betaFeedback

WorksPhysics and Astronomy

Anatomy of LyC Escape in MEGATRON: Connecting Feedback Physics to Demographics, Leakage Timescales, and Explorations of ISM Density as a Tracer of Escape Fraction

Abstract

Constraining the Lyman Continuum (LyC) escape fraction (fescf_{\rm esc}) of high-redshift galaxies is essential to connect the abundant populations revealed by JWST to cosmic reionization. We investigate this interplay using {\small MEGATRON}, a suite of zoom simulations with detailed radiation/chemistry and varied feedback prescriptions. We find that feedback physics simultaneously regulates UV source counts, LyC escape fractions, and escaping ionizing luminosities (N˙ion,esc\dot{N}_{\rm ion,esc}), producing different reionization histories. These differences are governed by the delay between a starburst and LyC escape: efficient early feedback and lower-density environments enable photons to escape while pre-supernovae stellar populations remain strongly ionizing. Consequently, models that produce fewer UV-bright galaxies yield earlier LyC escape and reionize sooner than models with more abundant UV sources and delayed LyC emission. Furthermore, UV-faint galaxies follow a common physical sequence in ISM structure with increasing fescf_{\rm esc} irrespective of feedback physics: stronger leakers (1%≤fesc≤50%1\%\leq f_{\rm esc}\leq 50\%) exhibit evidence of strong photoionization and SNe feedback alongside post-starburst star formation histories, while extreme leakers (fesc≥50%f_{\rm esc}\geq 50\%) have also cleared their cold/dense ISM. We find that electron densities inferred from the [O II] λλ3727\rm [O\,{II}]\,λ\lambda3727 doublet are a candidate diagnostic of fescf_{\rm esc}, with lower densities around H II\rm H~{\small II} regions selecting stronger leakers. Moreover, this density can trace the early evolution of fescf_{\rm esc}, whereas UV slope-based calibrations over-predict N˙ion,esc\dot{N}_{\rm ion,esc} for individual young galaxies by up to two orders of magnitude. Overall, the delayed duty cycle of LyC leakage must be accounted for when UV source counts are interpreted in the context of reionization.

Introduction

The Epoch of Reionization describes the period during which neutral hydrogen in the intergalactic medium (IGM) became ionized by the first galaxies. This put an end to the dark ages which had long dominated the Universe for the first billion years (z∼5.3z \sim5.3, Robertson et al. 2015; Kulkarni et al. 2019; Keating et al. 2020; Becker et al. 2021). A variety of probes have been used to constrain this process, including the existence of Gunn-Peterson troughs in quasar spectra (Keating et al. 2020; Becker et al. 2021; Bosman et al. 2022), the damping wings of high-redshift quasars (Davies et al. 2018; Durovčíková et al. 2020), the decrease in number density of Lyα\alpha emitters (LAEs, Stark et al. 2010; Pentericci et al. 2011; Mason et al. 2018, 2019), the opacity of the Lyα\alpha forest of quasars (Fan et al. 2006; Fan, Carilli & Keating 2006) and galaxies (Meyer et al. 2025), and measurements from the CMB (e.g., Heinrich & Hu 2021).

The sources of ionizing Lyman Continuum (LyC) photons remain largely debated, with various models appealing to differing contributions from UV-bright sources (e.g., Naidu et al. 2020a) and abundant UV-faint galaxies (e.g., Finkelstein et al. [48]). In either case, typical models for reionization driven by star-forming galaxies rely on three quantities, which together provide an estimate of the total ionizing photon production rate, given by n˙ion=ρUVξionfesc\dot{n}_{\mathrm{ion}} = \rho_{\mathrm{UV}}\xi_{\mathrm{ion}}f_{\mathrm{esc}} (see review by Robertson 2022). These include the redshift-dependent total UV luminosity density (ρUV\rho_{\mathrm{UV}}), the ionizing photon production efficiency (ξion\xi_{\mathrm{ion}}), giving the number of ionizing photons produced by a stellar population per UV luminosity, and the LyC escape fraction (fescf_{\mathrm{esc}}), reporting the number of these photons which escape their host galaxy. Another approach is to combine these quantities and to instead try to constrain the ‘escaping ionizing luminosity’ (N˙ion,esc\dot{N}_{\mathrm{ion,esc}}) of individual galaxies (see discussion by Choustikov et al. 2025). Once estimated for each source and aggregated across a survey volume, the LyC production rate can be estimated by computing an ionizing emissivity function, ϕion(N˙ion,esc,z)\phi_{\mathrm{ion}}(\dot{N}_{\mathrm{ion,esc}}, z), which can then be integrated to compute n˙ion\dot{n}_{\mathrm{ion}} (e.g., Giovinazzo et al. [54]; Summerfield et al. [198]).

Since the launch of the James Webb Space Telescope (JWST), much effort has gone to placing significant constraints on each of these quantities. In particular, many studies have tried to constrain the high-redshift UV luminosity function (Naidu et al. [135]; Austin et al. [6]; Donnan et al. [36, 37]; Castellano et al. [20, 19]; Chemerynska et al. [22, 21]; Finkelstein et al. [49]; Adams et al. [2]; Whitler et al. [218]; Atek et al. [4]; McLeod et al. [128]), all finding that there is a larger than expected abundance of UV-bright galaxies at z≳10z \gtrsim10. Similarly, the ionizing photon production efficiency of high-redshift galaxies has also been constrained, finding that typical high-redshift galaxies produce a large number of ionizing photons (Endsley et al. (2021); Stefanon et al. (2022); Tang et al. [200]; Simmonds et al. [183, 181]; Harshan et al. [61]; Saxena et al. [169]; Mascia et al. [122]; Pahl et al. [145]; Hayes et al. (2025)), which have since been mildly revised when large mass-complete samples are used (e.g., Simmonds et al. [182]; Austin et al. [7]). In contrast, the LyC escape fraction is impossible to measure directly in the Epoch of Reionization due to the neutral IGM (Inoue et al. 2014). Nevertheless, combining measurements of ρUV\rho_{\mathrm{UV}} and ξion\xi_{\mathrm{ion}} with reasonable expectations for fescf_{\mathrm{esc}} of ∼10%\sim10\% (e.g., Robertson et al. (2015); Rosdahl et al. [162]; Jecmen et al. [75]) or extrapolating fescf_{\mathrm{esc}} models based on low-redshift analogues (e.g., Chisholm et al. [25]) suggest that there may be too many ionizing photons: causing the Universe to reionize far earlier than expected based on historic measurements (Muñoz et al. 2024, though see discussion by Papovich et al. [148]).

A variety of models have since been proposed to explain the observed UV luminosity functions and slow decline in the redshift-evolution of the UV luminosity density beyond redshift z∼10z \sim10 (see review by Stark et al. 2026). Popular theories include an increase in the halo-scale star-formation efficiency due to weaker stellar feedback (Dekel et al. 2023; Qin, Balu & Wyithe 2023; Somerville et al. 2025, c.f., Ferrara, Manzoni & Ntor- mousi 2025), a decrease in the presence of dust (e.g., Ferrara, Pallottini & Dayal 2023), allowing burstier star for- mation histories (SFHs) to increase scatter in the LUVL_{\mathrm{UV}}–MhaloM_{\mathrm{halo}} relation (Mason, Trenti & Treu 2023; Shen et al. 2023; Sun et al. 2023; Kravtsov & Belokurov 2024), modifying the star formation rate–UV luminosity relation by including the nebular continuum (Katz et al. [83]), including a top-heavy IMF (e.g., Yung et al. [226]; Cueto et al. (2024)), modifying the physical properties of stars (Katz et al. 2024), and including contributions from AGN (e.g., Hegde, Wyatt & Furlanetto 2024).

Due to the fact that it can not be directly detected at high-redshift, the general approach to understand the leakage of ionizing radiation in high-redshift galaxies is to study low-redshift analogue systems (e.g., Flury et al. [50]) and simulations (see review by Jaskot 2025). Once identified, analogue leakers can be used to search for indirect diagnostics: quantities which may be shown to correlate with fescf_{\mathrm{esc}}. A huge number of potential diagnostics have been proposed over the last decade, including high [O iii] λ5007\lambda5007/[O ii] λ3727\lambda3727 (Jaskot & Oey 2013; Nakajima & Ouchi 2014), Lyα\alpha peak separations (Verhamme et al. 2017; Izotov et al. 2020), Lyα\alpha equivalent width (Steidel et al. 2018; Pahl et al. 2021), Lyα\alpha escape fractions (Dijkstra, Gronke & Venkatesan 2016; Verhamme et al. 2017), Mg II λλ2796,2804\lambda\lambda2796, 2804 doublet ratios (Chisholm et al. (2020); Katz et al. [85]), [S II] deficits (Wang et al. [216]), strong C IV λλ1488,1486\lambda\lambda1488, 1486 emission (Schaerer et al. [177]; Saxena et al. [172]), UV slope (Chisholm et al. [25]), Lyα\alpha–to–Hα\alpha surface brightness profiles (Choustikov et al. [28]), and the Lyα\alpha halo fraction (Saldana-Lopez et al. [166]). Furthermore, multivariate approaches have also been explored (e.g. Maji et al. [117]; Mascia et al. [123]; Choustikov et al. [27]; Jaskot et al. [73]), though there is still considerable disagreement. Finally, machine-learning approaches trained on the SPHINX20^{20} (Rosdahl et al. [162]; Katz et al. [90]) and THESAN-ZOOM simulations (Kannan et al. [81]) to directly predict N˙ion,esc\dot{N}_{\mathrm{ion,esc}} based on JWST photometry has also been proposed by Choustikov et al. 2025 and Summerfield et al. [198]. These efforts are complimentary to ongoing attempts to constrain LyC leakage with traditional SED-fitting codes (e.g., Giovinazzo et al. [54]; Stoffers et al. (2026); Papovich et al. [148], see review by Iyer et al. 2026).

On the theoretical side, considerable effort has gone into simulating the process of LyC escape in both extremely high-resolution GMC-scale simulations (e.g., Kimm et al. 2019, 2022; Kakichi & Gronke 2021; Menon et al. 2025) and cosmological radiation hydrodynamics simulations (e.g., Gnedin, Kravtsov & Chen 2008; Wise & Cen 2009; Kimm & Cen 2014; Paardekooper, Khochfar & Dalla Vecchia 2015; Xu et al. 2016; Trebitsch et al. 2017a, 2018; Rosdahl et al. 2018, 2022; Ma et al. 2020; Katz et al. 2020, 2022a, 2023b, 2026b; Kannan et al. 2022, 2025; Yeh et al. 2023; Bhagwat et al. 2024a; Choustikov et al. 2024a,b; Yuan et al. 2024; Giovinazzo et al. 2024; Farcy et al. 2025; Zier et al. 2026). Here, the difficulty lies in the need to simultaneously resolve the disruption of stellar birth clouds, the multi-phase nature of the interstellar medium (ISM), and the cosmological context of the host galaxy. Of particular relevance here, many potential solutions to the bright galaxy problem rely on modifications to feedback physics – something which fundamentally modifies the ISM structure in and observable properties of high-redshift galaxies [26].

In this work, we aim to address these challenges by investigating how changes to star-formation/feedback physics inspired by solutions to the bright galaxy problem influence the production and escape of ionizing radiation. To do this, we make use of the suite of high-resolution MEGATRON simulations [88, 16, 156]. These are cosmological radiation hydrodynamics zoom-in simulations of a Milky Way-mass progenitor equipped with a state-of-the-art galaxy formation model. In particular, MEGATRON features a detailed ISM model coupled through multigroup radiation transport to a 81-species non-equilibrium thermochemical model of primordial species, metals, and molecules – producing a dynamic and realistic multi-phase ISM for which spectroscopic properties are direct predictions.

This paper is arranged as follows. In § 2, we describe the MEGATRON simulations as well as our predictions for LyC escape fractions and emission lines. We discuss the demographics of galaxies that contribute to reionization in § 3. The ISM structure of LyC leakers is investigated in § 4, based on which we propose ISM tracers such as electron density as potential diagnostics for the LyC escape fraction in § 5. Finally, we conclude in § 6.

