization histories between two semi-analytical models. To do this, I built a “simulator translator” using an optimized Pytorch emulator (Kern, N. S., Liu, A., Parsons, A. R., Mesinger, A., Greig, B. 2017, ApJ, 848, 23) to predict the simulated ionization histories from the free parameters of a simulator and a chi-square minimization process to match corresponding ionization histories. Because the comparison between ionization histories from the two emulators is done via a least-squares fit, it is possible to use the chi-squared as a measure of the goodness-of-fit between two parameter sets (i.e., whether 21cmFAST can produce the same shape of ionization history as Zreion or not) and to use the covariance of the fitted parameters as a measure of the uniqueness of the match. Furthermore, all of our simulations share as much as possible in common: density field, average brightness temperature, the ionization history and differ only in the way the ionization field is implemented. Indeed, we made similarity of the ionization history our figure of merit for the translator, specifically with the covariance matrix and chi-squared values.