A team at Tsinghua University has developed an AI system that effectively triples the discovery rate of the universe’s earliest galaxies by removing noise from James Webb Space Telescope images. The technique, published in Science, found over 160 candidate galaxies from the first 500 million years after the Big Bang—a period astronomers call the Cosmic Dawn.
How ASTERIS Works
ASTERIS (Astronomical Spatiotemporal Enhancement and Reconstruction for Image Synthesis) takes a novel approach to astronomical noise reduction. Rather than processing images frame by frame, the system treats telescope data as a 3D spatiotemporal volume, learning to distinguish genuine astronomical signals from detector noise without human labeling.
The AI uses what researchers call a “photometric adaptive screening mechanism” combined with filtering that preserves scientific accuracy while revealing fainter objects. The result: ASTERIS can detect celestial sources roughly 2.5 times fainter than previous methods—a full magnitude deeper.
In practical terms, this means the AI needs less than two hours of Webb’s observation time to identify the same features that previously required more than a day and a half.
The Galaxy Harvest
Applied to images from the JWST Advanced Deep Extragalactic Survey (JADES), ASTERIS turned up 162 candidate high-redshift galaxies, tripling the number found using conventional approaches. The team validated 75% of these as genuine distant galaxies, yielding roughly 82 new discoveries from the universe’s earliest epoch.
These galaxies exist at redshifts between z≈9 and z≈16, meaning their light has traveled more than 13 billion years to reach us. For the first time, researchers obtained precise luminosity measurements for extremely faint objects at these distances.
What This Means
The technique addresses a fundamental constraint in deep-space astronomy: telescope time is finite and expensive. ASTERIS effectively multiplies the value of existing observations by extracting information that was always present but hidden in noise.
For astronomical surveys, this could mean discovering more distant objects without requiring longer exposure times. For our understanding of cosmic history, it opens a window into the earliest era of galaxy formation—when the first stars were lighting up a previously dark universe.
The Fine Print
The research comes from a cross-disciplinary team spanning Tsinghua’s automation and astronomy departments. Co-first authors Yuduo Guo, Hao Zhang, and Mingyu Li worked with corresponding authors Qionghai Dai, Zheng Cai, and Jiamin Wu.
While the results are promising, the 162 candidates are still candidates—confirmation requires spectroscopic follow-up. The 75% validation rate suggests the AI is reliable but not perfect. And ASTERIS was trained on specific Webb survey data; its performance on other instruments or observations may vary.
Still, for a field where every detected galaxy from the Cosmic Dawn helps constrain theories of how the universe evolved, tripling the discovery rate is a substantial advance.