Methods

The MEGATRON Simulations

For this work, we make use of results from the high-redshift suite of MEGATRON simulations, described in Katz et al. [88]; Cadiou et al. [16]; Rey et al. [156]. These are high-resolution (∼5\sim5 pc at z∼8.5z \sim8.5) cosmological simulations of a Milky Way progenitor halo simulated using RAMSES-RTZ1 (Katz 2022). These simulations include a full non-equilibrium chemistry network including primordial species (H I-II, He I-III, and e−e^{-}), metals (C I-VI, N I-VII, O I-VIII, Ne I-X, Mg I-X, Si I-XI, S I-XI, and Fe I-XI), and molecules (H2_2 and CO) coupled using on-the-fly multi-frequency (8 groups) radiation transport to a detailed model for ISM heating and cooling processes Katz et al. [87]. As a result, MEGATRON captures the non-equilibrium state of the multi-phase ISM, which has been shown to reproduce the observed spectroscopic properties of high-redshift galaxies (Katz et al. 2026a,b; Choustikov et al. 2026, in prep.), the mass-metallicity relation of local ultra-faint dwarf galaxies Rey et al. [156], FIR scaling relations (Matsumoto et al., in prep.), properties of extreme emission line galaxies (Quilt et al., in prep.), as well as making predictions for the observability of Population III stars (Storck et al. 2026), LyC-leaking galaxies (Xing et al., in prep.), and quenched dwarf galaxies during the Epoch of Reionization (Attard et al. 2026).

The high-redshift MEGATRON simulation suite includes four re-simulations of the same initial conditions, equipped with different models for star formation and feedback. These are each designed to capture a different potential solution to the bright galaxy problem and include:

  • Efficient Star Formation: This ‘fiducial’ simulation has the weakest feedback among the four models, leading to the highest global star formation efficiency. Stellar SEDs follow BPASS (Eldridge et al. 2017; Stanway & Eldridge 2018) assuming a Kroupa IMF with a maximum mass of 300 M⊙_{\odot}. This model produces smooth SFHs and galaxies with a very dense ISM along with the widest range of ionization parameters.

  • Bursty Star Formation: This model increases the energy per supernova by a constant factor of five compared to the fiducial run. Galaxy stellar masses are better regulated and the SFHs become significantly burstier. Stellar SEDs follow BPASS (Eldridge et al. 2017; Stanway & Eldridge 2018) assuming a Kroupa IMF with a maximum mass of 300 M⊙_{\odot}.

  • Variable IMF: This model uses a variable IMF following Marks et al. (2012) with a fixed maximum mass of 120 M⊙_{\odot}. As a result, the upper mass slope becomes shallower at higher gas densities and lower metallicities, leading to a more top-heavy IMF. Some high mass stars explode as hypernovae, releasing ∼10×\sim10 \times the energy with commensurate yields (Nomoto et al. 2006). Stellar SEDs follow Starburst99 Leitherer et al. [109].

  • High ϵff\epsilon_{\mathrm{ff}}, HN: Stars form in lower-density gas with a local efficiency per free-fall time of 100%. A subset of massive stars also explode as hypernovae, as for the variable IMF run. Stellar SEDs follow BPASS (Eldridge et al. 2017; Stanway & Eldridge 2018) assuming a Kroupa IMF with a maximum mass of 300 M⊙_{\odot}.

As each model includes different versions of stellar feedback, the typical ISM of a galaxy in each simulation will also be different – resulting in different emission line ratios and observed properties driven by changes to the density and temperature structure of the ISM (see discussion in Choustikov et al. [26]). We make use of these re-simulations to understand the impact of varying star formation/feedback physics on the process and history of reionization. Angle-averaged monochromatic (900Å–912Å) escape fractions are measured at 0.75 RvirR_{\mathrm{vir}} using RASCAS (Michel-Dansac et al. 2020) as described in Choustikov et al. [26].

Direct Predictions for Synthetic Observations

Across the four simulations, we predict a large library of spectra (∼38,000\sim38,000 with MUV≤−13M_{\mathrm{UV}} \leq-13 at redshift z≳8.5z \gtrsim8.5, see also Figure 4 of Katz et al. [88]), each accounting for the stellar continuum, nebular continuum and nebular emission lines. For the purpose of this work, we predict the full intrinsic UV-optical spectrum of each galaxy, alongside fluxes for a variety of specific optical lines (e.g., Hβ\beta, [O II] λ3726\lambda3726, 3728, [O III] λ4363\lambda4363, and [O III] $\lambda5007). To do so, we make use of the fact that RAMSES-RTZ tracks the number density of all relevant ionic species (e.g., H+^{+}, O+^{+}, O2+^{2+}) as well as the electron number density (accounting for ionized metal species) on-the-fly during the simulation. We use PyNeb (Luridiana, Morisset & Shaw 2015) to compute the emissivity of each UV/optical emission line on a cell-by-cell basis given the temperature and electron number density2^{2} (nen_{\mathrm{e}}) of the cell, alongside atomic data from Version 10 of the CHIANTI database (Dere et al. 1997, 2019).

Figure 4. Escape fraction, ionising production efficiency, and escaping ionising luminosity population demographics. We show distributions for ξion\xi_{\mathrm{ion}} (both the intrinsic value alongside that inferred from Hα\alpha), fescf_{\mathrm{esc}}, and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} for all galaxies with MUV≤−13M_{\mathrm{UV}} \leq-13 across all four MEGATRON simulations. Changes to star formation/feedback physics modify escape fraction distributions while leaving ξion(Hα)\xi_{\mathrm{ion}}(\mathrm{H}\alpha) largely unchanged. However, the mapping to N˙ion,esc\dot{N}_{\mathrm{ion,esc}} is non-trivial. For example, while the bursty star formation model produces the largest number density of strong leakers (fesc≳30%f_{\mathrm{esc}} \gtrsim30\%), the ionizing output of these galaxies is systematically lower than those of the high ϵff\epsilon_{\mathrm{ff}}, HN simulation.

For specific measurements, all gas cells within 25% of the virial radius of a particular halo are considered3^{3}. This is done in order to avoid the impact of satellite galaxies, though it is important to note that more faithful predictions would require the use of a realistic aperture/slit to account for losses (e.g., Bhagwat et al. 2024b). Finally, we utilise a subgrid post-processing correction for unresolved Strömgren spheres by using a grid of spherical idealised2 CLOUDY photoionization models (Ferland et al. 2017). The UV slope (β\beta) is measured by fitting the continuum with a power law of the form fλ∝λβf_\lambda\propto\lambda^\beta within the wavelength range 1350Å – 2700Å. Following the approach of Saxena et al. (2026), we include the nebular continuum but exclude nebular emission lines.

Measuring ISM Densities

In order to be consistent with observational approaches, we also use a variety of strong line ratios to measure the properties of the ISM (see review by Kewley, Nicholls & Sutherland 2019). In particular, we focus on the density-sensitive emission line doublet ratio of O2 = [O II] λ3726/[O II] λ3728 (see Figure 1 of Kewley et al. 2019). To measure the densities, we again make use of PyNeb alongside the same atomic data described above. As the combination of temperature-sensitive and density-sensitive lines is fairly uncommon at high redshift (though see discussions in Topping et al. 2024; Arellano-Córdova et al. 2026), we follow the common procedure of assuming a fixed temperature when measuring densities. Specifically, we adopt a fiducial temperature of Te=1.5×104 KT_{\mathrm{e}} = 1.5 \times10^{4}\ \mathrm{K}.5

IGM-scale view of the efficient star formation MEGATRON simulation at redshift $z \sim 11$

Figure 1. An IGM-scale view of the efficient star formation MEGATRON simulation at redshift z∼11z \sim11. A variety of projections through the MEGATRON simulation are shown. These include intrinsic properties such as the dark matter density field (light purple), gas density (dark purple), temperature (orange), and LyC flux (white), as well as direct predictions of Hβ\beta (blue) and [O III] λ5007\lambda5007 (red) emission lines. A composite of these features is also shown in the top panel, inheriting the same colours.

Throughout the present work, we frequently make use of two different density measurements, defined as follows. The ‘intrinsic’ gas density (⟨nH⟩[OII]\langle n_{\mathrm{H}}\rangle_{\mathrm{[OII]}}) is the average hydrogen density weighed by the luminosity of [O II] λλ3727 as computed within the ISM of each galaxy. In contrast, the ‘observed’ electron density (ne([OII])n_{\mathrm{e}}(\mathrm{[OII]})) is the inferred density as measured from the [O II] doublet as described above. While the former is not directly observable, it is a useful feature to consider in a theoretical context as it is not affected by any observational bias. A comparison between these two quantities is provided in Appendix § 9.

Ionizing Photon Production Efficiency

To compute ionizing photon production efficiencies (ξion\xi_{\mathrm{ion}}), we follow the typical approach of assuming Case B recombination with Te=1.5×104 KT_{\mathrm{e}} = 1.5 \times10^{4}\ \mathrm{K}, ne=100 cm−3n_{\mathrm{e}} = 100\ \mathrm{cm}^{-3}, fesc=0f_{\mathrm{esc}} = 0 so that (Osterbrock & Ferland 2006):

ξion(Hα) / [Hz/erg]=7.7×1011L(Hα)L1500,int(1)\xi_{\mathrm{ion}}(\mathrm{H}\alpha)\ /\ [\mathrm{Hz}/\mathrm{erg}] = \frac{7.7 \times10^{11} L(\mathrm{H}\alpha)}{L_{1500,\mathrm{int}}} \tag*{(1)}

where L(Hα)L(\mathrm{H}\alpha) is the flux of the Hα line and L1500,intL_{1500,\mathrm{int}} is the intrinsic UV luminosity density (in the LνL_\nu system, Maseda et al. 2020). This efficiency should in principle be corrected for the escape fraction of LyC radiation6. However, to best compare with observational inferences, we ignore this effect (see discussion in Saxena et al. 2024b). The intrinsic ionizing photon production efficiency, ξionint\xi_{\mathrm{ion}}^{\mathrm{int}}, is instead computed using the intrinsic ionizing luminosity of all stars within the central 0.25Rvir0.25R_{\mathrm{vir}} of the halo in question.

The Sources of Reionization in MEGATRON

Our goal is to understand the sources and history of reionization in the high-redshift MEGATRON simulations. Figure 1 shows a compilation of projections through the cosmic web and IGM of the efficient star formation simulation at redshift z=11z = 11, showing the appearance of an ionized bubble centred on the dark matter overdensity, driven by a large number of sources. In particular, we show the dark matter density, gas density, mass-weighted temperature and LyC photon flux, as well as predictions for Hβ and [O III] λ5007 luminosities. At this stage, the reionization process has only just begun (xHII≲0.1x_{\mathrm{HII}} \lesssim0.1, see below). Furthermore, metal enrichment has not significantly polluted the IGM and oxygen emission is confined to star-forming galaxies. Finally, while some strong LyC sources appear co-spatial with strong emission lines, there are also some very strong sources that do not, consistent with a scenario where all line-emitting gas has been evacuated.

The evolution of the mass- and volume-weighted IGM ionized fraction, xHIIx_{\mathrm{HII}} as computed within the entire high-resolution zoom region is shown in the top panel of Figure 2 (see also Appendix § 11 of Rey et al. 2026). We find that the four different star formation/feedback prescriptions lead to dramatically different reionization histories. For example, this process is almost entirely complete in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation by redshift z∼8.5z \sim8.5, while only a third of the efficient star formation simulation volume has been ionized at this time. These differences are interesting when contrasted against the distribution of source UV magnitudes for each simulation at redshifts z≲16z \lesssim16 (bottom panel of Figure 2). Here, we find that the efficient star formation simulation features by far the most (by an order of magnitude) UV-bright (MUV≤−18M_{\mathrm{UV}} \leq-18) sources, in comparison to the high ϵff\epsilon_{\mathrm{ff}}, HN simulation which features the smallest number of galaxies at nearly any MUVM_{\mathrm{UV}}. We compare each simulation to observed UV luminosity functions (UVLFs) from Whitler et al. (2025) and Atek et al. (2026), each normalised7 to the number of galaxies in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation at MUV=−19M_{\mathrm{UV}} = -19. Interestingly, the efficient star formation and high ϵff\epsilon_{\mathrm{ff}}, HN simulations both prefer the shape of the flatter faint-end slope of Atek et al. (2026). In contrast, the bursty star formation simulation prefers the steeper slope of Whitler et al. (2025), while the variable IMF simulation falls somewhere in the middle. These findings align with previous simulations have shown that modifications to feedback physics such as stellar SEDs (Rosasdahl et al. 2018), IMF (Cueto et al. 2024), stellar feedback strength/timing (Bhagwat et al. 2024a), and cosmic rays (Farcy et al. 2025, see also discussion by Martin-Alvarez, Wechsler & Sijacki 2026) can jointly modify predicted UVLFs alongside the process of reionization. Nevertheless, it is instructive to understand the physical reasons for these differences, especially given the naive expectation that a greater number of UV sources would lead to a faster reionization history.

