By mapping the large-scale thermal and ionization stucture of the intergalactic medium (IGM), cosmic 21-cm interferometry is set to revolutionize our understanding of the Epoch of Reionization (EoR) and the Cosmic Dawn (CD). However, interpreting upcoming observations relies on our ability to accurately model early galaxies and their cosmic radiation fields which drive the signal. The first galaxies are likely very rare and biased, with their abundances modulated by long wavelength modes of the density field. Moreover, the radiation fields from these galaxies interact with the IGM over a large range of scales. Therefore, the limited volume of 21-cm simulations can underestimate the amount of structure in the signal. In this work we quantify the minimum box size for simulating the power spectrum (PS) of the cosmic 21-cm signal. We perform multiple realizations of the initial conditions, for a range of box sizes. We quantify convergence with respect to a mock observation of box length 1.1 Gpc, with thermal noise computed for a 1000h observation with SKA1-low. We find that simulations of box lengths 200-300 Mpc typically do not show a bias in the PS during the EoR; however, they do tend to underestimate the large-scale power during the earlier epoch of CD by ~7-9 % on average. The absolute error on the PS from small-box realizations goes up to ~7 mK^2, or analogously up to ~4 sigma of the total noise. As expected, this bias and scatter between different realizations decreases with increasing box size. We conclude that box lengths of 250 Mpc are needed to converge at the level of <~1 sigma of the total noise.

Galaxy formation during the first billion years of our Universe remains a challenging problem at the forefront of astrophysical cosmology. Although these $z \geq 6$ galaxies are likely responsible for the last major phase change of our Universe, the epoch of reionization (EoR), detailed studies are possible only for relatively rare, bright objects. Characterizing the fainter galaxies which are more representative of the population as a whole is currently done mainly through their non-ionizing UV luminosity function (LF). Observing the faint end of the UV LFs is nevertheless challenging, and current estimates can differ by orders of magnitude. Here we propose a methodology to combine disparate high-$z$ UV LF data sets in a Bayesian framework: Bayesian Data Averaging (BDA). Using a flexible, physically-motivated galaxy model, we compute the relative evidence of various $z=6$ UV LFs within the magnitude range $-20 \leq M_{\rm UV} \leq -15$ which is common to the data sets. Our model, based primarily on power-law scalings of the halo mass function, naturally penalizes systematically jagged data points as well as mis-estimated errors. We then use the relative evidence to weigh the posteriors obtained from disparate LF observations during the EoR, $6 \leq z \leq 10$. The resulting LFs suggest that the star formation rate density (SFRD) integrated down to a UV magnitude of -17 represent $60.9^{+11.3}_{-9.6}\%$ / $28.2^{+9.3}_{-10.1}\%$ / $5.7^{+4.5}_{-4.7}\%$ of the total SFRD at redshifts 6 / 10 / 15. The BDA framework we introduce enables galaxy models to leverage multiple, analogous observational data sets.

We present a deep learning model trained to emulate the radiative transfer during the epoch of cosmological reionization. CRADLE (Cosmological Reionization And Deep LEarning) is an autoencoder convolutional neural network that uses two-dimensional maps of the star number density and the gas density field at z=6 as inputs and that predicts 3D maps of the times of reionization $\mathrm{t_{reion}}$ as outputs. These predicted single fields are sufficient to describe the global reionization history of the intergalactic medium in a given simulation. We trained the model on a given simulation and tested the predictions on another simulation with the same paramaters but with different initial conditions. The model is successful at predicting $\mathrm{t_{reion}}$ maps that are in good agreement with the test simulation. We used the power spectrum of the $\mathrm{t_{reion}}$ field as an indicator to validate our model. We show that the network predicts large scales almost perfectly but is somewhat less accurate at smaller scales. While the current model is already well-suited to get average estimates about the reionization history, we expect it can be further improved with larger samples for the training, better data pre-processing and finer tuning of hyper-parameters. Emulators of this kind could be systematically used to rapidly obtain the evolving HII regions associated with hydro-only simulations and could be seen as precursors of fully emulated physics solvers for future generations of simulations.

