Lumina Simulation Recreates the First 2.2 Billion Years of the Universe at Record Scale

Since the James Webb Space Telescope (JWST) began observing the deep universe, astronomers have faced a new puzzle. In the early universe, only a few hundred million years after the Big Bang, one candidate galaxy after another has turned up that appears brighter or more massive than conventional theory predicts. But the light of distant galaxies captured by a telescope cannot, on its own, reveal the full picture: what kind of gas filled the universe at that time, and how the first stars affected the surrounding intergalactic space. To make sense of the observations, we need large-scale numerical simulations that reproduce, from physical laws, the motion of gravity and fluids and the interaction between gas and the light emitted by stars and galaxies.

A research team led by Oliver Zier at the Harvard-Smithsonian Center for Astrophysics has developed "Lumina," a cosmological radiation-hydrodynamics simulation that simultaneously tracks the large-scale structure of the early universe and the ionization of its gas. The results were published on October 26, 2026 in the journal Monthly Notices of the Royal Astronomical Society (MNRAS, Vol. 552, Issue 4, DOI: 10.1093/mnras/stag1469). Lumina is a numerical model that reproduces the history of "cosmic reionization," in which hydrogen and helium in the early universe were ionized by light from the first stars and active galactic nuclei (AGN), on a spatial scale far larger than previous efforts.

However, the universe that Lumina depicts is not a past state reconstructed directly from observational data. It is a theoretical prediction, produced by combining calculations based on physical laws with various assumptions and numerical approximations. What matters is that it can generate statistical predictions for a vast region of the universe that can be compared with results from the latest space telescopes.

The Puzzle of Cosmic Reionization and the Limits of Simulation

According to the standard cosmology, about 380,000 years after the Big Bang the universe cooled enough for electrons and nuclei to combine into neutral atoms, an event known as recombination (the "clearing of the universe"). The era that followed, until the first stars were born, is called the "Dark Ages," when no light sources such as stars yet existed. Eventually gravity gathered gas together and the first stars and galaxies formed, and the ultraviolet light they emitted began to spread into the surrounding space. Powerful radiation produced by gas accreting onto giant black holes added to this, and electrons were stripped from the neutral gas spread through intergalactic space. This process is "cosmic reionization."

Reionization did not happen all at once; it is thought to have passed through different stages for hydrogen and helium. Hydrogen reionization proceeded mainly through ultraviolet light emitted by stars in early galaxies. Removing the second electron from helium, on the other hand, requires photons with energies of at least 54.4 eV, and light from ordinary stars alone would struggle to drive this process to completion across the universe. For that reason, the reionization of helium from singly ionized (He II) to doubly ionized (He III) is thought to have been caused mainly by the intense high-energy radiation of AGN. What astronomers want to determine is when these changes occurred, which objects drove them, and how fast they proceeded.

Reproducing this phenomenon on a computer requires satisfying conflicting demands. To capture the whole picture of reionization, a huge computational domain spanning hundreds of millions of light-years is needed. At the same time, accurately calculating the behavior of the galaxies that produce photons, and of the dense gas that absorbs their light, demands sufficiently fine resolution. Adding "radiative transfer," which calculates the propagation and absorption of photons, on top of the hydrodynamics that solves gas motion makes the computational cost soar. In earlier studies, the trade-off between widening the domain and sharpening the detail was a major constraint.

Lumina greatly eases this constraint, following the evolution of the universe from redshift $z=49$ (about 50 million years after the Big Bang) to $z=3$ (about 2.2 billion years). Redshift is a measure of how much the wavelength of light has been stretched by cosmic expansion; larger values correspond to the more distant past. This span covers everything from hydrogen reionization led by stars in early galaxies to the near-completion of helium reionization led by AGN. The aim of the research is not to reproduce the Big Bang itself, but to clarify how the galaxies and black holes that were born changed the space around them.

A 500 cMpc Simulation Enables Statistical Analysis of the Universe

Lumina's greatest feature is the breadth of its computational domain. The team set up a cube 500 cMpc (comoving megaparsecs, about 1.63 billion light-years) on a side. Comoving distance is a scale that removes the change in distance caused by cosmic expansion, and must be distinguished from the actual physical distance at each epoch.

Within this vast volume, Lumina places $2\times 6000^3$, about 432 billion, resolution elements. The calculation uses "AREPO," which solves hydrodynamics on a moving unstructured mesh, and incorporates the existing "IllustrisTNG" physics model for galaxy formation. For radiative transfer, it uses a GPU-accelerated solver based on an approximation called the M1 closure, tracking the radiation's energy in multiple frequency bands. According to the Smithsonian's announcement, the computation was run on the "Frontier" supercomputer at Oak Ridge National Laboratory in the United States. According to the paper, it used 105,504 CPU cores and 7,536 AMD Instinct MI250X accelerators across 1,884 nodes, equivalent to roughly 15,000 GPU dies.