Figure 2. Reionization history and UV number counts. Redshift evolution of the IGM ionized fraction (xHIIx_{\mathrm{HII}}, top) till the end of the simulation at z∼8.5z \sim8.5 as well as the measured distribution of UV magnitudes (bottom) among snapshots below redshift z≲16z \lesssim16 for each MEGATRON simulation. We include observed UVLFs renormalised to the high ϵff\epsilon_{\mathrm{ff}}, HN simulation at MUV=−19M_{\mathrm{UV}} = -19 for comparison. While the high ϵff\epsilon_{\mathrm{ff}}, HN simulation features the fewest galaxies, it produces the fastest reionization history. The opposite is true for the efficient star formation simulation.

Demographics of LyC Leakers in MEGATRON

Due to the fact that the process of reionization depends on the ionizing emissivity, an accelerated reionization history can canonically be attained by either increasing the number of sources, the ionizing photon production efficiency, or the escape fraction. In reality however, these three parameters are intrinsically linked by the processes of star formation and feedback – something that we now

aim to disentangle. As the MEGATRON simulations follow a zoomed-in unrepresentative region of the Universe, we can not reliably report on traditional volume-averaged quantities (see discussions in Lovell et al. 2021, 2026). Instead, we leverage the fact that each simulation uses identical initial conditions and therefore all feature the same controlled volume – allowing us to make meaningful benchmark comparisons.

To understand the demographics of LyC sources, Figure 3 shows cumulative distribution functions (CDFs) of the escaping ionising luminosity (N˙ion,esc\dot{N}_{\mathrm{ion,esc}}) for each simulation as a function of UV magnitude (top) and halo mass (bottom). Magnitudes and masses corresponding to 50% of the contributed LyC photons are denoted by dashed vertical lines. We find that the LyC budget in the variable IMF simulation is dominated by UV-bright galaxies (due to the strong MUV−MvirM_{\mathrm{UV}} - M_{\mathrm{vir}} relation, see Appendix § 10), in contrast to the high ϵff\epsilon_{\mathrm{ff}}, HN simulation where ∼40%\sim40\% of LyC photons are produced by galaxies with MUV≤−16M_{\mathrm{UV}} \le-16 (see discussion below). In contrast, we find that the relative contributions of LyC-contributing DM halos are less sensitive to feedback, owing to the fact that the median N˙ion,esc−Mvir\dot{N}_{\mathrm{ion,esc}} - M_{\mathrm{vir}} appears comparably robust (albeit with more scatter than in other relations, see also Appendix § 10 for a discussion of scaling relations). This conclusion brings a novel angle to the ongoing debate of whether reionization is a democratic process or driven by a small number of massive, UV-bright halos (e.g., [48] (Naidu et al. 2020b)), by highlighting that the MUV−MvirM_{\mathrm{UV}} - M_{\mathrm{vir}} relation is linked to the physics of LyC escape.

Figure 3. Galactic contributions to Reionization. Cumulative fractions of the total escaping LyC photons released by galaxies below redshift z≤14z \le14 as a function of UV magnitude (top) and halo virial mass (bottom) for each MEGATRON simulation. Magnitudes and masses corresponding to 50% of the contributed LyC photons are denoted by dashed vertical lines. Different feedback models lead to changes in the relative importance of UV bright/faint galaxies, while leaving the dependence on halo mass less impacted.

Figure 4 shows the distributions of intrinsic and Hα\alpha-inferred ξion\xi_{\mathrm{ion}}, fescf_{\mathrm{esc}}, and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} for each simulation, considering all galaxies with MUV≤−13M_{\mathrm{UV}} \leq-13. It is important to note that while Population III-dominated galaxies are included in this selection, these are unlikely to dominate due to the low UV luminosity of such systems (see discussion in Storck et al. 2026). We find that all four simulations produce similar ionizing photon production efficiencies (up to differences in SEDs used and changes to the density structure of the ISM, see also Katz et al. [83] for discussions of impacts due to the nebular continuum). However, there is a slight bias in that the strongest intrinsic producing galaxies (ξion≳1026\xi_{\mathrm{ion}} \gtrsim10^{26} Hz/erg) appear to be comparatively less common in the variable IMF and high ϵff\epsilon_{\mathrm{ff}}, HN simulations when inferred by Balmer emission, as compared to the frequency of such sources in the efficient star formation and bursty star formation simulations. This is attributed to differences in the nebular continuum driven by whether or not the ISM has experienced feedback from the SNae as well as the impact of a non-zero ionizing escape fraction. Furthermore, we find markedly different distributions in fescf_{\mathrm{esc}} as well as N˙ion,esc\dot{N}_{\mathrm{ion,esc}}. This demonstrates that the implementation of feedback has a significant effect on the creation of under-dense ionized channels through which LyC photons can escape (see also discussions in (Kimm et al. 2022; Bhagwat et al. 2024a; Farcy et al. 2025)). In particular, we find that the bursty star formation simulation produces the largest number density of galaxies with fesc≳10%f_{\mathrm{esc}} \gtrsim10\%, followed by the high ϵff\epsilon_{\mathrm{ff}}, HN and efficient star formation simulations. Interestingly, the variable IMF simulation features by far the fewest number density of galaxies with fesc≳10%f_{\mathrm{esc}} \gtrsim10\%.

To synthesize these findings, Figure 5 shows the median N˙ion,esc\dot{N}_{\mathrm{ion,esc}} as a function of MUVM_{\mathrm{UV}} for galaxies below redshift z≤16z \leq16 across all four MEGATRON simulations. We find that both the normalisation and scatter (shaded regions describe the 16th--84th percentiles) are sensitive to feedback physics. In particular, while the bursty star formation and variable IMF simulations follow a common trend (which agrees with findings from THESAN-ZOOM, Summerfield et al. [198]), there is a marked step change in behaviour around MUV∼−16M_{\mathrm{UV}} \sim-16 in the other two simulations. Namely, galaxies fainter than this limit in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation contribute ∼1\sim1 dex more ionizing radiation, while galaxies brighter than this in the efficient star formation simulation release ∼2\sim2 dex fewer ionizing photons than comparable systems – mirroring the opposite of the observed behaviour in the distribution of UV sources for each simulation (Figure 2).

Figure 5. Ionizing Luminosity as a function of Magnitude. We show running medians for the escaping ionizing luminosity of galaxies as a function of MUVM_{\mathrm{UV}} for galaxies below redshift z≤16z \leq16 across all four MEGATRON simulations. Both the normalisation and scatter (shown by shaded areas which mark the 16th and 84th percentiles) in this relation are sensitive to feedback physics and can be conditional on MUVM_{\mathrm{UV}}. Among very faint galaxies (MUV∼−10M_{\mathrm{UV}} \sim-10), there is a boost due to the ionizing luminosity of Population III stars. As a comparison, we show the running median from the THESAN-ZOOM simulation (Summerfield et al. [198]).

Nevertheless, galaxies brighter than MUV≤−17M_{\mathrm{UV}} \leq-17 in the efficient star formation simulation still account for 50% of the total ionizing background due to their increased abundance. Interestingly, there is also a noticeable up-scatter in N˙ion,esc\dot{N}_{\mathrm{ion,esc}} at MUV∼−10M_{\mathrm{UV}} \sim-10 due to the presence of Population III stars, which release more ionizing photons per stellar mass (Schaerer 2002; Larkin, Gerasimov & Burgasser 2023).

Finally, it is crucial to highlight that the distribution of LyC escape fractions does not translate to effective contributions to reionization. This is due to the existence of remnant leakers (galaxies with large escape fractions but no LyC-producing stars, Katz et al. [91]). Similarly, models which are capable of effectively quenching dwarf galaxies by expelling all gas will produce more remnant leakers, stopping such galaxies from contributing to reionization. In particular, Attard et al. (2026) recently showed that the bursty star formation and high ϵff\epsilon_{\mathrm{ff}}, HN simulations produce significantly larger fractions of quenched galaxies. As a result, while the bursty star formation simulation produces significantly more galaxies with fesc>10%f_{\mathrm{esc}} > 10\% than the variable IMF simulation, both simulations produce the same number density of galaxies with N˙ion,esc≳1051 s−1\dot{N}_{\mathrm{ion,esc}} \gtrsim10^{51}\ \mathrm{s}^{-1}. Ultimately, these differences are driven by the timescale synchronicity of LyC production and leakage.

The Importance of Timing for Effective Leakage of LyC Photons

The question of LyC escape from galaxies is one of timing. Immediately following a burst of star formation, the race is on to clear channels in the ISM before the sources of ionising radiation disappear. In the extreme case, we can understand this by considering a stellar population consisting of a single O star. Without early feedback8, the first chance that LyC photons will have to escape will come after the star goes supernova – at which point there will necessarily be no more photons to escape. In this scenario, there will be a step change at ∼3 Myr\sim3\ \mathrm{Myr}, at which point the escape fractions increase significantly while the ionising luminosity reaches N˙ion,esc=0\dot{N}_{\mathrm{ion,esc}} = 0. If instead we consider early feedback (e.g., winds, radiation pressure, photoionization) then this transition will be smoothed – allowing the O star to contribute. In reality, we must contend with multiple concurrent sources of both feedback and LyC photons, as well as the necessity to clear channels beyond the birth cloud.

To understand the timescales for effective leakage produced in a realistic context, Figure 6 shows running medians of fescf_{\mathrm{esc}}, the intrinsic ionising luminosity (N˙ion,int\dot{N}_{\mathrm{ion,int}}) and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} for galaxies with −13≥MUV≥−17-13 \geq M_{\mathrm{UV}} \geq-17 (dashed) and MUV≤−17M_{\mathrm{UV}} \leq-17 (solid) in each simulation as a function of ionising luminosity-weighted mean stellar age9. We find that simulations that feature hypernovae begin to leak more quickly (Nomoto et al. 2006) and sustain a higher level of leakage for all subsequent times. In particular, the high ϵff\epsilon_{\mathrm{ff}}, HN simulation features the shortest

Figure 6. Time-evolution of LyC production and escape. We show running medians of fescf_{\mathrm{esc}}, N˙ion,int\dot{N}_{\mathrm{ion,int}}, and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} as a function of the ionizing luminosity-weighted mean stellar age for galaxies in each MEGATRON simulation with MUV≤−17M_{\mathrm{UV}} \leq-17 (solid) and −17≤MUV≤−13-17 \leq M_{\mathrm{UV}} \leq-13 (dashed). Changes to feedback physics (e.g., burstier star formation, stronger early feedback, and preferentially forming stars in low-density gas) all act to shorten the time delay between a particular starburst and the onset of effective leakage. As the intrinsic ionizing luminosity decreases with time, a shorter gap will lead to more effective leakage. To highlight this, we include a dashed vertical line at ∼3.5 Myr\sim3.5\ \mathrm{Myr}, at which point the first Type II SNe begin. Interestingly, the UV-bright population is capable of releasing similar numbers of ionizing photons, while the output of UV-faint galaxies is strongly dependent on feedback.

time delay between burst and leakage. This is driven by the fact that stars form in lower-density gas in this simulation, leading to more effective coupling of radiation and SNe feedback to the ISM (see discussions in Blondin et al. 1998; Thornton et al. 1998; Kannan et al. 2020). Following this, the median galaxy in the bursty star formation simulation also shows boosted leakage at earlier times, driven by the boosted SNe energy (Katz et al. [88]). The median efficient star formation simulation galaxy only reaches the highest escape fractions after ∼10\sim10 Myr. As highlighted by Katz et al. [91] and Choustikov et al. [27], high levels of leakage are inconsequential if stellar populations are not producing any ionising photons. Overall, this means that burstier star formation, stronger early feedback, modifications to stellar physics which lead to star formation in lower density environments or the inclusion of hypernovae will all lead to earlier leakage as well as higher escape fractions at all times. Finally, it is interesting to note that while UV-bright (MUV≤−17M_{\rm UV} \leq-17) galaxies in the bursty star formation and variable IMF simulations begin leaking earlier than UV-faint (−17≤MUV≤−13-17 \leq M_{\rm UV} \leq-13) systems, owing to the high gas densities of UV-faint systems undergoing their first major burst of star formation. In contrasts, the trends discussed above are unchanged with luminosity for the efficient star formation and high ϵff\epsilon_{\rm ff}, HN models, as they either all have very high gas densities, or low gas densities respectively.

