How to Understand the Early Universe

The first data sent back by the James Webb Space Telescope (JWST) in 2022 revealed far more galaxies than expected. Brant Robertson's team at UC Santa Cruz faced the reality that understanding the early universe would require interpreting terabytes of light that had crossed the universe for over 1.3 billion years.

The team built an analysis pipeline combining AI and GPU. It automates classification, data reduction, catalog generation, anomaly detection, and simulation, with development handled on the campus's Lux cluster and NVIDIA DGX Station, and large-scale execution processed on government supercomputers. Underpinning this is a $1.6 million NSF grant.

The core model, Morpheus, uses semantic segmentation to distinguish bulges and disks at the pixel level rather than treating galaxies as a whole. When applied to JWST's larger, more precise images, this model produced results showing that rotating disk galaxies—thought not to have existed in the early universe—were actually present early on, a finding that has since been independently confirmed multiple times.

• GalaxyFriends UMAP, created by Anavi Uppal, groups about 90,000 galaxies into similarity-based neighborhoods, making similar objects and outliers visible at a glance.
• For work related to the Vera C. Rubin Observatory, the team is training AI to remove distortions from Earth's atmosphere and, in a manner similar to NVIDIA DLSS, restore sharper images. Rubin is expected to produce about 20TB of raw data every night.
• This effort extends to NASA's Nancy Grace Roman Space Telescope and the proposed Habitable Worlds Observatory, with the team validating observations through large-scale cosmological simulation and releasing data on nearly 500,000 galaxies for anyone to use.

Computation is transforming the early universe from something only experts can interpret into a data asset that everyone can use.