Comparison with "Thesan-1," a leading earlier cosmological radiation-hydrodynamics simulation, shows Lumina's scale. Thesan-1's domain was 95.5 cMpc on a side, with $2\times 2100^3$ (about 18.5 billion) resolution elements. Lumina covers a volume about 143.5 times larger.

{
  "type": "bar",
  "title": "主要な宇宙論的輻射流体シミュレーションの計算体積比較",
  "unit": "×10^6 cMpc^3",
  "categories": ["Thesan-1 (2022)", "Lumina (2026)"],
  "series": [
    { "name": "シミュレーション体積", "values": [0.87, 125.0] }
  ],
  "caption": "一辺の共動長（95.5 cMpc対500 cMpc）から算出した計算体積",
  "source": "MNRAS Vol. 552, Issue 4 (2026)"
}

This expansion in volume means more than simply competing on computing scale. The universe contains objects, such as extremely bright quasars, that can only be found in sufficient numbers by surveying a very large volume. When examining the distribution of the huge ionized regions formed by reionization, or large-scale fluctuations in the 21 cm line emitted by neutral hydrogen, a computational domain that is too small biases the results. By calculating a vast region at once, Lumina makes it possible to evaluate the statistical properties of the universe, including these rare objects and large-scale structures, more appropriately than before.

That said, a 143.5-fold increase in computational volume does not mean Lumina surpasses Thesan-1 in every respect. Lumina's mass resolution is $3.6\times 10^6{M}_{\odot }$ (solar masses) for baryons, the ordinary matter, and $1.9\times 10^7{M}_{\odot }$ for dark matter. Thesan-1's baryon mass resolution, by contrast, is about $5.8\times 10^5{M}_{\odot }$, meaning each gas element represents about one-sixth the mass of Lumina's. In other words, Thesan-1 can track small-scale structure in finer detail in terms of mass resolution, while Lumina is designed to prioritize volume. Fine-grained physical phenomena around individual stars and black holes depend on "subgrid models," which treat processes below the resolvable scale in an approximate way.

The main differences in specifications, based on the comparison table in the research paper, are as follows.

In the visualization of the universe's large-scale structure released by the team, fine web-like filaments represent the distribution of dark matter, while orange and yellow regions show places where gas moves at high speed between galaxy clusters. These are not photographs taken by a telescope, but scientific visualizations in which physical quantities obtained from the simulation are translated into color and brightness.

Lumina's Predicted History of Hydrogen and Helium Reionization

Lumina offers concrete predictions about how the reionization of hydrogen and helium proceeded in the early universe. However, these do not mean that new observations have pinned down the timing of reionization. They are simply the results of numerical calculations based on the physical model the team adopted.

Hydrogen reionization, driven mainly by star formation in early galaxies, was completed relatively late in the simulation. The team divided the simulation domain into 125 sub-regions of 100 cMpc on a side and examined the ionization progress of each. In the median region, hydrogen reionization was complete by redshift $z\approx 5.2$ (about 1 billion years after the Big Bang). Progress varied widely from region to region, however, and regions of neutral hydrogen (H I) that remained un-ionized to the end persisted until redshift $z\approx 4.75$ (about 1.2 billion years).

In helium reionization, AGN play a more important role than stars. Lumina tracked the high-energy photons emitted in association with the activity of supermassive black holes and calculated the process by which they ionize He II into He III. As a result, helium reionization progressed gradually as the universe evolved and was nearly complete by $z=3$ (about 2.2 billion years), the end point of the simulation.

The team also confirmed that the temperature of the intergalactic medium rises again with helium reionization. When high-energy photons ionize He II, the surplus energy released heats the surrounding gas. In the early universe, temperatures jumped sharply with hydrogen reionization, then the gas cooled through cosmic expansion and other effects, and was heated again as helium reionization began. Lumina showed that it can calculate these multi-stage temperature changes of the intergalactic medium in connection with the ionization histories of hydrogen and helium.

This is also an important difference from earlier large-scale reionization simulations. Thesan-1, for example, focused on the era in which hydrogen reionization proceeds and did not go as far as self-consistently tracking the reionization of He II to He III. Lumina treats hydrogen reionization by stars in early galaxies and the subsequent helium reionization by AGN continuously within the same region of the universe.