In the central panel of Figure 6, we can see that on average, the ionising luminosity of galaxies in all simulations decays as a function of their age, especially after ∼3\sim3 Myr at which point the first O stars start to die (Yusof et al. [227]; Groh et al. 2014; Chisholm et al. 2019). However, it is important to note that we find two key patterns of behaviour. First, galaxies in the variable IMF simulation tend to produce fewer ionising photons at later ages. This is due to the fact that this simulation uses Starburst99 SEDs (Leitherer et al. [109]) (as opposed to the others, which use BPASS, Eldridge et al. 2017; Stanway & Eldridge 2018), which do not consider binary stars and limit the maximum mass to 120 M⊙M_{\odot}. While a top-heavy IMF will help to bridge this gap by increasing the fraction of massive stars in young stellar populations, binarity acts to increase the magnitude and duration of LyC production (Stanway, Eldridge & Becker 2016; Götberg, de Mink & Groh 2017; Götberg et al. 2020) and are known to have a significant impact on simulations of reionization (Ma et al. 2016; Rosdahl et al. [163]). Furthermore, very massive stars can also be an important factor by significantly boosting the ionizing output and early feedback10 of a given stellar population (e.g., Schaefer et al. 2025). Next, we find that UV-faint galaxies in the high ϵff\epsilon_{\rm ff}, HN simulation also produces systematically fewer ionising photons at early times, owing to the inclusion of hypernovae physics, allowing stars with masses beyond 25 M⊙M_{\odot} to explode much earlier (see also Kobayashi, Springel & White 2007). In contrast, UV-faint galaxies in the efficient star formation simulation tend to produce more ionising photons at all ages. This is due to the fact that smoother SFHs

(see e.g., Figure 5 of Katz et al. [88]) lead to multiple concurrent starbursts with more continuous LyC production. At this stage, it is relevant to note that an ionising luminosity-weighted stellar age is not a strict measure of time since starburst in galaxies with extended SFHs (see also discussion in Mehrotra et al. [129]). To this end, we shall discuss an alternate probe below.

Finally, the bottom panel of Figure 6 shows the median N˙ion,esc\dot{N}_{\rm ion,esc} as a function of ionising luminosity-weighted stellar age in each simulation. Here, we find that changes to feedback physics affect galaxies differently as a function of their UV magnitude. Among the UV-bright systems (MUV≤−17M_{\rm UV} \leq-17), all four models converge to similar escaping ionizing luminosities (N˙ion,esc∼1052 s−1\dot{N}_{\rm ion,esc} \sim10^{52}\ {\rm s}^{-1}), though after very different time delays. Both simulations featuring HN e reach this stage after ∼2.5\sim2.5 Myr, while the bursty star formation and efficient star formation simulations begin leaking at this rate after ∼4\sim4 Myr and ∼11\sim11 Myr respectively. In contrast, UV-faint (−17≤MUV≤−13-17 \leq M_{\mathrm{UV}} \leq-13) galaxies in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation release more ionising photons at all times, in particular ∼3.5\sim3.5 Myr into the bursty star formation cycle, as indicated by the vertical dotted line. As a direct consequence, this simulation begins to reionize the IGM significantly earlier than the others, despite the fact that it features fewer galaxies at a given MUVM_{\mathrm{UV}} (see Figure 2). Next, we find that typical UV-faint galaxies in the variable IMF and bursty star formation simulations follow very similar trends, though the former has slightly more leakage before ∼3.5\sim3.5 Myr, before the latter begins to dominate at later times due to the presence of binary stars which increase the production efficiency at older stellar populations. Finally, we find that the time evolution of N˙ion,esc\dot{N}_{\mathrm{ion,esc}} is fundamentally different in the efficient star formation simulation. This is due to the fact that feedback is weaker, meaning that galaxies in this simulation can over-cool – requiring multiple successive starbursts to clear a particularly dense ISM (see also discussion in Choustikov et al. [26]).

It is important to note that the ionizing luminosity-weighted stellar age is sensitive to the SFH, particularly for galaxies with complex or extended SFHs, and may therefore not directly correspond to the time since the last starburst. Therefore, to confirm our findings, Figure 7 shows a compilation of smoothed (with a standard deviation of 1 Myr) median SFHs computed for stacks in fescf_{\mathrm{esc}} (top) and fescξionf_{\mathrm{esc}}\xi_{\mathrm{ion}} (bottom) considering all galaxies with MUV≤−15M_{\mathrm{UV}} \leq-15 in the bursty star formation simulation. Regions where the median SFH appears flat is caused by the fact that we are averaging over many galaxies which would have experienced bursts at different times. In both cases, we find that galaxies which can and do leak more ionizing photons tend to show a post starburst feature, in contrast to non-leakers which feature rapidly rising SFHs representative of an ongoing burst of star formation. These results echo the trends highlighted for the bursty star formation simulation in Figure 6: that there is often a characteristic time delay between the onset of a starburst and the conditions necessary for effective leakage. Furthermore, we note that this time delay is similar to that of the ‘bursty’ leakers found in the SPHINX20^{20} simulation (see Figure 3 of Katz et al. [91]).

Figure 7. Typical SFHs of LyC-leaking galaxies. We show median SFHs of galaxies with MUV≤−15M_{\rm UV} \leq-15 in the bursty star formation simulation, stacked as a function of fescf_{\rm esc} (top) and fescintf_{\rm esc}^{\rm int} (bottom). These have been smoothed with a Gaussian kernel of standard deviation ∼1\sim1 Myr and renormalised at the present day. Rapidly rising SFHs are emblematic of non-leakers, while on average, LyC-leaking systems appear to be in the post-starburst phase due to the need for a strong feedback episode to clear channels through the neutral ISM.

Sensitivity of LyC Leakage Timescales toFeedback Physics

Interestingly, we have found that the presence of earlier feedback alongside the drive to form stars in lower-density gas (as used in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation) means that effective leakage can start earlier than the ∼3.5\sim3.5 Myr timescale discussed in Choustikov et al. [27], as shown by a dotted line in Figure 6. In contrast, the average galaxy in the other simulations begins leaking at or after this timescale, potentially driven by the time taken to photoionize and/or clear channels through the more dense ISM produced by these models (Choustikov et al. [26], in prep.). While numerous studies based around cosmological simulations have shown that there are characteristic feedback-dependent delays between starbursts and high escape fractions (e.g., Paardekooper, Khochfar & Dalla Vecchia 2015; Trebitsch et al. 2017b; Ma et al. 2020; Barrow et al. [8]; Katz et al. [91]; Choustikov et al. [27]), the exact timescale is not settled. For example, the simulations of Kimm & Cen (2014) found a characteristic leakage timescale of ∼10\sim10 Myr, equal to the mean main-sequence lifetime of a SNe progenitor assumed therein (Schaller et al. 1992).

Furthermore, the question of whether SNe are the key feedback mechanism is also debated. For example, high resolution simulations of GMC-scale simulations (e.g., Kim, Kim & Ostriker 2019; Kimm et al. 2019, 2022) seemingly agree that early stellar feedback is more important, particularly focusing on the impact of photoionization heating (see Section 4.1 of Rosdahl et al. [162] for a discussion of this physics in SPHINX20^{20}). Furthermore, Menon et al. (2025) find that the timescale to clear the birth cloud is shorter for star-forming regions with higher gas surface densities, owing to the shortening of collapse and gas consumption timescales alongside the added impact of radiation pressure in such high-density environments. While this is physics that operates on resolutions that are not always resolved in MEGATRON, it is important to note that while GMC-scale simulations follow the star-forming environment with exquisite detail, they do not include the bulk ISM of the host galaxy and therefore can only speak to the feedback necessary for radiation to escape past the birth cloud (see Figures 8 and 9 of Yoo, Kimm & Rosdahl 2020).

At the same time, several recent low-redshift observations have pointed out that direct radiation pressure may be the key feedback process to induce leakage, with observed LyC emitters featuring ages indicative of pre-SNe stellar populations (e.g., Komarova et al. [104]; Carr et al. [17]). However, it is possible that this is a temporary line-of-sight effect driven by the lack of LyC production following the death of massive O stars, with other observations suggesting that SNe-driven outflows can provide effective escape channels for neighbouring stellar populations (Peng et al. 2023; Martin, Peng & Li 2024). Nevertheless, it has been well established that the pre-processing effect of radiation feedback is vital (e.g., Rosdahl et al. 2015; Petterson et al. 2025) and further observations to better constrain the impact of early stellar feedback on angle-averaged fescf_{\mathrm{esc}} are needed. Finally, the efficiency of pre-SNe feedback in the context of a dense ISM characteristic to starburst galaxies in the early Universe (Isobe et al. 2023; Redey et al. 2023; Abdurro’uf et al. 2024; Senchyna et al. 2024; Topping et al. [205, 204]; Li et al. [110]) remains to be seen, where physics not presently included in MEGATRON such as Lyα\alpha pressure (e.g., Smith, Bromm & Loeb 2016; Kimm et al. 2018; Nebrin et al. 2025, 2026; Menon & Smith 2026) and cosmic rays (e.g., Rodríguez Montero et al. 2024; Farcy et al. 2025) may also be vital pieces of the puzzle.

Altogether, we find that more effective, earlier feedback leads to a scenario where a smaller number of galaxies (across all UV magnitudes) are capable of driving significantly earlier reionization. In contrast, smoother SFHs with less effective feedback lead to a much larger number of bright sources but a much slower reionization process owing to the delay before effective leakage in individual galaxies. Furthermore, feedback in lower-density gaseous environments will also lead to escape of ionizing radiation earlier into the burst cycle. However, irrespective of feedback, the DM halos that typically contribute to

reionization remain unchanged.

The ISM Structure of UV-faint LyC Leakers in MEGATRON

Across all four MEGATRON simulations, we find that UV-faint, low mass DM halos can contribute significantly to reionization of their local environment. Therefore, it is instructive to understand the physics that drives leakage in such systems.

To accomplish this, we investigate the structure of the multi-phase ISM, making use of the detailed PRISM ISM model used in MEGATRON (Katz et al. [87] (2026c)). For each simulation, we compute mass-weighted density-temperature phase diagrams for a mass-matched sample (see e.g., Kim & Ostriker 2017; Katz et al. [87] for a discussion). Specifically, we consider all gas within the ISM (R<0.25RvirR < 0.25 R_{\mathrm{vir}}) of all galaxies with 106≤M∗/M⊙≤106.510^{6} \leq M_{*}/M_{\odot} \leq10^{6.5}, before then stacking these based on LyC escape fractions of each halo. Figure 8 shows the resulting phase diagrams, organised such that each column denotes a different range of fescf_{\mathrm{esc}} (stronger leakers on the right), with each simulation shown on a unique row. Focusing first on the left-most column, we confirm that while each MEGATRON simulation produces a multi-phase ISM, the exact structure and properties of the ISM are different in each simulation, owing to the different implementations of feedback and leading directly to differences in their observed properties (Katz et al. [84]; Choustikov et al. [26]; Attard et al. 2026).

Figure 8. Multi-phase ISM structure of non-leaking galaxies compared to intermediate and strong LyC leakers. In each panel, we show mass-weighed density-temperature phase diagrams of all gas in the ISM of a mass-matched (106≤M∗/M⊙≤106.510^{6} \leq M_{\ast}/M_{\odot} \leq10^{6.5}) sample of galaxies. Each column corresponds to a stack of galaxies with LyC escape fractions within a specific range (larger fescf_{\mathrm{esc}} from left to right). Similarly, each row corresponds to the relevant stack for each MEGATRON simulation. While the intrinsic ISM structure is different for each simulation as a function of changes to star-formation/feedback physics, the mechanisms that drive the production of escape channels are unchanged. Strong photoionization feedback is present in intermediate leakers (1%≤fesc≤10%1\% \leq f_{\mathrm{esc}} \leq10\%), strong leakers (10%≤fesc≤50%10\% \leq f_{\mathrm{esc}} \leq50\%) display a combination of photoionization and SNe feedback, while the cold/dense ISM of extreme leakers (fesc≥50%f_{\mathrm{esc}} \geq50\%) is almost completely disrupted.

