MEGATRON: Reproducing the Diversity of High-Redshift Galaxy Spectra with Cosmological Radiation Hydrodynamics Simulations

MEGATRON presents a suite of cosmological radiation hydrodynamics simulations designed to reproduce the diversity of high-redshift galaxy spectra, which is crucial for testing models against JWST observations. This work introduces the first set of cosmological simulations that couples a vast non-equilibrium thermochemistry network with on-the-fly radiation transport to directly predict the spectral properties of early galaxies. By initializing these simulations at zero metallicity and resolving haloes well below the atomic cooling threshold, MEGATRON aims to address key science themes such as star formation at cosmic dawn, galaxy formation during reionization, and the circumgalactic medium towards cosmic noon.

The MEGATRON Suite Overview

The MEGATRON suite consists of seven cosmological radiation hydrodynamics simulations run in the Lagrange region around a Milky Way-mass environment. These simulations utilize the MEGATRON galaxy formation model developed within the RAMSES-RTZ adaptive mesh refinement (AMR) code, which is unique for its connection to on-the-fly, multi-frequency radiation transport and its ability to reproduce equilibrium results computed with 1D photoionization codes while maintaining a non-equilibrium aspect. The project is designed to advance understanding in four key science areas:

1. Predicting the demographics and spectral properties of the first generation of Population III (Pop. III) stars.

2. Quantifying how observable ISM properties respond to different galaxy and stellar evolution models at low metallicity.

3. Studying how non-equilibrium chemistry, local radiation fields, and mass assembly affect the emission and absorption observables of the circumgalactic medium (CGM) at cosmic noon.

4. Deciphering the archaeological traces left by high-redshift galaxy formation physics in a local volume environment of a Milky Way-mass galaxy.

Numerical Methods and Initial Conditions

The simulations are constructed using cosmological, zoomed initial conditions (ICs) for a Milky Way-mass halo with a virial mass of approximately 1012 M⊙ at z = 0. These ICs are generated using the genetIC software and a flat Planck Collaboration cosmology. To maximize the sample size and ensure adequate resolution, the researchers employ a quadratic genetic modification technique to modify progenitor collapse histories, such as the 'Early Forming' IC used for the high-redshift suite. The simulations utilize a constant comoving spatial resolution for the high-redshift suite, allowing refinement up to 1.7 pc h−1 when Pop. III stars form at z ∼ 27 and 5 pc h−1 at z = 8.5, while the cosmic noon suite employs a constant physical resolution of ≈ 20 pc h−1.

Physical Modeling and Chemistry

The core innovation of MEGATRON is the coupling of detailed physics networks. The simulations employ an explicit modeling of Pop. III star formation from zero-metallicity initial conditions facilitated by non-equilibrium H2 cooling and a high spatial resolution (≈pc). A non-equilibrium chemistry model for primordial species, metals, and molecules is included, which is not generally captured by many previous high-redshift simulations. Heating and cooling processes include photoheating, photoelectric heating, H2 formation heating/cooling, CO cooling, dust recombination cooling, and metal line cooling. Stellar feedback includes an energetic feedback model calibrated on Milky Way-mass galaxies at z = 0 that produces a realistic stellar mass-halo mass relation at z = 0.

Simulation Suites and Results

The suite comprises four high-redshift simulations (Efficient SF, Bursty SF, Variable IMF, HN, High εff) and three cosmic noon simulations (Early Collapse, Fiducial Collapse, Late Collapse). Across the high-redshift suite of >175,000 spectra are produced. The results demonstrate how the diversity of galaxy spectra seen by JWST is naturally reproduced in a ΛCDM cosmology. Key findings include:

- Pop. III galaxies and cooling haloes that have strong emission H and He emission lines, no (or very weak) metal lines, and a dominant nebular continuum.

Star Formation Histories of EELGs

The simulations produce numerous Emission Line Galaxies (EELGs), defined as those with an Hα or [O III] λ5007 EW > 750 Å. The fraction of bright galaxies that would be considered EELGs are similar across all runs at ∼ 50%. A key result is that in the variable IMF simulation, [O III] EWs can reach > 4000 Å for galaxies in the magnitude range ≲ −19 undergoing extreme bursts of star formation.

Spectral Diversity and Galaxy Types

Dimensionality reduction (UMAP8) of the >175,000 spectra reveals several galaxy types:

1. Pop. III galaxies and cooling haloes with strong H and He emission lines, no (or very weak) metal lines, and a dominant nebular continuum.

As a fastidious and diligent researcher, I have analyzed the MEGATRON suite simulations to identify several key areas where advancements in AI systems could yield significant, measurable improvements in astrophysical discovery and modeling.

Here are the specific improvements for AI systems:

) 1. Supervised Learning for Spectral Classification of High-Redshift Galaxies (Based on Figure 8 & Section 4)

The MEGATRON suite produces >175,000 intrinsic spectra across four different galaxy classes (Pop. III, Mini-quenched, Post-starburst, Starforming).

