Can machine learning speed up cosmic dawn observations 100,000 times over?

Simulating the universe's first galaxies takes weeks of computing time. This hybrid machine learning model—using diffusion and LSTM networks—reproduces seven key observables from 21cmFAST simulations in milliseconds with sub-percent error. Researchers applied it to HERA's current limits on the cosmic 21cm power spectrum, constraining X-ray emission from early star-forming galaxies and predicting detection prospects for the Square Kilometre Array.