Comparing the different fescf_{\mathrm{esc}} bins for the efficient star formation simulation (top row), we find that galaxies with 1%≤fesc≤10%1\% \leq f_{\mathrm{esc}} \leq10\% show significantly more gas at 15,000 K −- 20,000 K, as compared to the weakest leakers considered, where the majority of photoionized gas sits at 10,000 K11 (e.g., Rosdahl et al. 2015). The difference here amounts to whether the photoionized gas is mixed with the neutral ISM, with pure photoionized gas sitting on the second track at larger temperatures (e.g., Rey et al. [157]). Continuing to the strongest leakers (fesc≥50%f_{\mathrm{esc}} \geq50\%), we find a completely different ISM structure. In all cases, we find that photoionization heating (especially in dense environments) has been incredibly impactful, heating a significant mass fraction of the ISM to close to ∼30,000\sim30{,}000 K. Among these galaxies, the cold (T∼100T \sim100 K), dense, star-forming ISM is almost effectively depleted −- a strong signature of SNe feedback that has effectively coupled to the ISM, heating gas to ∼106\sim10^{6} K (Gent 2012; Kim & Ostriker 2017). Furthermore, we note that the bursty star formation and high ϵff\epsilon_{\mathrm{ff}}, HN simulations are capable of efficient ejective feedback, clearing gas beyond the central 25% of the virial radius (see Figure 8 of Attard et al. 2026). Interestingly, the dense photoionized line-emitting gas (nH≳103 cm−3n_{\mathrm{H}} \gtrsim10^{3}\ \mathrm{cm}^{-3} and T∼104T \sim10^{4} K, see Figure 2 of Choustikov et al. [26]) has also been depleted, consistent with the fact that extremely strong LyC leakers show weak emission line equivalent widths (e.g., Nakajima & Ouchi 2014; Steidel et al. 2018; Choustikov et al. [27]).

Building on our findings for the efficient star formation simulation, we can repeat this exercise for the other MEGATRON simulations, shown in the other rows. Despite the fact that the intrinsic ISM structure and fescf_{\mathrm{esc}} distributions (see Figure 4) are completely different across the suite, we find that the trends described above are consistent. In general, strong photoionization feedback is a hallmark of intermediate leakers (1%≤fesc≤10%1\% \leq f_{\mathrm{esc}} \leq10\%), strong leakers (10%≤fesc≤50%10\% \leq f_{\mathrm{esc}} \leq50\%) show a combination of effective photoionization and SNe feedback, and the strongest leakers (fesc≥50%f_{\mathrm{esc}} \geq50\%) show a lack of dense star-forming and line-emitting gas, with effectively all of the gas having been photoionized and cleared by SNe. This is a crucial finding, as it shows that while changes to feedback physics can modify fescf_{\mathrm{esc}}, the mechanisms by which it achieves the production of escape channels through the dense ISM are unchanged. To further highlight this, Figure 9 shows an integrated variant of the same phase diagrams. Specifically, the left/right columns show density/temperature distributions from the weakest leakers (red) to the strongest (blue). Again, we find that while the structure of the ISM is different between the four MEGATRON simulations, the signatures of higher fescf_{\mathrm{esc}} are the same: less cold neutral/dense gas (T≲103T \lesssim10^3 K and nH≳102 cm−3n_{\mathrm{H}} \gtrsim10^2\,\mathrm{cm}^{-3}) and more warm ionized/low-density gas (T≳104.5T \gtrsim10^{4.5} K and nH≲1 cm−3n_{\mathrm{H}} \lesssim1\,\mathrm{cm}^{-3}). Moreover, it is interesting to note that in contrast to the other simulations, some dense gas remains intact even in extreme leakers (fesc≥50%f_{\mathrm{esc}} \ge50\%) in the efficient star formation simulation. Looking towards the future, it would also be interesting to explore the turbulent properties of each ISM phase in LyC-leaking galaxies, as under idealised conditions the presence of turbulence can occasionally produce low-density channels, promoting leakage without the need for any feedback whatsoever (e.g., Xing et al. 2026, see also discussion by Kakiichi & Gronke 2021). Furthermore, early feedback is more effective at ejecting and ionizing lower-density channels in a turbulent medium while leaving higher-density channels unaffected (e.g., Thompson & Krumholz 2016; Menon et al. 2025), thereby increasing the patchiness of the ISM and associated LyC escape. We however leave such discussions to future work, as MEGATRON is unable to resolve such physics at present.

Figure 9. Density/temperature structure of LyC leakers. We show ISM gas mass fractions as a function of density (left) and temperature (right) for each MEGATRON simulation (rows). Each line is computed by integrating the relevant phase diagram of Figure 8. The ISM of galaxies with larger LyC escape fractions feature progressively less dense/cold gas and significantly more hot gas. Similarly, the ratio of peaks at ∼104\sim10^{4} K and ∼104.5\sim10^{4.5} K is sensitive to fescf_{\mathrm{esc}}, indicating the presence of strong photoionization feedback.

To consolidate these findings, Figure 10 shows fescf_{\mathrm{esc}} as a function of the hydrogen gas density weighted by [O II] λλ3727\lambda\lambda3727, ⟨nH⟩[O II]\langle n_{\mathrm{H}}\rangle_{\mathrm{[O\,II]}}. Each pixel shows the median gas fraction for galaxies across all four MEGATRON simulations with MUV≤−15M_{\mathrm{UV}} \le-15. Crucially, we find that all four simulations reproduce the same running median – confirming that gas density contains some information about the leakage of ionizing radiation. Interestingly, we find that hydrogen density weighted by either the [O II] λλ3727\lambda\lambda3727 or [S II] λλ6717,6731\lambda\lambda6717,6731 doublets carry considerably more information about fescf_{\mathrm{esc}} than densities weighted by C III] λλ1907,1909\lambda\lambda1907,1909, or N IV] λλ1483,1486\lambda\lambda1483,1486. The reason for this is likely due to the fact that [O II] and [S II] are preferentially sensitive to lower-density gas (see discussions in Kewley et al. 2019; Choustikov et al. [26]). Furthermore, for a given density profile, a low electron density measured from [O II] implies a larger ionized region due to the fact that O+^+ is a bivalent species owing to the charge exchange interaction with hydrogen (Jin, Kewley & Sutherland 2022; Xing et al. 2026). Altogether, ⟨nH⟩[O II]\langle n_{\mathrm{H}}\rangle_{\mathrm{[O\,II]}} carries more information about the density structure on the edge of the H II regions. Finally, we find that the pure leakers with fesc∼100%f_{\mathrm{esc}} \sim100\% tend to be extremely gas-poor, having almost completely evacuated their ISM within 25% of the virial radius (see also Figure 8 of Attard et al. 2026). Such pure leakers tend to have weak emission line equivalent widths as well as high values of [O III] λ5007\lambda5007/[O II] λλ3727\lambda\lambda3727 for a given gas density (see Figure 9 of Choustikov et al. [26]).

Connecting LyC escape to gas density

Figure 10. Connecting LyC escape to gas density. We show fescf_{\mathrm{esc}} as a function of the [O II] λλ3727\lambda\lambda3727-weighted hydrogen gas density for galaxies across all MEGATRON simulations with MUV≤−15M_{\mathrm{UV}} \le-15. Each pixel is coloured by the median gas fraction of individual galaxies, highlighting that halos with fesc∼100%f_{\mathrm{esc}} \sim100\% have expelled a significant portion (≳90%\gtrsim90\%) of their gas. Running medians are provided for each individual simulation, highlighting that while the distribution of escape fractions is strongly modified by feedback physics, the connection between fescf_{\mathrm{esc}} and ⟨nH⟩[O II]\langle n_{\mathrm{H}}\rangle_{\mathrm{[O\,II]}} remains unchanged.

A visual representation of this discussion is presented in Figure 11. Here, we show projections along each axis in gas density (purple), electron density (red), LyC luminosity (white) and He+^+-ionizing luminosity (yellow) for the most massive galaxy at redshift z∼8.5z \sim8.5 in the efficient star formation (top) and bursty star formation (bottom) simulations. In the composite images, we find that LyC leakage is highly anisotropic in both simulations, in agreement with previous studies (Gnedin, Kravtsov & Chen 2008; Wise & Cen 2009; Paardekooper, Khochfar & Dalla Vecchia 2015; Trebitsch et al. 2017a; Kim et al. [99]; Choustikov et al. [28]; Sollvaug et al. 2025; Ejdetjärn et al. 2026). Sightlines featuring high hydrogen column densities readily absorb ionizing flux, leaving behind regions of high electron number densities. Furthermore, it is important to note that there are significant morphological differences in the galaxies produced by these models (see discussion in Matsumoto et al. in prep.). For example, galaxies in the efficient star formation simulation feature cold, dense discs with high electron number densities and biconical LyC escape. In contrast, galaxies in the bursty star formation simulation are considerably more clumpy and disturbed by strong feedback, with LyC absorption occurring also due to filaments and ejected material in the CGM. This trend agrees with previous work with the SPICE simulations, which showed that dispersion-dominated galaxies tend to feature systematically larger escape fractions (Bhagwat et al. 2024a). Finally, it is interesting to note that some galaxies in the efficient star formation simulation leak

Figure 11. Anisotropic LyC emission from galaxies in the efficient/bursty star formation simulations at redshift z∼8.5z \sim8.5. We show projections of hydrogen gas density (purple), electron density (red), LyC luminosity (white), and He+\mathrm{He}^{+}-ionizing luminosity (yellow) along each axis of the most massive galaxy at redshift z∼8.5z \sim8.5 in the efficient star formation (top) and bursty star formation (bottom) simulations. In each case, a composite of the four quantities is also provided. In both simulations, we find that LyC leakage is highly anisotropic and frequently attenuated along sightlines with high gas column densities, where high electron number densities can be found.

He+^{+}-ionizing radiation in only a subset of LyC-leaking sightlines. As the former is hard radiation (here12 produced by lower-metallicity young massive stars, [173] (Xiao, Stanway & Eldridge 2018; Stanway & Eldridge 2019)), this implies that such galaxies are releasing LyC late into a starburst cycle, as demonstrated in Figure 6 and Figure 7.

Probing Galactic Contributions to Reionization with ISM Density

Given the findings of § 4, we now proceed to investigate whether directly observable probes of ISM properties might carry information about fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}}.

ISM Density as a Probe of LyC Escape

To begin, Figure 12 shows fescf_{\mathrm{esc}} as a function of the electron number density inferred from the [O II] λλ3727\lambda\lambda3727 doublet, nen_{\mathrm{e}}([O II]), for galaxies with MUV≤−17M_{\mathrm{UV}} \leq-17 and EW(Hβ\beta) >50> 50Å across each MEGATRON simulation. An added cut of EW(Hβ\beta) ≥50\geq50Å is included to account for the fact that it is difficult to reliably measure ISM properties from weak emission lines. We find that a trend does emerge, driven by the presence of strong leakers with nen_{\mathrm{e}}([O II]) ∼30−100 cm−3\sim30 - 100\ \mathrm{cm}^{-3} and dust-free blue UV continuum slopes (β≲−2.6\beta\lesssim-2.6, [25]) alongside non-leakers with nen_{\mathrm{e}}([O II]) ∼1000 cm−3\sim1000\ \mathrm{cm}^{-3} and red UV continuum slopes (β≳−2.4\beta\gtrsim-2.4) driven by the nebular continuum ([83]). Interestingly, we find several galaxies with blue UV slopes but low escape fractions. Such systems all feature ISM densities lower than the median relation and feature blue slopes due to a nebular continuum that has been weakened by non-equilibrium physics ([84]).

Predicting LyC escape with electron density

Figure 12. Predicting LyC escape with electron density. We show fescf_{\mathrm{esc}} as a function of electron number densities measured from the [O II] λλ3727\lambda\lambda3727 emission line doublet, with individual pixels coloured by the UV continuum slope, β\beta. We include all galaxies with MUV≤−17M_{\mathrm{UV}} \leq-17 and EW(Hβ\beta) >50> 50Å in order to produce an observationally comparable sample. Running medians are provided for each individual simulation, alongside a quadratic fit as given in Equation (2). We find this trend continues to be robust to stellar feedback physics, highlighting that nen_{\mathrm{e}}([O II]) contains useful information about the leakage of ionizing radiation.