Next generation observatories will enable us to study the first billion years of our Universe in unprecedented detail. Foremost among these are 21-cm interferometry with the HERA and the SKA, and high-$z$ galaxy observations with the James Webb Space Telescope (JWST). Taking a basic galaxy model, in which we allow the star formation rates and ionizing escape fractions to have a power-law dependence on halo mass with an exponential turnover below some threshold, we quantify how observations from these instruments can be used to constrain the astrophysics of high-$z$ galaxies. For this purpose, we generate mock JWST LFs, based on two different hydrodynamical cosmological simulations; these have intrinsic luminosity functions (LFs) which turn over at different scales and yet are fully consistent with present-day observations. We also generate mock 21-cm power spectrum observations, using 1000h observations with SKA1 and a moderate foreground model. Using only JWST data, we predict up to a factor of 2-3 improvement (compared with HST) in the fractional uncertainty of the star formation rate to halo mass relation and the scales at which the LFs peak (i.e. turnover). Most parameters regulating the UV galaxy properties can be constrained at the level of $\sim 10$% or better, if either (i) we are able to better characterize systematic lensing uncertainties than currently possible; or (ii) the intrinsic LFs peak at magnitudes brighter than $M_{\rm UV} \lesssim -13$. Otherwise, improvement over HST-based inference is modest. When combining with upcoming 21-cm observations, we are able to significantly mitigate degeneracies, and constrain all of our astrophysical parameters, even for our most pessimistic assumptions about upcoming JWST LFs. The 21-cm observations also result in an order of magnitude improvement in constraints on the EoR history.

The 21-cm power spectrum (PS) has been shown to be a powerful discriminant of reionization and cosmic dawn astrophysical parameters. However, the 21-cm tomographic signal is highly non-Gaussian. Therefore there is additional information which is wasted if only the PS is used for parameter recovery. Here we showcase astrophysical parameter recovery directly from 21-cm images, using deep learning with convolutional neural networks (CNN). Using a database of 2D images taken from 10,000 21-cm lightcones (each generated from different cosmological initial conditions), we show that a CNN is able to recover parameters describing the first galaxies: (i) Tvir , their minimum host halo virial temperatures (or masses) capable of hosting efficient star formation; (ii) {\zeta} , their typical ionizing efficiencies; (iii) LX/SFR , their typical soft-band X-ray luminosity to star formation rate; and (iv) E0 , the minimum X-ray energy capable of escaping the galaxy into the IGM. For most of their allowed ranges, log Tvir and log LX/SFR are recovered with < 1% uncertainty, while {\zeta} and E0 are recovered within 10% uncertainty. Our results are roughly comparable to the accuracy obtained from Monte Carlo Markov Chain sampling of the PS with 21CMMC for the two mock observations analyzed previously, although we caution that we do not yet include noise and foreground contaminants in this proof-of-concept study.

The properties of the first galaxies, expected to drive the Cosmic Dawn (CD) and the Epoch of Reionization (EoR), are encoded in the 3D structure of the cosmic 21-cm signal. Parameter inference from upcoming 21-cm observations promises to revolutionize our understanding of these unseen galaxies. However, prior inference was done using models with several simplifying assumptions. Here we introduce a flexible, physically-motivated parametrization for high-$z$ galaxy properties, implementing it in the public code 21cmFAST. In particular, we allow their star formation rates and ionizing escape fraction to scale with the masses of their host dark matter halos, and directly compute inhomogeneous, sub-grid recombinations in the intergalactic medium. Combining current Hubble observations of the rest-frame UV luminosity function (UV LFs) at high-$z$ with a mock 1000h 21-cm observation using the Hydrogen Epoch of Reionization Arrays (HERA), we constrain the parameters of our model using a Monte Carlo Markov Chain sampler of 3D simulations, 21CMMC. We show that the amplitude and scaling of the stellar mass with halo mass is strongly constrained by LF observations, while the remaining galaxy properties are constrained mainly by 21-cm observations. The two data sets compliment each other quite well, mitigating degeneracies intrinsic to each observation. All eight of our astrophysical parameters are able to be constrained at the level of $\sim 10\%$ or better. The updated versions of 21cmFAST and 21CMMC used in this work are publicly available.

Photoheating associated with reionization suppressed star formation in low-mass galaxies. Reionization was inhomogeneous, however, affecting different regions at different times. To establish the causal connection between reionization and suppression, we must take this local variation into account. We analyze the results of CoDa (`Cosmic Dawn') I, the first fully-coupled radiation-hydrodynamical simulation of reionization and galaxy formation in the Local Universe, in a volume large enough to model reionization globally but with enough resolving power to follow all atomic-cooling galactic halos in that volume. For every halo identified at a given time, we find the redshift at which the surrounding IGM reionized, along with its instantaneous star formation rate (`SFR') and baryonic gas-to-dark matter ratio ($M_\text{gas}/M_\text{DM}$). The average SFR per halo with $M < 10^9 \text{ M}_\odot$ was steady in regions not yet reionized, but declined sharply following local reionization. For $M > 10^{10} \text{ M}_\odot$, this SFR continued through local reionization, increasing with time, instead. For $10^9 < M < 10^{10} \text{ M}_\odot$, the SFR generally increased modestly through reionization, followed by a modest decline. In general, halo SFRs were higher for regions that reionized earlier. A similar pattern was found for $M_\text{gas}/M_\text{DM}$, which declined sharply following local reionization for $M < 10^9 \text{ M}_\odot$. Local reionization time correlates with local matter overdensity, which determines the local rates of structure formation and ionizing photon consumption. The earliest patches to develop structure and reionize ultimately produced more stars than they needed to finish and maintain their own reionization, exporting their `surplus' starlight to help reionize regions that developed structure later.

