A cross-disciplinary research team from Tsinghua University has announced a breakthrough in astronomical imaging with the development of ASTERIS (Astronomical Spatiotemporal Enhancement and Reconstruction for Image Synthesis), a pioneering AI model that significantly enhances our ability to see into the deepest reaches of the cosmos.
The research, published in the journal Science on February 20 with the title of “Deeper detection limits in astronomical imaging using self-supervised spatiotemporal denoising”, demonstrates how "self-supervised spatiotemporal denoising" can effectively increase the detection depth of the James Webb Space Telescope (JWST) by 1,0 magnitude (effectively enabling the telescope to detect objects 2.5 times fainter than previously possible). This "intelligent gain" has allowed scientists to map what is being called the most profound "Extreme Deep Field" in human history.
Overcoming the Cosmic Fog
The quest to understand the origins of the universe requires capturing light from incredibly faint, distant objects. However, even the most advanced telescopes must contend with a "fog" of background noise—a combination of zodiacal light from our solar system, diffuse light from the Milky Way, and thermal radiation from the telescope itself.
Traditional noise-reduction techniques rely on stacking multiple exposures, assuming noise is uniform and sometimes correlated. In reality, deep-space noise is complex and varies across both time and space. ASTERIS addresses this by reconstructing deep-space images as a 3D spatiotemporal volume. By using a "photometric adaptive screening mechanism," the model identifies the subtle fluctuations of noise and distinguishes them from the true, ultra-faint signals of distant stars and galaxies.
Science-First Artificial Intelligence
A common pitfall of standard AI image enhancement is the tendency to over-smooth data, which can inadvertently erase the dimmest signals or distort the actual shape of celestial bodies. To combat this, the Tsinghua team established a new evaluation paradigm: Science-driven AI.
The primary task of an AI model must be to ensure the scientific integrity and rigor of the data. ASTERIS utilizes a dual optimization strategy—"inter-frame median, intra-frame average"—to filter out transient interference while maximizing the signal-to-noise ratio of legitimate astronomical signals. This ensures that the resulting images are not just visually clearer, but scientifically accurate.
Unlocking the Cosmic Dawn
The results of this technological breakthrough are already reshaping our understanding of the early universe:
Triple Number of Early Galaxies: Using ASTERIS, the team identified over 160 candidate high-redshift galaxies from the "Cosmic Dawn" (200–500 million years after the Big Bang), tripling the number of discoveries compared to previous methods. This allowed the first precise measurement of the luminosity function for extremely faint galaxies at redshifts z≈9–16.
New Physical Insights: The findings reveal that the early universe contained a significantly higher proportion of faint galaxies than current theoretical models predict, providing crucial data for the study of "cosmic reionization" and challenging existing models.
Beyond these early galaxies, ASTERIS has proven highly effective at restoring the delicate spiral arms of distant galaxies and recovering the faint arcs of light created by gravitational lensing.
Research Leadership and Collaborations
This milestone is the result of a robust collaboration between Astronomy and the Departments of Automation at Tsinghua University. The paper’s co-first authors include postdoctoral researchers in the department of automation Yuduo Guo and Hao Zhang, and Department of Astronomy Ph.D. student Mingyu Li. The study's corresponding authors are Professor Qionghai Dai (Department of Automation), Associate Professor Zheng Cai (Department of Astronomy), and Associate Professor Jiamin Wu (Automation). Other key contributors include Yuhan Hao (Automation), and several members in the department of Astronomy, including associate professor Song Huang, postdoctoral researcher Yunjing Wu, and Ph.D. students Fujiang Yu, Yunjing Wu, and Xiaojing Lin.
As the technology demonstrates cross-platform compatibility—already showing success with data from the Subaru Telescope—it is poised to become a universal platform for next-generation missions such as the MUltiplexed Survey Telescope (MUST).
Paper link: DOI:10.1126/science.ady9404