• AI Improvement: Develop a deep learning classifier (e.g., a Variational Autoencoder or Convolutional Neural Network) trained on the UMAP embeddings and the associated spectral features (Hα/OIII Equivalent Widths, UV slope β, Balmer Jump strength).

• Improved AI Capability: The system can automatically classify new JWST spectra with high precision into one of the four physically motivated galaxy types. It can also quantify the probability of a spectrum belonging to a specific class (e.g., This spectrum is 92% likely to be an EELG). This drastically reduces the manual effort required for initial sample selection and allows for rapid, unbiased statistical studies across large JWST datasets.

) 2. Generative Modeling of Star Formation Histories (Based on Figure 13 & Section 4.2)

The simulations show that Hα EW is an extremely strong predictor of the recent Star Formation History (SFH), with extreme changes in SFR over short timescales (50 Myr).

• AI Improvement: Train a recurrent neural network (RNN) or a Transformer model to ingest the observed emission line equivalent widths (Hα, [OIII]) and predict the entire SFH curve (SFR vs. time) for a given galaxy class.

• Improved AI Capability: The system can rapidly infer the recent starburst history of an observed galaxy with high fidelity, moving beyond simple instantaneous SFR measurements to understand the dynamic interplay between feedback and gas consumption in real-time.

) 3. Inverse Modeling for ISM Physics (Based on Section 2 & Appendix A)

The simulations couple non-equilibrium thermochemistry, radiation transport (RASCAS), and hydrodynamics.

• AI Improvement: Implement a Bayesian inference framework or a physics-informed neural network (PINN) to perform inverse modeling. The AI would take observed line ratios and continuum shapes as input and output the most likely underlying ISM parameters (e.g., electron temperature, metallicity, dust column density).

• Improved AI Capability: This allows researchers to move beyond simple photoionization models by inferring the complex, non-equilibrium thermal state of the gas in high-redshift galaxies directly from JWST data, bypassing the systematic uncertainties inherent in traditional photoionization modeling.

) 4. Parameter Space Exploration via Genetic Algorithms (Based on Section 1 & Fig. 6)

The MEGATRON suite demonstrates that galaxy properties are sensitive to fundamental subgrid models (IMF slope, feedback efficiency, star formation efficiency).

• AI Improvement: Use a Bayesian Optimization or Genetic Algorithm approach to systematically search the vast parameter space of subgrid physics (e.g., varying the feedback energy per SN, or the upper mass slope of the IMF) and map these variations to observable outcomes (like UV luminosity function shapes or stellar mass-halo mass relations).

• Improved AI Capability: The system can efficiently guide future simulations toward regions where new physical constraints are most likely to be found, accelerating the discovery of which specific subgrid physics parameters are most critical for reproducing observed galaxy populations.

) 5. Automated Identification of Exotic Populations (Based on Section 6 & Conclusion 1)

The study struggles to produce certain objects like massive mini-quenched systems and high-redshift nitrogen emitters.

• AI Improvement: Develop anomaly detection algorithms (e.g., using Isolation Forests or Autoencoders trained on the normal population) to flag spectra that deviate significantly from the main sequence, specifically targeting features indicative of these rare objects (like extremely low sSFR or unique spectral breaks).

• Improved AI Capability: The system can serve as an early warning system, automatically flagging potential candidates for rare classes like massive mini-quenched galaxies or Pop. III remnants that are currently missed by traditional selection methods.

By implementing these improvements, the resulting AI system will transition from a descriptive tool to a predictive and diagnostic engine capable of handling the spectral complexity and physical depth revealed by JWST, directly addressing the core challenge highlighted in Section 5: moving toward a detailed comparison between simulations and observations.

We present the MEGATRON suite of cosmological radiation hydrodynamics simulations following the formation of Milky Way-mass galaxies from the earliest cosmic epochs when Population III stars form to Cosmic Noon. The suite represents the first set of cosmological simulations that couples a vast non-equilibrium thermochemistry network of primordial species, metals, and molecules to multifrequency, on-the-fly radiation transport, allowing us to directly predict the spectral properties of early galaxies. By initializing the simulations at zero metallicity, resolving haloes well below the atomic cooling threshold, reaching parsec-scale resolution, and modeling a Milky Way-mass environment, we aim to address four key science themes: 1) Star formation at cosmic dawn, 2) Galaxy formation and the interstellar medium in the epoch of reionization, 3) The circumgalactic medium towards cosmic noon, and 4) Reionization in a local volume environment and near-field cosmology. In this introductory work, we present an overview of the physical characteristics of high-redshift MEGATRON galaxies and their environment at z>8. We present a library of >175,000 simulated galaxy spectra and demonstrate how much of the diversity of galaxy spectra seen by JWST is naturally reproduced in the context of a Λ CDM cosmology. Caveats are discussed, such as the lack of AGN in our simulations and the limitations of our adopted stellar population and chemical yield models. This project represents a step towards making more direct comparisons between simulations and observations and is particularly applicable for optimizing methods to infer galaxy properties from existing high-redshift JWST spectra and imaging data.