Cosmic reionization by starlight from early galaxies affected their evolution, thereby impacting reionization, itself. Star formation suppression, for example, may explain the observed underabundance of Local Group dwarfs relative to N-body predictions for Cold Dark Matter. Reionization modelling requires simulating volumes large enough [~(100Mpc)^3] to sample reionization "patchiness", while resolving millions of galaxy sources above ~10^8 Msun , combining gravitational and gas dynamics with radiative transfer. Modelling the Local Group requires initial cosmological density fluctuations pre-selected to form the well-known structures of the local universe today. Cosmic Dawn ("CoDa") is the first such fully-coupled, radiation-hydrodynamics simulation of reionization of the local universe. Our new hybrid CPU-GPU code, RAMSES-CUDATON, performs hundreds of radiative transfer and ionization rate-solver timesteps on the GPUs for each hydro-gravity timestep on the CPUs. CoDa simulated (91Mpc)^3 with 4096^3 particles and cells, to redshift 4.23, on ORNL supercomputer Titan, utilizing 8192 cores and 8192 GPUs. Global reionization ended slightly later than observed. However, a simple temporal rescaling which brings the evolution of ionized fraction into agreement with observations also reconciles ionizing flux density, cosmic star formation history, CMB electron scattering optical depth and galaxy UV luminosity function with their observed values. Photoionization heating suppressed the star formation of haloes below ~2 x 10^9 Msun , decreasing the abun- dance of faint galaxies around MAB_1600 = [-10,-12]. For most of reionization, star formation was dominated by haloes between 10^10 - 10^11 Msun , so low-mass halo suppression was not reflected by a distinct feature in the global star formation history. (Abridged)

We search for vast planes of satellites (VPoS) in a high resolution simulation of the Local Group performed by the CLUES project, which improves significantly the resolution of former similar studies. We use a simple method for detecting planar configurations of satellites, and validate it on the known plane of M31. We implement a range of prescriptions for modelling the satellite populations, roughly reproducing the variety of recipes used in the literature, and investigate the occurence and properties of planar structures in these populations. The structure of the simulated satellite systems is strongly non-random and contains planes of satellites, predominantly co-rotating, with, in some cases, sizes comparable to the plane observed in M31 by Ibata et al.. However the latter is slightly richer in satellites, slightly thinner and has stronger co-rotation, which makes it stand out as overall more exceptional than the simulated planes, when compared to a random population. Although the simulated planes we find are generally dominated by one real structure, forming its backbone, they are also partly fortuitous and are thus not kinematically coherent structures as a whole. Provided that the simulated and observed planes of satellites are indeed of the same nature, our results suggest that the VPoS of M31 is not a coherent disc and that one third to one half of its satellites must have large proper motions perpendicular to the plane.

We use high resolution simulations of the formation of the local group post-processed by a radiative transfer code for UV photons, to investigate the reionization of the satellite populations of an isolated Milky Way-M31 galaxy pair in a variety of scenarios. We use an improved version of ATON which includes a simple recipe for radiative feedback. In our baseline models, reionization is initiated by low mass, radiatively regulated haloes at high redshift, until more massive haloes appear, which then dominate and complete the reionization process. We investigate the relation between reionization history and present-day positions of the satellite population. We find that the average reionization redshift (zr) of satellites is higher near galaxy centers (MW and M31). This is due to the inside-out reionization patterns imprinted by massive haloes within the progenitor during the EoR, which end up forming the center of the galaxy. Thanks to incomplete dynamical mixing during galaxy assembly, these early patterns survive down to present day, resulting in a a clear radial gradient in the average satellites reionization redshift, up to the virial radius of MW and M31 and beyond. In the lowest emissivity scenario, the outer satellites are reionized about 180 Myr later than the inner satellites. This delay decreases with increasing source model emissivity, or in the case of external reionization by Virgo or M31, because reionization happens faster overall, and becomes spatially quasi-uniform at the highest emissivity.