Similar to the results of Figure 10, we find that running medians from the four simulations tend to agree (albeit with comparatively more scatter), suggesting that there is still a consistent trend. A line of best fit is computed (amalgamating data from all four simulations) to capture this trend, given by:

log⁡(fesc)=−1.094+1.890x−0.939x2,(2)\log(f_{\mathrm{esc}}) = -1.094 + 1.890x - 0.939x^2, \tag*{(2)}

where x=log⁡(ne([O II])/[cm−3])x = \log(n_{\mathrm{e}}([\mathrm{O\ II}]) / [\mathrm{cm}^{-3}]). We re-iterate that this trend is not predictive, especially when compared to published multi-variate relations for fescf_{\mathrm{esc}} ((e.g., Mascia et al. 2023; Choustikov et al. 2024a,b; Jaskot et al. 2024a,b)) and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} ([198] (Choustikov et al. 2025)). Furthermore, there is reason to expect that this relation should weaken in more massive systems. Here, the combination of dust and the presence of multiple emitting regions are likely to wash out this tracer to a typical galactic-averaged value of nen_{\mathrm{e}}([O II]) ∼300−1000 cm−3\sim300 - 1000\ \mathrm{cm}^{-3} as line emission is dominated by dense gas ((Xing et al. 2026, see also discussion in Katz et al. 2026a)). Similarly, we find that while nen_{\mathrm{e}}([O II]) traces the global (angle-averaged) escape fraction, the trend does not hold for line-of-sight escape fractions (see discussion in Appendix § 11). Nevertheless, nen_{\mathrm{e}}([O II]) is potentially a useful indicator when applied to population-scale samples, especially given published trends for the observed redshift evolution of ISM densities ([e.g. 204, 121, 195] (Isobe et al. 2023; Abdurro’uf et al. 2024)).

We have also investigated a number of density-sensitive doublets as potential tracers of fescf_{\mathrm{esc}}. We find that indicators sensitive to low-density gas (e.g., [S II] 6716/6731, [O III] 88μ\mum/52μ\mum, and [S III] 33.5μ\mum/18.7μ\mum) can also provide insight about LyC escape (see discussion in Xing et al., in prep.). However, these ratios (especially [S II]) are less robust to feedback physics due to contributions from the diffuse ionised gas. In contrast, UV tracers of dense gas (e.g., C III] 1906/1908 and N IV] 1483/1486) are not as sensitive to the effective coupling of feedback in the ISM, as they are biased to extremely dense environments ([204, 121, 11]).

Spurred on by these findings, we investigate the dependence of N˙ion,esc\dot{N}_{\mathrm{ion,esc}} on nen_{\mathrm{e}}([O II]) in Figure 13. On the left, each panel shows this relation for a different bin in absolute UV magnitude, coloured by the burstiness of the recent SFH (defined using SFR1_1/SFR3_3). Furthermore, we retain the cut of EW(Hβ\beta) ≥50\geq50Å as discussed above. We find that in all cases, there is a trend wherein galaxies which appear to have a low-density ISM and a recently declining SFH are likely to be releasing the most ionizing photons into the IGM (in agreement with our findings in § 3 and § 4). Interestingly, galaxies with −20≤MUV≤−19-20 \leq M_{\mathrm{UV}} \leq-19 show the tightest relation. In contrast, fainter galaxies feature a more shallow slope with increased scatter. For each UV magnitude bin, we fit a power law of the form x=c0+c1y+c2y2x = c_{0}+c_{1}y+c_{2}y^{2}, where y=log⁡(N˙ion,esc/[#/s])y = \log(\dot{N}_{\mathrm{ion,esc}} / [\#/ \mathrm{s}]) and x=log⁡(ne([O II])/[cm−3])x = \log(n_{\mathrm{e}}([\mathrm{O\ II}]) / [\mathrm{cm}^{-3}]) using the same approach that is typically applied to strong-line metallicity calibrations (e.g., [139, 108, 167, 67 Choustikov et al., in prep.] (Sanders et al. 2024)). Each line is displayed in the relevant panel of Figure 13 and compiled on the rightmost panel alongside the underlying distribution of all galaxies with MUV≤−15_{\mathrm{UV}} \le-15 and EW(Hβ\beta) ≥50\ge50Å. Coefficients are provided in Table 1, though it is important to highlight that the fits provided in Table 1 are not strongly predictive (especially at the high-density limit). This means that they should be applied with caution and are best-used for population-level analyses.

MUVM_{\mathrm{UV}} Rangec0c_0c1c_1c2c_2
−20≤MUV<−19-20 \le M_{\mathrm{UV}} < -19−77.52-77.523.4063.406−0.03595-0.03595
−19≤MUV<−18-19 \le M_{\mathrm{UV}} < -18−112.9-112.94.8854.885−0.05137-0.05137
−18≤MUV<−17-18 \le M_{\mathrm{UV}} < -17−116.6-116.65.0685.068−0.05359-0.05359
−17≤MUV<−16-17 \le M_{\mathrm{UV}} < -16−108.9-108.94.8534.853−0.05190-0.05190

Table 1. Fitted functions for lines shown in Figure 13. All functions are of the form x=c0+c1y+c2y2x = c_0 + c_1y + c_2y^2, where x=log⁡(ne([O II])/[cm−3])x = \log(n_{\mathrm{e}}([\mathrm{O\,II}]) / [\mathrm{cm}^{-3}]) and y=log⁡(N˙ion,esc/[#/s])y = \log(\dot{N}_{\mathrm{ion,esc}} / [\#/\mathrm{s}]).

Figure 13. Connecting the escaping ionizing luminosity to emission line-derived ISM electron densities. We show N˙ion,esc\dot{N}_{\mathrm{ion,esc}} as a function of electron densities measured from the [O II] λλ3727\lambda\lambda3727 emission line doublet and coloured by the burstiness of the recent SFH (defined using SFR1/SFR3\mathrm{SFR}_{1}/\mathrm{SFR}_{3}). On the left, each panel shows a subset of galaxies in a specific UV magnitude range. We only include galaxies with EW(Hβ\beta) ≥50\ge50Å to account for the difficulty of measuring ISM properties from weak emission lines. We fit quadratics of the form log⁡(ne([O II]/[cm−3]))=c0+c1y+c2y2\log(n_{\mathrm{e}}([\mathrm{O\ II}] / [\mathrm{cm}^{-3}])) = c_{0} + c_{1}y + c_{2}y^{2} (where y=log⁡(N˙ion,esc/[#/s])y = \log(\dot{N}_{\mathrm{ion,esc}} / [\#/ \mathrm{s}])) to each subset of galaxies, with coefficients given in Table 1. Finally, we show the distribution of all galaxies with MUV≤−15_{\mathrm{UV}} \le-15 and EW(Hβ\beta) ≥50\ge50Å in this space, highlighting each of the fits. In general, galaxies with low electron densities (≲300 cm−3\lesssim300\ \mathrm{cm}^{-3}) and recently declining SFHs tend to release more ionizing photons into the IGM, with UV-bright galaxies showing less scatter in the trend. Nevertheless, these functions are not strongly predictive, meaning that they should mainly be used for population-level analyses.

Ambient ISM Properties of LyC-Leaking Galaxies: The TUne Diagram

Building on our findings that electron densities contain information about fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}}, we proceed to explore a multi-variate approach (see also Choustikov et al. [27]). To this end, we introduce the temperature-ionization parameter-density (TUne) diagram. Figure 14 shows O33_{3} = [O III] λ4363\lambda4363/[O III] λ5007\lambda5007 (a tracer of ISM temperature, Kewley, Nicholls & Sutherland 2019) as a function of O32_{2} = [O III] λ5007\lambda5007/[O II] λλ3727\lambda\lambda3727 (a tracer of ionization parameter and density, Martinez et al. [121]) for galaxies with MUV≤−15_{\mathrm{UV}} \le-15 across all four MEGATRON simulations, coloured by the median fescf_{\mathrm{esc}} (left), N˙ion,esc\dot{N}_{\mathrm{ion,esc}} (centre), and ne([O II])n_{\mathrm{e}}([\mathrm{O\ II}]) (right) of each pixel. We find that galaxies with fesc≳10%f_{\mathrm{esc}} \gtrsim10\% and N˙ion,esc≳1050.5 s−1\dot{N}_{\mathrm{ion,esc}} \gtrsim10^{50.5}\ \mathrm{s}^{-1} are confined to a particular portion of the diagram corresponding to a hot, low density ISM (consistent with the findings of Figure 8 and Figure 9). Furthermore, the strongest leakers (fesc∼100%f_{\mathrm{esc}} \sim100\%) have an extremely hot ISM, though much of this is driven by them also being very gas poor (making emission lines unobservable). An example visually defined boundary is shown in cyan, given by:

Identifying LyC leakers with ISM properties: the TUne diagram

Figure 14. Identifying LyC leakers with ISM properties: the TUne diagram. We provide a line ratio diagnostic diagram of all galaxies with MUV≤−15M_{\mathrm{UV}} \le-15 across each MEGATRON simulation. Specifically, we use the temperature-sensitive O33=[O III] λ4363/[O III] λ5007\mathrm{O33} = [\mathrm{O\,III}]\,\lambda4363/[\mathrm{O\,III}]\,\lambda5007 and ionization parameter/density-sensitive O32=[O III] λ5007/[O II] λλ3727\mathrm{O32} = [\mathrm{O\,III}]\,\lambda5007/[\mathrm{O\,II}]\,\lambda\lambda3727 ratios, colouring each pixel by the median value of fescf_{\mathrm{esc}} (left) and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} (centre), and ne([OII])n_{\mathrm{e}}([\mathrm{OII}]) (right). Significant outliers have low gas fractions, leading to extremely weak emission line equivalent widths. Finally, we indicate a potential helpful but insufficient classifier in cyan given by log⁡(O33)>0.47log⁡(O32)−2.1\log(\mathrm{O33}) > 0.47\log(\mathrm{O32}) - 2.1. This selection produces a sample of which >50%> 50\% of galaxies have fesc>10%f_{\mathrm{esc}} > 10\%, capturing 88%88\% of all such halos. Similarly, this selection captures the majority of galaxies with N˙ion,esc≳1052 s−1\dot{N}_{\mathrm{ion,esc}} \gtrsim10^{52}\ \mathrm{s}^{-1}, though galaxies with slightly larger values of O32 are still able to contribute to reionization, despite featuring LyC escape fractions below ∼1%\sim1\%.

log⁡(O33)>0.47×log⁡(O32)−2.1.(3)\log(\mathrm{O3}_{3}) > 0.47 \times\log(\mathrm{O3}_{2}) - 2.1. \tag*{(3)}

This helpful but insufficient criterion produces a sample of which 52% (14%) have fesc>10%f_{\mathrm{esc}} > 10\% (N˙ion,esc>1052 s−1\dot{N}_{\mathrm{ion,esc}} > 10^{52}\ \mathrm{s}^{-1}) and that manages to capture 90% (82%) of all such systems across all four MEGATRON simulations. Looking to the central panel, we find that galaxies releasing the most ionising photons into the IGM tend to have O32∼5−15_{2} \sim5 - 15 (consistent with the findings of Choustikov et al. [27]). At lower O32_{2}, N˙ion,esc\dot{N}_{\mathrm{ion,esc}} falls as fewer ionizing photons are being produced, while at larger O32_{2} the dense ISM has not yet been cleared and [O II] λλ3727\lambda\lambda3727 is collisionally de-excited. Nevertheless, we find that for galaxies with a fixed value of O32_{2} ratios within this range, the temperature of the ISM is a strong indicator of whether a galaxy is effectively leaking LyC photons.

Time-Evolution of Observational Diagnostics of LyC Escape

Another interesting potential benefit of ISM structure-based tracers is their potential ability to track the time-evolution of fescf_{\mathrm{esc}} over an individual burst cycle. For example, the electron density of a starburst will track both the ionization parameter as well as the process of feedback. This combination should therefore be sensitive to both density-bound and ionisation-bound leakage modes (Zackrisson, Inoue & Jensen 2013). Similarly, O33O_{33} is indicative of the high-energy ionisation parameter (e.g., traced by C IV λλ1550\lambda\lambda1550/C III] λλ1908\lambda\lambda1908, Berg et al. 2021), as shown by Figure 9 of Choustikov et al., (in prep.).

To investigate this aspect, Figure 15 shows smoothed medians for fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} (using the true intrinsic N˙ion,int\dot{N}_{\mathrm{ion,int}}) as a function of ionising luminosity weighted stellar age. We include smoothed (over a Gaussian of standard deviation 1 Myr) running medians considering the true LyC escape fraction (dotted), the ne([O II])n_{\mathrm{e}}([\mathrm{O\ II}])-based escape fraction given by Equation (2) (bold), the β\beta-based values using Equation 11 of Chisholm et al. [25], and the recent UV magnitude-based calibration13 using galaxies from the GLIMPSE survey (Jecmen et al. [75]). Here, we find that fesc(β)f_{\mathrm{esc}}(\beta) and fesc(MUV)f_{\mathrm{esc}}(M_{\mathrm{UV}}) are both roughly constant in time, with the former reproducing the population average value of ⟨fesc⟩=0.14\langle f_{\mathrm{esc}}\rangle= 0.14 reported by Jecmen et al. [75]. In contrast, we find that fesc(ne([O II]))f_{\mathrm{esc}}(n_{\mathrm{e}}([\mathrm{O\ II}])) begins to capture the time-dependent nature of fescf_{\mathrm{esc}} in MEGATRON galaxies. While this particular relation is less accurate for the variable IMF and high ϵff\epsilon_{\mathrm{ff}}, HNHN simulations (owing to their inclusion of hypernovae), the overall mean time evolution of N˙ion,esc\dot{N}_{\mathrm{ion,esc}} is much better captured than by the traditional approach of Chisholm et al. [25]. Finally, we note that while this method would over-predict the escaping ionising luminosity for the high ϵff\epsilon_{\mathrm{ff}}, HNHN simulation, this is not unexpected given the fact that the ISM of this model is characteristically less dense (see Figure 7 of Choustikov et al. [26]).