Reducing the Speed of Light to One-Fifth: Numerical Approximations and Model Limits

Lumina can calculate a vast region of the universe over a long period of evolution, but its results must be interpreted with the effects of approximations introduced to hold down computational cost in mind. Two matter especially: the "Reduced Speed of Light Approximation" used in radiative transfer, and the switch of physical models partway through the calculation.

The first constraint, the reduced speed of light approximation, is a technique that speeds up computation by setting the propagation speed of photons lower than the real value. Using the actual speed of light $c$ would require repeating the calculation in extremely short time steps matched to the time light takes to cross the computational grid, demanding enormous computing resources. To reduce this load, Lumina sets the speed of light in its radiative transfer solver to $0.2c$, 20% of the real value. The team acknowledges that this approximation may affect the speed at which the boundaries of ionized regions expand and the amount of neutral hydrogen remaining in the final stages of reionization. In low-density regions such as voids in particular, there is concern that systematic bias may arise in the results.

The second constraint is that the calculation method changes at redshift $z=4.75$. Before that point, radiation from stars, AGN and other sources is divided into multiple energy bands, and ionization and heating of hydrogen and helium are calculated in coupling with gas motion. After $z=4.75$, however, to reduce computational load, ionization and heating by photons below 54.42 eV are replaced with a spatially uniform ultraviolet background radiation model. At the same time, the radiation tracked explicitly is narrowed to high-energy photons originating from AGN, and the calculation switches to a scheme centered on He II reionization. Consequently, not all light sources and ionization processes are tracked with the same accuracy throughout the entire simulated period.

Despite these constraints, the team has verified the validity of the model by comparison with observational data. In particular, the "Thomson scattering optical depth," which expresses how strongly the cosmic microwave background (CMB) is scattered by free electrons in space, agreed very well with the value measured by the Planck satellite. The temperature evolution of the intergalactic medium and the intensity of the hydrogen-ionizing background radiation were also broadly consistent with existing observational constraints.

However, agreement in optical depth alone does not prove that the fine physical processes of galaxy formation and reionization are correctly reproduced. Lumina's galaxy formation relies on IllustrisTNG's subgrid models, and phenomena that cannot be calculated directly, such as star formation efficiency, supernova feedback, and black hole growth and energy release into the surroundings, involve modeling assumptions. The team also ran two kinds of smaller simulations to examine the effects of X-ray preheating, as well as a lower-resolution dark-matter-only calculation. These, however, are comparison runs for investigating the effects of specific numerical effects or physical models, and do not mean that the results of Lumina as a whole have been reproduced by an independent research team.

Linking JWST and Next-Generation Observations with Theory

Lumina's key achievement is not a final answer to the mystery of cosmic reionization, but the establishment of a computational foundation for detailed comparison between theoretical predictions and observational data. At its center is a technique called "forward modeling."

In forward modeling, conditions such as a telescope's field of view, sensitivity and observing wavelengths are applied to the distributions of galaxies and gas obtained from a theoretical model, to predict what data would actually be obtained if observed. As JWST observations of early galaxies advance, and wide-field surveys by the Nancy Grace Roman Space Telescope (Roman) begin in the future, large amounts of data are expected on the number and brightness distribution of distant galaxies and how galaxies cluster together. Large-scale simulations like Lumina can generate mock observational data for directly comparing such results with theory.

If Lumina's predictions agree with observations, that would be evidence supporting the validity of the galaxy formation model and reionization scenario adopted. However, because different physical models may produce similar observational results, agreement cannot be taken to mean that a particular theory is the only correct one. Conversely, if clear discrepancies between predictions and observations are found, they would offer clues for revisiting assumptions built into the model, such as star formation efficiency and the radiation properties of AGN. Furthermore, if the discrepancies cannot be explained by modifying ordinary galaxy formation models alone, they could lead to consideration of more fundamental physics, including the nature of dark matter.

The team plans complementary smaller-scale, higher-resolution simulations to examine in detail how the relative velocity between baryons and dark matter (the streaming velocity) in the early universe affects star formation. It also intends to continue analyzing the flow of gas between galaxies and the evolution of regions of active star formation. At this stage, however, Lumina has not resolved the discrepancy between JWST observations of early galaxies and theory, nor has it verified future Roman observations.

Using about 432 billion resolution elements to follow the evolution of the universe up to roughly 2.2 billion years after the Big Bang, Lumina becomes a new computational foundation for studying the early universe. The focus going forward is how closely the distribution of ionized regions and the temperature changes of the intergalactic medium predicted by the model match actual observations. Comparison with radio observations probing the 21 cm line of neutral hydrogen, and with observational data on distant galaxies and quasars, will be especially important in evaluating Lumina's predictive accuracy and in unraveling the history of cosmic reionization.