Can AI generate realistic 3D maps of the early universe's radio glow?

Researchers used conditional diffusion models—the same technology behind image generation—to synthesize three-dimensional maps of 21 cm radiation from the early universe. The challenge: 3D data is massive, skewed, and memory-hungry. They tested preprocessing methods and found that data normalization matters more than model depth; Yeo-Johnson compression plus careful scaling produced the most faithful outputs across brightness maps, power spectra, and global signals. The emulations look plausible but retain measurable biases in higher-order statistics, marking a baseline for accelerating expensive cosmological simulations.