Lumina simulates the universe’s first 2 billion years at record scale

The story

A new simulation called Lumina has followed the first 2 billion years of cosmic evolution across a virtual region 500 megaparsecs — about 1.6 billion light-years — wide. The peer-reviewed project, published October 26 in Monthly Notices of the Royal Astronomical Society, combines roughly half a trillion dark-matter particles and gas cells with models of star formation, black-hole growth and radiation. The scale matters because early galaxies are small, while the ionized regions they collectively create can stretch across enormous distances.

Lumina begins about 45 million years after the Big Bang, when neutral hydrogen made the universe opaque to much ultraviolet light. It then calculates how gravity draws dark matter and gas into filaments and haloes, how galaxies and stars form inside them, and how radiation travels through the surrounding material. The run ends around redshift 3, when the universe was about 2 billion years old. It is a model of how the early universe could have evolved, not a recording of what happened.

The calculation links two transitions that are often modeled separately. During hydrogen reionization, ultraviolet light from young stars stripped electrons from neutral hydrogen and gradually made intergalactic space transparent. Later, harder radiation associated with accreting supermassive black holes ionized singly ionized helium. In Lumina, stars dominate the first process and active galactic nuclei drive the second. The paper reports hydrogen reionization largely complete near redshift 5.2, with residual neutral patches persisting longer, while helium reionization approaches completion by redshift 3.

That result is not simply a consequence of adding more computer power. The team built on the IllustrisTNG galaxy-formation framework and the moving-mesh code AREPO, then coupled it to radiation transport that tracks different photon-energy groups. This lets the same calculation connect small-scale processes inside galaxies with the expanding ionized bubbles between them. At peak scale on the Frontier supercomputer at Oak Ridge National Laboratory, MIT says the work brought together computing equivalent to more than 100,000 CPU cores and about 15,000 GPUs.

The volume is about 100 times larger than earlier high-resolution reionization simulations cited by the researchers, while the element count is about 30 times higher than IllustrisTNG. A bigger box samples rare luminous objects and representative cosmic environments; finer resolution gives the model more credible galaxies. Using both at once is what makes population-level comparisons possible. The paper’s multiscale visualizations can move from the 500-megaparsec cosmic web to individual developing galaxies without switching to a separate simulation.

Observations are the test. The James Webb Space Telescope is already finding unexpectedly bright and compact objects in the young universe, while the Nancy Grace Roman Space Telescope is designed to survey much larger areas. Lumina can generate populations of simulated galaxies and ionized regions under a consistent set of physical assumptions. Researchers can apply an instrument’s selection rules to that synthetic universe, then ask whether the predicted counts, clustering, spectra and reionization history resemble the sky.

The caveat is that no calculation resolves every star, black-hole accretion disk or escape path for ionizing photons across a volume this large. Lumina therefore uses sub-grid prescriptions for unresolved physics, including star formation, feedback and radiation escaping galaxies. Its active-galactic-nucleus population and photon production are calibrated choices. Agreement with broad constraints is encouraging, but it does not prove that every mechanism or timeline is correct; different assumptions can sometimes produce similar global histories.

Independent reporting by The Times described Lumina as the most detailed look yet at the universe’s infancy, while the paper itself supplies the more useful standard: a specific numerical experiment whose outputs can be compared with data. Claims such as ‘first’ or ‘largest’ depend on which combination of volume, resolution and included physics is measured. The defensible achievement is that one run connects galaxy formation with both hydrogen and helium reionization across an unusually large representative volume.

INNOVOX analysis: Lumina is best understood as a forward model for the next generation of cosmology, not a digital reconstruction. Its innovation is to produce a synthetic sky large enough for survey statistics while retaining enough local detail to test astrophysical explanations. If the collaboration releases reproducible catalogues, light cones and ionization maps, other teams can challenge the assumptions instead of merely admiring the images. The discrepancies could be more valuable than the matches, because they would identify missing physics or biased interpretations of early galaxies.

What to watch is the conversion of compute into falsifiable predictions. Near-term tests include Webb galaxy counts and spectra, Roman’s wide-field samples, and 21-centimeter measurements designed to trace neutral hydrogen during reionization. The project also points to forecasts for gravitational-wave backgrounds from early galaxies and black holes. Lumina has made an unusually expansive hypothesis about the cosmic dawn calculable; telescopes now have to decide how much of that virtual universe survives contact with the real one.

INNOVOX analysis

Lumina’s important advance is the union of representative cosmic volume, galaxy-scale detail and radiation transport. That combination can turn a simulation into a falsifiable synthetic survey, provided its assumptions, data products and selection effects remain open to independent testing.

What to watch

Watch for the promised community data release, tests against large galaxy samples from the James Webb Space Telescope, and forecasts for the Nancy Grace Roman Space Telescope and 21-centimeter experiments. The most revealing outcome may be where the simulated universe fails to match the observed one.