An emulator for the ionizing photon mean free path in ultra-high resolution simulations: the implications of mean free path measurements for the reionization history

Measurements of the mean free path of ionizing photons from high-redshift quasar spectra at $z \sim 5$-$6$ constrain the reionization history, but interpreting them requires modeling the kiloparsec-scale clumping that large-volume reionization simulations cannot resolve. We present a deep learning emulator for the mean free path (MFP) trained on high-resolution cosmological radiative transfer simulations of ionization fronts sweeping through small 2 comoving~Mpc/h volumes. Using a residual multi-layer perceptron neural network, we predict the MFP at a given redshift as a function of the reionization redshift, photoionization rate, wavelength, and box-scale density, achieving a median relative error of 1.6% across nearly four orders of magnitude in MFP. Integrating its predictions over box-scale overdensity and an extended reionization history allows the emulator to predict the global MFP. We apply the emulator to extended reionization histories constrained by observed photoionization rates, finding that models prefer late reionization with substantial neutral fractions persisting at $z \lesssim 6$. Fitting a parametric ionization history yields a midpoint of reionization of $z_{\rm re} = 6.8\pm 1.2$ for reionization durations consistent with Planck and kinetic Sunyaev-Zeldovich constraints, and the universe being $10%$ neutral still at $z_{\rm re} < 5.8 (6.3)$ at 1(2)$σ$. Global ionizing emissivity inferences using measurements of the photoionization rate and MFP plus our emulator, which avoids common power-law assumptions, suggest a factor of $2-3$ decline between $z = 6$ and $4.8$, in agreement with previous studies. Our method provides an efficient (and more converged) alternative to large-volume radiative-hydrodynamic simulations of reionization for interpreting MFP measurements, and can also serve as a subgrid prescription for the ionizing opacity within such simulations.

💡 Research Summary

This paper presents a novel deep‑learning emulator for the ionizing photon mean free path (MFP) that bridges the gap between the kiloparsec‑scale clumping physics required to interpret high‑redshift quasar spectra and the gigaparsec‑scale volumes needed for reionization studies. The authors run a suite of ultra‑high‑resolution cosmological radiation‑hydrodynamic simulations in 2 cMpc h⁻¹ boxes with 1024³ cells, achieving a spatial resolution of 2 h⁻¹ ckpc. These simulations resolve the Jeans scale, capture self‑shielding, and model ionization fronts propagating through gas with a power‑law ionizing spectrum (Jν ∝ ν⁻¹) across six photon energy bins (13.6–39.5 eV). By varying three key environmental parameters—reionization redshift (z₍re₎ = 5–15), hydrogen photo‑ionization rate (Γ₋₁₂ = 0.03–30), and large‑scale overdensity (δ/σ = 0, ±√3)—they generate 126 distinct simulations that span the full dynamic range of MFP (≈0.1–1000 cMpc). To account for fluctuations on scales larger than the box, they employ a “DC mode” approach combined with three‑point Gauss‑Hermite quadrature, weighting the three overdensity runs to reproduce the cosmic density distribution.

From each simulation they compute the MFP using the flux‑weighted distance definition of Becker et al. (2013), averaging over 10 000 random sightlines and confirming consistency with two alternative estimators (absorption‑rate based and segment‑based). The resulting dataset provides MFP as a function of redshift, z₍re₎, Γ₋₁₂, photon energy, and overdensity.

The emulator is a residual multi‑layer perceptron (ResMLP). Inputs are the five physical parameters; the target is log₁₀(λₘ𝚏ₚ). Log‑transforming compresses the four‑order‑of‑magnitude range, and the loss function is the Huber loss, which is robust to outliers. The network architecture consists of an initial dense layer mapping the five inputs to 128 hidden units, followed by four residual blocks each containing two fully‑connected layers with layer‑normalization and Mish activations, plus a skip connection. A dropout of 0.1 mitigates over‑fitting. Inputs are scaled with scikit‑learn’s RobustScaler (median and IQR) to reduce sensitivity to extreme values. Training uses AdamW (lr = 0.005, weight decay = 0.001) with a cosine‑annealing schedule and warm restarts, for up to 5 000 epochs with early stopping (patience = 450). The data split is 80 % training, 10 % validation, 10 % test. On the held‑out test set the emulator achieves R² = 0.95 and a median relative error of 1.6 %, accurately reproducing MFP across the entire dynamic range. Validation on a simulation with z₍re₎ = 6.5 (not in the training set) yields a 1.7 % error, demonstrating strong generalization.

To obtain a global (volume‑averaged) MFP, the emulator predictions for each overdensity are weighted according to the Gauss‑Hermite quadrature coefficients and summed, effectively integrating over the large‑scale density field. This yields a redshift evolution of the mean free path that matches observational measurements from quasar spectra at z ≈ 5–6, including the rapid decline near the end of reionization.

The authors then combine the emulator‑derived MFP with observed hydrogen photo‑ionization rates (Γ) to infer the global ionizing emissivity ε(z). Unlike many previous works that assume a simple power‑law dependence of opacity on frequency (λ ∝ ν⁻³), the emulator retains the full wavelength dependence derived from the simulations. The resulting emissivity shows a decline by a factor of 2–3 between z = 6 and z = 4.8, consistent with earlier studies but derived without imposing a power‑law opacity model.

For reionization history constraints, the authors adopt a parametric tanh model for the ionized fraction, characterized by a midpoint redshift z₍re₎ and a duration Δz. Using a Markov Chain Monte Carlo (MCMC) analysis that simultaneously fits the observed Γ(z) and MFP(z) (with the emulator providing the theoretical MFP for any set of parameters), they find a preferred reionization midpoint of z₍re₎ = 6.8 ± 1.2. The duration is compatible with Planck CMB optical depth measurements and kinetic Sunyaev‑Zel’dovich (kSZ) constraints. Notably, the model predicts that the universe remains ≈10 % neutral down to z < 5.8 at 1σ (z < 6.3 at 2σ), indicating a relatively late and extended reionization process.

The paper concludes that the deep‑learning emulator offers an efficient, highly accurate alternative to running thousands of expensive radiative‑hydrodynamic simulations when exploring reionization parameter space. It can be used as a sub‑grid prescription for ionizing opacity in large‑volume simulations, dramatically reducing computational cost while preserving the essential small‑scale physics. The authors acknowledge limitations: the training data are confined to 2 cMpc boxes (potential boundary effects), the assumption of instantaneous, spatially uniform reionization within each box, and the current neglect of correlations among input variables. Future work could expand the training set to larger volumes, incorporate more realistic, patchy reionization histories, and develop architectures that model joint variable correlations. Nonetheless, the presented emulator constitutes a powerful tool for interpreting MFP measurements and constraining the timing and duration of cosmic reionization.