Figure 15. Time-evolution of LyC leakage from galaxies using a variety of indirect models for the escape fraction. We show running means (smoothed with a Gaussian with standard deviation 1 Myr) of fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} as a function of the ionizing luminosity-weighted mean stellar age for galaxies in each MEGATRON simulation with MUV≤−13M_{\mathrm{UV}} \le-13. We include lines based on the true escape fraction (dotted), as well as based on indirect methods such as the O2-inferred electron number density (solid, Equation (2)), the UV continuum slope (dashed, Chisholm et al. [25]), and the UV magnitude (dot-dashed, Jecmen et al. [75]). We find that fesc(β)f_{\mathrm{esc}}(\beta) and fesc(MUV)f_{\mathrm{esc}}(M_{\mathrm{UV}}) are constant with time, compared to the significant early time evolution in the true fescf_{\mathrm{esc}}. This leads to systematic over-predictions in N˙ion,esc\dot{N}_{\mathrm{ion,esc}} by up to two orders of magnitude. In contrast, fesc(ne([O II]))f_{\mathrm{esc}}(n_{\mathrm{e}}([\mathrm{O\,II}])) is capable of reproducing much of this early time evolution, especially when applied to the efficient star formation and bursty star formation simulations. While the ISM density-based approach systematically over-predicts N˙ion,esc\dot{N}_{\mathrm{ion,esc}} at late times in the high ϵff\epsilon_{\mathrm{ff}}, HN simulation, this is a result of the fact that this model produces galaxies with a characteristically less dense ISM, therefore biasing the inferred escape fractions.

Altogether, we emphasise that the early time-evolution of fescf_{\mathrm{esc}} and ξion\xi_{\mathrm{ion}} of UV-faint galaxies must be jointly accounted for in models of the reionization process. Another potential avenue to accomplish this may be to condition traditional escape fraction diagnostics on the inferred recent star formation history. We leave such discussions to future work.

Conclusions

Understanding the process of cosmic reionization – and by extension – the escape of Lyman Continuum (LyC) photons from star-forming galaxies is one of the key questions of modern astrophysics. In particular, it is vital to reconcile observed high-redshift UV luminosity functions and ionizing production efficiencies of early stellar populations with reionization histories that have been measured from parallel probes. Here, LyC escape fractions continue to be the main degree of uncertainty – traditionally assumed constant or inferred by indirect diagnostics that have been calibrated on low-redshift analogues.

To address this challenge, we conduct a study of the process of reionization in the MEGATRON simulations (Katz et al. [88]; Cadiou et al. [16]; Rey et al. [156]): a suite of high-resolution cosmological zoom-in simulations of a Milky Way-mass progenitor that feature on-the-fly radiation transfer coupled to an extensive non-equilibrium thermochemical network (81 species) and detailed model for heating and cooling processes in the ISM. In particular, we utilise a number of re-simulations which feature modifications to star-formation/feedback physics, inspired by various potential solutions to the observed number of UV-bright galaxies and capable of reproducing the observed ISM properties of high-redshift galaxies (Choustikov et al. [26]). With this as a basis, we investigate both the demographics and ISM structure of LyC leaking halos. Our results are as follows:

  • Number counts of UV sources are directly linked to the distribution of LyC escape fractions by feedback physics. Changes to star formation/feedback physics will directly mediate the number of galaxies at a given MUVM_{\mathrm{UV}}, the contribution of each galaxy to the ionizing photon background, and ultimately the reionization history as a whole. As a result, the presence of a large number of UV sources does not imply a commensurate increase in the total escaping ionizing luminosity, with the two features appearing to anti-correlate.

  • Understanding the time-delay between a starburst and the onset of effective LyC escape is the key challenge. The primary physical factor connecting the number of galaxies to the ionizing photon budget is the relative timescale of processes that allow LyC photons to escape compared to the lifetime of LyC-producing stellar populations. More effective early feedback (e.g., radiation and hypernovae) released in lower-density environments leads to a scenario where a smaller number of galaxies are capable of driving an earlier reionization. In contrast, efficient star formation in dense environments leads to longer delays between a starburst and effective leakage – leading to a slow reionization irrespective of an increase to UV source number counts.

  • UV-faint LyC leakers tend to feature post-starburst star formation histories. Similarly, UV-faint galaxies with rapidly rising, starburst, SFHs do not leak LyC photons. When the production efficiency of ionizing radiation is accounted for, post-starburst systems are still the dominant UV-faint LyC leakers.

  • The characteristic ISM structure of UV-faint LyC leakers is similar across models with different feedback physics. While changes to feedback fundamentally modify the structure of the ISM, the physical mechanisms that drive LyC escape are similar in all four simulations. Strong photoionization feedback is a hallmark of intermediate (1%≤fesc≤10%1\% \leq f_{\mathrm{esc}} \leq10\%) leakers, strong leakers (10%≤fesc≤50%10\% \leq f_{\mathrm{esc}} \leq50\%) show a combination of effective coupling from photoionization and SNe feedback, while extreme leakers (fesc≥50%f_{\mathrm{esc}} \geq50\%)

show a lack of dense star-forming and line-emitting gas. This leads to a relationship between fescf_{\mathrm{esc}} and the average hydrogen gas density that is consistent across all four simulations.

  • ISM Electron number density measured from the [O II] λλ3727\lambda\lambda3727 emission line doublet is a potential indicator of LyC escape in UV-faint galaxies. Based on our finding that gas density contains information about the production of escape channels, we propose that the electron number density measured from the [O II] λλ3727\lambda\lambda3727 doublet may be a useful indicator of fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}} in UV-faint galaxies. Crucially, this trend appears to be robust to changes to feedback physics. While these trends are not strictly predictive, they may provide a viable method to identify leakers among large populations of galaxies. Furthermore, we note that ne([O II])n_{\mathrm{e}}([\mathrm{O\ II}]) can be polluted by emission from multiple emitting regions (e.g., in more massive galaxies) and is generally more indicative of global fescf_{\mathrm{esc}} as compared to the value measured along a single sight-line.

  • At a fixed ionization parameter, the ISM of LyC leakers is hotter and less dense. This is summarised in the TUne diagram, which combines O33 = [O III] λ4363\lambda4363/[O III] λ5007\lambda5007 and O32 = [O III] λ5007\lambda5007/[O II] λλ3727\lambda\lambda3727. Galaxies with high escape fractions cluster in a region of the O33-O32 space that is associated with lower electron densities and high temperatures. Furthermore, galaxies with large escaping ionizing luminosities tend to show larger values of O32 but still fall in the same region.

  • Traditional UV continuum slope-based escape fraction calibrations over-predict escaping ionizing luminosities by failing to capture the time evolution of the feedback processes. In contrast to traditional indicators based on the UV continuum slope (β\beta), ne([O II])n_{\mathrm{e}}([\mathrm{O\ II}]) is able to better capture the time evolution of fescf_{\mathrm{esc}} and N˙ion,esc\dot{N}_{\mathrm{ion,esc}}. At early times (τ∗≲5\tau_{\ast} \lesssim5 Myr), the fesc(β)f_{\mathrm{esc}}(\beta) calibration of Chisholm et al. [25] may over-predict the luminosity of escaping ionising radiation from individual sources by ∼2\sim2 orders of magnitude. However, this calibration tends to the correct value at later times. The delayed duty cycle of LyC leakage must be accounted for in models of reionization that are based on observed UV source number counts.

This work has highlighted that reionization-relevant observables such as UV luminosity functions, ionizing photon production efficiencies, and LyC escape fractions are inextricably linked by time-delayed duty cycles informed by feedback physics. Furthermore, we argue that tracers of the multi-phase ISM structure such as electron densities and temperatures should be explored as potential fescf_{\mathrm{esc}} indicators – especially in the era of exquisite spectra observed by JWST. Nevertheless, future work will be needed to characterise the impact of other feedback channels (e.g., Lyα\alpha pressure, cosmic rays, and AGN) on the time-dependent process of Lyman Continuum escape. Furthermore, we will ultimately need to address the divide between ultra-high resolution GMC-scale star formation simulations and cosmological simulations of reionization with a resolved multi-phase ISM. MEGATRON represents good progress in this direction, but we must continue to leverage the synergy between high-resolution simulations, synthetic observations, and state-of-the-art observational facilities to understand the contributions of individual galaxies to cosmic reionization.

Data Availability

The data employed in this manuscript will be shared upon reasonable request to the corresponding author.

Acknowledgements

The authors would like to thank Lily Whitler and Christopher Cain for insightful discussions.

NC acknowledges support from the Kavli Institute for Cosmology, Cambridge and the Kavli Foundation. TK were supported by the National Research Foundation of Korea (RS-2022-NR070872 and RS-2025-00516961). MJH is fellow of the Knut & Alice Wallenberg Foundation. MPR acknowledges support from UKRI Science and Technology Facilities Council (STFC) grant UKRI4103. Support for program JWST-AR-08709 (AS) was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-03127.

Megatron made extensive use of the dp016, dp265, dp373, and dp379 projects on the DiRAC ecosystem. This work was performed using the DiRAC Data Intensive service at Leicester, operated by the University of Leicester IT Services, which forms part of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The equipment was funded by BEIS capital funding via STFC capital grants ST/K000373/1 and ST/R002363/1 and STFC DiRAC Operations grant ST/R001014/1. This work used the DiRACODurham facility managed by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The equipment was funded by BEIS capital funding via STFC capital grants ST/P002293/1, ST/R002371/1 and ST/S002502/1, Durham University and STFC operations grant ST/R000832/1. This work was performed using resources provided by the Cambridge Service for Data Driven Discovery (CSD3) operated by the University of Cambridge Research Computing Service (www.csd3.cam.ac.uk), provided by Dell EMC and Intel using Tier-2 funding from the Engineering and Physical Sciences Research Council (capital grant EP/T022159/1), and DiRAC funding from the Science and Technology Facilities Council (www.dirac.ac.uk). DiRAC is part of the National e-Infrastructure. The authors thank Jonathan Patterson for smoothly running the Glamdring Cluster hosted by the University of Oxford, where part of the data processing was performed. The authors also acknowledge financial support from Oriel College’s Research Fund. The material in this manuscript is based upon work supported by NASA under award No. 80NSSC25K7009.

The authors thank Romain Teyssier and Leo Michel-Dansac for both developing and open-sourcing RAMSES and RASCAS, respectively. We thank the developers and maintainers of pynbody ((Pontzen et al. 2013)), yt ((Turk et al. 2011)), NumPy ([212, 60]), SciPy ([215]), jupyter ((Ragan-Kelley et al. 2014)), matplotlib ((Hunter 2007)), cmasher ([211]), pandas ((pandas development team 2020; Wes McKinney 2010)), the Astrophysics Data Service, and the arXiv pre-print repository for providing open-source software and services that were used extensively in this work.

Intrinsic ISM Gas Density and Observed Electron Density

In Figure 16, we show a comparison between the electron number density inferred from [O II] λλ3727 and the intrinsic average hydrogen gas density weighed by [O II] for all galaxies with MUV≤−15M_{\mathrm{UV}} \leq-15 across each MEGATRON simulation. We find that there is a characteristic offset of ∼0.8\sim0.8 dex, with the electron density under-predicting the typical gas density of these regions. This is consistent with the fact that the [O II] doublet is sensitive to a less-dense (owing to the low critical densities, [121] (Kewley et al. 2019)) and lower ionization parameter regime.

Figure 16. A comparison between the intrinsic gas density and inferred electron density traced by O+^{+}. We show ⟨nH⟩[O II]\langle n_{\mathrm{H}}\rangle_{\mathrm{[O\,II]}} as a function of ne([O II])n_{\mathrm{e}}(\mathrm{[O\,II]}) for all galaxies with MUV≤−15M_{\mathrm{UV}} \leq-15 across all four MEGATRON simulations.

Impact of star-formation/feedback physics on scaling relations

Median scaling relations for the halo mass (MvirM_{\mathrm{vir}}), stellar mass (after mass loss, M∗M_{*}), and UV magnitudes (MUVM_{\mathrm{UV}}) of galaxies at redshift 8.5≤z≤9.58.5 \leq z \leq9.5 across all four MEGATRON simulations are shown in Figure 17. We also include the median relation for N˙ion,esc\dot{N}_{\mathrm{ion,esc}} as function of halo mass. As a comparison point, we show the scaling relations for a baryon conversion efficiency, f∗=0.01f_{*} = 0.01 of Kravtsov & Belokurov (2024), a semi-analytic model that was able to reproduce observed UVLFs by leveraging scatter in the SFHs of galaxies (see also Mason, Trenti & Treu (2023)). The bursty star formation and high ϵff\epsilon_{\mathrm{ff}}, HN simulations produce galaxies of a similar stellar mass (and UV luminosity) for a given halo mass. In contrast, galaxies in the efficient star formation simulation form stars much more efficiently (see discussion in Katz et al. [88]), while the considerable stellar mass loss due to feedback in the variable IMF simulation leads to a lower normalisation. Similarly, the four models produce a very similar MUV−M∗M_{\mathrm{UV}} - M_{*} relations (up to their scatter), with the variable IMF simulation featuring a boost at the higher masses due to the assumed SPS models (Leitherer et al. [109]; Eldridge et al. (2017); Stanway & Eldridge (2018)). Finally, we find that at a fixed halo mass, there is a large degree of scatter in the escaping ionizing luminosity, with the four simulations’ median relations agreeing more than expected.

Figure 17. Halo mass, stellar mass, UV magnitude, and escaping ionizing luminosity scaling relations in MEGATRON. We show the joint scaling relations between the DM halo virial mass (MvirM_{\mathrm{vir}}), stellar mass (after mass loss, M∗M_\ast), UV magnitude (MUVM_{\mathrm{UV}}), and escaping ionizing luminosity N˙ion,esc\dot{N}_{\mathrm{ion,esc}} for galaxies at redshift 8.5≤z≤9.58.5 \le z \le9.5 across all four MEGATRON simulations. In each case, the running median is shown alongside the single standard deviation scatter. As a comparison point, we show the scaling relations produced by the semi-analytic model of Kravtsov & Belokurov (2024) that was able to reproduce observed UVLFs.

Robustness Test of Electron Density as a Tracer of Escape Fraction

To further test the robustness of the relation between fescf_{\mathrm{esc}} and ne([O II])n_{\mathrm{e}}(\mathrm{[O\,II]}) proposed in § 5, we proceed to investigate this relation in galaxies from the SPHINX20^{20} simulation ([163, 162]). To do so, we use synthetic observations provided in the SPHINX20^{20} Public Data Release v1 (SPDRv1, [90]). We measure electron number densities from the intrinsic [O II] λλ3727\lambda\lambda3727 luminosities of the 696 galaxies at redshift z≥6z \ge6 with SFR10≥0.5 M⊙yr−1\mathrm{SFR}_{10} \ge0.5\,M_\odot\mathrm{yr}^{-1} using the method outlined in § 2. Finally, we make use of both the global (i.e., angle-averaged) and line-of-sight LyC escape fractions (fescLOSf_{\mathrm{esc}}^{\mathrm{LOS}}) that were computed using RASCAS ((Michel-Dansac et al. 2020, see discussions in Choustikov et al. 2024a,b)).

Figure 18 shows the global fescf_{\mathrm{esc}} (left) and line-of-sight fescf_{\mathrm{esc}} (right) as a function of ne([O II])n_e([\mathrm{O\,II}]) for SPHINX20\mathrm{SPHINX}^{20} galaxies. We find that the trend with global fescf_{\mathrm{esc}} holds. Given the fact that SPHINX20\mathrm{SPHINX}^{20} uses a different stellar feedback prescription, a simpler treatment of heating and cooling physics in the ISM and slightly worse resolution of ∼10\sim10 pc [163, 162], this confirms that fesc−ne([O II])f_{\mathrm{esc}}-n_e([\mathrm{O\,II}]) appears to be a robust trend in simulated galaxies during the Epoch of Reionization. Interestingly however, we find that no such trend exists when the line-of-sight escape fractions are considered. This speaks to the highly anisotropic nature of LyC escape (see Appendix § 10 of Choustikov et al. [28] for a discussion of this in the context of SPHINX20^{20}). Specifically, galaxies with global LyC escape fractions >10%> 10\% are typically dominated by a few ‘successful’ sightlines while also frequently featuring sightlines with fescLOS∼0f_{\rm esc}^{\rm LOS} \sim0 – thereby weakening the relation. Another factor which should be considered here is the impact of morphology and ISM geometry (see discussions by Bhagwat et al. 2024a), which would fundamentally alter the distribution of fescf_{\rm esc} for a given feedback event and ISM structure.

Figure 18. Testing the robustness of the escape fraction – density relation with the SPHINX20^{20} simulation. Considering all galaxies with SFR10≥0.5_{10} \ge0.5 M⊙_\odotyr−1^{-1} at redshifts z≥6z \ge6 in the SPHINX20^{20} simulation, we show the global LyC escape fraction (left) and line-of-sight LyC escape fraction (right) as a function of nen_e([O II]). In each panel, we provide the best fit relation for MEGATRON as presented in Figure 12 and Equation (2). We find that this relation holds for the global escape fractions but disappears for line-of-sight escape fractions, owing to the highly anisotropic nature of LyC leakage.

References

  1. [2]Adams N. J. et al., 2024, ApJ, 965, 169
  2. [4]Atek H. et al., 2026, arXiv e-prints, arXiv:2604.23823DOI
  3. [6]Austin D. et al., 2023, ApJ, 952, L7
  4. [7]Austin D. et al., 2025, arXiv e-prints, arXiv:2512.10839arxiv.org/abs/2512.10839
  5. [8]Barrow K. S. S., Robertson B. E., Ellis R. S., Nakajima K., Saxena A., Stark D. P., Tang M., 2020, ApJ, 902, L39
  6. [11]Berg D. A. et al., 2026, ApJ, 1003, 112
  7. [16]Cadiou C. et al., 2026, The Open Journal of Astrophysics, 9, 69640
  8. [17]Carr C. et al., 2025a, arXiv e-prints, arXiv:2510.21197arxiv.org/abs/2510.21197
  9. [19]Castellano M. et al., 2025, arXiv e-prints, arXiv:2504.05893DOI
  10. [20]Castellano M. et al., 2023, ApJ, 948, L14
  11. [21]Chemerynska I. et al., 2025, arXiv e-prints, arXiv:2509.24881arxiv.org/abs/2509.24881
  12. [22]Chemerynska I. et al., 2023, arXiv e-prints, arXiv:2312.05030DOI
  13. [25]Chisholm J. et al., 2022, MNRAS, 517, 5104
  14. [26]Choustikov N. et al., 2026, The Open Journal of Astrophysics, 9, 58199
  15. [27]Choustikov N. et al., 2024a, MNRAS, 529, 3751
  16. [28]Choustikov N. et al., 2024b, MNRAS, 532, 2463
  17. [36]Donnan C. T. et al., 2023, MNRAS, 518, 6011
  18. [37]Donnan C. T. et al., 2024, MNRAS, 533, 3222
  19. [48]Finkelstein S. L. et al., 2019, ApJ, 879, 36
  20. [49]Finkelstein S. L. et al., 2024, ApJ, 969, L2
  21. [50]Flury S. R. et al., 2022a, ApJS, 260, 1
  22. [54]Giovinazzo E. et al., 2026b, A&A, 707, A352
  23. [60]Harris C. R. et al., 2020, Nature, 585, 357
  24. [61]Harshan A. et al., 2024, MNRAS, 532, 1112
  25. [67]Isobe Y. et al., 2026, arXiv e-prints, arXiv:2606.11345arxiv.org/abs/2606.11345
  26. [73]Jaskot A. E. et al., 2024a, ApJ, 972, 92
  27. [75]Jecmen M. C. et al., 2026, ApJ, 1001, 234
  28. [81]Kannan R. et al., 2025, arXiv e-prints, arXiv:2502.20437DOI
  29. [83]Katz H. et al., 2025, The Open Journal of Astrophysics, 8, 104
  30. [84]Katz H. et al., 2026a, The Open Journal of Astrophysics, 9, 69642
  31. [85]Katz H. et al., 2022a, MNRAS, 515, 4265
  32. [87]Katz H. et al., 2022b, arXiv e-prints, arXiv:2211.04626arxiv.org/abs/2211.04626
  33. [88]Katz H. et al., 2026b, The Open Journal of Astrophysics, 9, 69643
  34. [90]Katz H. et al., 2023a, The Open Journal of Astrophysics, 6, 44
  35. [91]Katz H. et al., 2023b, MNRAS, 518, 270
  36. [99]Kim K. J. et al., 2023b, ApJ, 955, L17
  37. [104]Komarova L. et al., 2025, arXiv e-prints, arXiv:2506.19623DOI
  38. [108]Laseter I. H. et al., 2024, A&A, 681, A70
  39. [109]Leitherer C. et al., 1999, ApJS, 123, 3
  40. [110]Li S. et al., 2025, ApJ, 979, L13
  41. [117]Maji M. et al., 2022, A&A, 663, A66
  42. [121]Martinez Z. et al., 2025, ApJ, 995, 204
  43. [122]Mascia S. et al., 2024, A&A, 685, A3
  44. [123]Mascia S. et al., 2023, A&A, 672, A155
  45. [128]McLeod D. J. et al., 2026, arXiv e-prints, arXiv:2604.16666arxiv.org/abs/2604.16666
  46. [129]Mehrotra P. et al., 2026, arXiv e-prints, arXiv:2610.30265arxiv.org/abs/2610.30265
  47. [135]Naidu R. P. et al., 2022, ApJ, 940, L14
  48. [139]Nakajima K. et al., 2022, ApJS, 262, 3
  49. [145]Pahl A. et al., 2025, ApJ, 981, 134
  50. [148]Papovich C. et al., 2026, ApJ, 1000, 111
  51. [156]Rey M. et al., 2026, The Open Journal of Astrophysics, 9, 69605DOI
  52. [157]Rey M. P. et al., 2025, MNRAS, 541, 1195
  53. [162]Rosdahl J. et al., 2022, MNRAS, 515, 2386
  54. [163]Rosdahl J. et al., 2018, MNRAS, 479, 994
  55. [166]Saldana-Lopez A. et al., 2025, arXiv e-prints, arXiv:2504.07074
  56. [167]Sanders R. L. et al., 2026, ApJ, 1003, 228
  57. [169]Saxena A. et al., 2024a, A&A, 684, A84
  58. [172]Saxena A. et al., 2022, MNRAS, 517, 1098
  59. [173]Saxena A. et al., 2020, A&A, 636, A47
  60. [177]Schaerer D. et al., 2022, A&A, 658, L11
  61. [181]Simmonds C. et al., 2024a, MNRAS, 527, 6139
  62. [182]Simmonds C. et al., 2024b, MNRAS, 535, 2998
  63. [183]Simmonds C. et al., 2023, MNRAS, 523, 5468
  64. [195]Stephenson M. G. et al., 2026, arXiv e-prints, arXiv:2608.23846arxiv.org/abs/2608.23846
  65. [198]Summerfield Z. et al., 2026, arXiv e-prints, arXiv:2606.14671arxiv.org/abs/2606.14671
  66. [200]Tang M. et al., 2023, MNRAS, 526, 1657
  67. [204]Topping M. W. et al., 2025, arXiv e-prints, arXiv:2502.08712arxiv.org/abs/2502.08712
  68. [205]Topping M. W. et al., 2024, MNRAS, 529, 3301
  69. [211]van der Velden E., 2020, The Journal of Open Source Software, 5, 2004
  70. [212]van der Walt S., Colbert S. C., Varoquaux G., 2011, Comput. Sci. Eng., 13, 22
  71. [215]Virtanen P. et al., 2020, Nature Methods, 17, 261
  72. [216]Wang B. et al., 2021, ApJ, 916, 3
  73. [218]Whitler L. et al., 2025, ApJ, 992, 63
  74. [226]Yung L. Y. A., Somerville R. S., Finkelstein S. L., Wilkins S. M., Gardner J. P., 2024, MNRAS, 527, 5929
  75. [227]Yusof N. et al., 2013, MNRAS, 433, 1114

Paper details

vlnc:W_myhqz5dk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents